Article(id=1325493736079241471, tenantId=1146029695717560320, journalId=1324759914146529478, issueId=1325493569112388051, articleNumber=null, orderNo=null, doi=10.7544/issn1000-1239.202440623, pmid=null, cstr=null, oa=null, hot=null, price=null, onlineType=0, articleFormat=0, articleType=null, articleTypeStr=research-article, receivedDate=1721404800000, receivedDateStr=2024-07-20, revisedDate=1745164800000, revisedDateStr=2025-04-21, acceptedDate=1749571200000, acceptedDateStr=2025-06-11, onlineDate=1793951933164, onlineDateStr=2026-11-06, pubDate=1769875200000, pubDateStr=2026-02-01, doiRegisterDate=null, doiRegisterDateStr=null, onlineIssueDate=1793951933164, onlineIssueDateStr=2026-11-06, onlineJustAcceptDate=null, onlineJustAcceptDateStr=null, onlineFirstDate=null, onlineFirstDateStr=null, sourceXml=null, magXml=null, createTime=1793951933164, creator=13701087609, updateTime=1793951933164, updator=13701087609, issue=Issue{id=1325493569112388051, tenantId=1146029695717560320, journalId=1324759914146529478, year='2026', volume='63', issue='2', pageStart='275', pageEnd='538', issueExtLink='null', onlineDate='null', pubDate='1769875200000', pubDateStr='2026-02-01', beforeIssueId=null, nextIssueId=null, price=null, status=1, issueComplete=1, articleOrder=1, issueType=-1, specialIssue=null, createTime=1793951893355, creator='13701087609', updateTime=1793952511329, updator='13701087609', preIssue=null, nextIssue=null, articleTotal=null, ext={EN=IssueExt(id=1325496161167090054, tenantId=1146029695717560320, journalId=1324759914146529478, issueId=1325493569112388051, language=EN, specialIssueTitle=, coverIllustrator=null, specialIssueEditor=, specialIssueAbout=), CN=IssueExt(id=1325496161167090055, tenantId=1146029695717560320, journalId=1324759914146529478, issueId=1325493569112388051, language=CN, specialIssueTitle=, coverIllustrator=null, specialIssueEditor=, specialIssueAbout=)}, issueFiles=null, downloadFileDto=null}, startPage=477, endPage=504, ext={EN=ArticleExt(id=1325493736477700352, articleId=1325493736079241471, tenantId=1146029695717560320, journalId=1324759914146529478, language=EN, title=Survey of RDMA Congestion Control Techniques for Data Center Networks, columnId=1325493671541474142, journalTitle=Journal of Computer Research and Development, columnName=Network and Information Security, runingTitle=null, highlight=null, articleAbstract=

Congestion control is one of the key technologies for realizing high-performance data center networks, and it affects important network performance indicators such as throughput, latency, and packet loss rate. Over the past 20 years, with the continuous expansion of the scale of data centers and the increasing requirements of upper-layer applications for network performance, the deployment of remote direct memory access (RDMA) technology based on lossless underlying networks has received widespread attention within the industry. However, the priority-based flow control (PFC) mechanism, while maintaining a lossless network, introduces problems such as head-of-line blocking, leading to a decline in network performance or even network paralysis. As a crucial auxiliary means for achieving a lossless network, how to design a practical RDMA congestion control mechanism has become a hot issue. By dividing the congestion control process into congestion awareness and congestion regulation, this paper comprehensively reviews the research achievements in this field: Firstly, from the perspectives of explicit feedback and latency, different representative algorithms for congestion awareness are elaborated and summarized in detail; Secondly, representative algorithms for congestion regulation are introduced in detail from the dimensions of rate and window, and their advantages and disadvantages are summarized; Some optimization work of algorithms and congestion control algorithms based on reinforcement learning methods are supplemented; Finally, the existing challenges in this field are summarized and discussed.

, authors=Yutao Zhang, Hui Yang, Tao Li*, Mandi Huang, Tianyun Li, Zhigang Sun, authorsList=Yutao Zhang, Hui Yang, Tao Li, Mandi Huang, Tianyun Li, Zhigang Sun, authorCompany=null, correspAuthors=Tao Li, authorNote=

born in 2000. Master. His main research interests include RDMA and data center congestion control

born in 1986. PhD, associate professor. Her main research interests include RDMA and network processor

born in 1983. PhD, associate professor. His main research interest includes high-performance network system and chip

born in 1999. Master. Her main research interest includes RDMA

born in 2000. Master. Her main research interests include data center network, RDMA congestion control, and artificial intelligence

born in 1973. PhD, professor. His main research interests include software-defined network, time-sensitive networking, network architecture, and network security

, correspAuthorsNote=null, copyrightStatement=Copyright © 2026 Journal of Computer Research and Development. All rights reserved., copyrightOwner=null, extLink=null, articleAbsUrl=null, sourceXml=null, magXml=null, pdfUrl=null, pdf=null, pdfFileSize=null, pdfExtLink=null, richHtmlUrl=null, mobilePdfUrl=null, reviewReport=null, pdfFirstPage=null, abstractGraph=null, abstractGraphContent=null, abstractVideo=null, citation=null, cebUrl=null, magXmlContent=null, mapNumber=null, fund=null), CN=ArticleExt(id=1325493744383963453, articleId=1325493736079241471, tenantId=1146029695717560320, journalId=1324759914146529478, language=CN, title=数据中心网络RDMA拥塞控制技术综述, columnId=1325493672690713440, journalTitle=计算机研究与发展, columnName=网络与信息安全, runingTitle=null, highlight=null, articleAbstract=

拥塞控制是实现高性能数据中心网络的关键技术之一,影响吞吐量、延迟、丢包率等重要网络性能指标。过去20年间,随着数据中心规模不断扩大,上层应用对网络性能的要求不断提高,基于无损底层网络的远程直接内存访问(remote direct memory access,RDMA)技术在数据中心的部署受到了业内广泛关注。然而,基于优先级的流控(priority-based flow control,PFC)机制在维护无损网络的同时会引入头阻塞等问题,导致网络性能下降甚至网络瘫痪。作为实现无损网络的关键辅助手段,如何设计实用的RDMA拥塞控制机制成为了热点问题。通过将拥塞控制过程划分为拥塞感知与拥塞调整,全面综述了该领域的研究成果:首先从显式反馈与延迟的角度详细阐述并总结了不同的拥塞感知代表算法;其次从速率和窗口的维度对拥塞调整代表算法进行了详细介绍并对其优缺点进行了总结;而后补充了部分算法的优化工作以及基于强化学习方法的拥塞控制算法;最后总结并讨论了该领域存在的挑战。

, authors=张毓涛, 杨惠, 李韬*, 黄曼蒂, 李天云, 孙志刚, authorsList=张毓涛, 杨惠, 李韬, 黄曼蒂, 李天云, 孙志刚, authorCompany=null, correspAuthors=李韬, authorNote=

2000年生。硕士。主要研究方向为RDMA、数据中心拥塞控制

1986年生。博士,副研究员。主要研究方向为RDMA、网络处理器

1983年生。博士,副研究员。主要研究方向为高性能网络系统及芯片

1999年生。硕士。主要研究方向为RDMA

2000年生。硕士。主要研究方向为数据中心网络、RDMA拥塞控制、人工智能

1973年生。博士,研究员。主要研究方向为软件定义网络、时间敏感网络、网络体系结构、网络安全

, correspAuthorsNote=
李韬()
, copyrightStatement=版权所有 © 2026 计算机研究与发展编辑部, copyrightOwner=null, extLink=null, articleAbsUrl=null, sourceXml=Kq+g9IGXSGiDoYv829Ilpg==, magXml=kV6I+JcZqvosD0jKD6gQDA==, pdfUrl=null, pdf=EoTfeqlry5ow+Kep/AqGBA==, pdfFileSize=1961977, pdfExtLink=null, richHtmlUrl=null, mobilePdfUrl=null, reviewReport=null, pdfFirstPage=null, abstractGraph=kw0eqh1wf9bcmXdCFI1XWQ==, abstractGraphContent=null, abstractVideo=null, citation=null, cebUrl=null, magXmlContent=5aSHAw1AOOCC5YOEeevbuQ==, mapNumber=null, fund=null)}, authors=[Author(id=1325493744815976770, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=0, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=zhang_yt@nudt.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1325493744887279940, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493744815976770, language=EN, stringName=Yutao Zhang, firstName=Yutao, middleName=null, lastName=Zhang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 2000. Master. His main research interests include RDMA and data center congestion control

"}, bioImg=null, bioContent=

born in 2000. Master. His main research interests include RDMA and data center congestion control

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493744954388805, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493744815976770, language=CN, stringName=张毓涛, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=国防科技大学计算机学院 长沙 410073, bio={"img":"MRsYvq4PBHe8RA7p4dYueA==","content":"

2000年生。硕士。主要研究方向为RDMA、数据中心拥塞控制

"}, bioImg=MRsYvq4PBHe8RA7p4dYueA==, bioContent=

2000年生。硕士。主要研究方向为RDMA、数据中心拥塞控制

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])]), Author(id=1325493745017303367, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=1, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1325493745088606537, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493745017303367, language=EN, stringName=Hui Yang, firstName=Hui, middleName=null, lastName=Yang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1986. PhD, associate professor. Her main research interests include RDMA and network processor

"}, bioImg=null, bioContent=

born in 1986. PhD, associate professor. Her main research interests include RDMA and network processor

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493745164104010, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493745017303367, language=CN, stringName=杨惠, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=国防科技大学计算机学院 长沙 410073, bio={"img":"AwoySCYfASLKXOQILp89Ag==","content":"

1986年生。博士,副研究员。主要研究方向为RDMA、网络处理器

"}, bioImg=AwoySCYfASLKXOQILp89Ag==, bioContent=

1986年生。博士,副研究员。主要研究方向为RDMA、网络处理器

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])]), Author(id=1325493745256378700, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=2, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=litao@nudt.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=1, authorType=1, ext={EN=AuthorExt(id=1325493745336070478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493745256378700, language=EN, stringName=Tao Li, firstName=Tao, middleName=null, lastName=Li, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=*, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1983. PhD, associate professor. His main research interest includes high-performance network system and chip

"}, bioImg=null, bioContent=

born in 1983. PhD, associate professor. His main research interest includes high-performance network system and chip

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493745407373647, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493745256378700, language=CN, stringName=李韬, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=*, address=国防科技大学计算机学院 长沙 410073, bio={"img":"/gCcIXfqnULENXvqBdI1tA==","content":"

1983年生。博士,副研究员。主要研究方向为高性能网络系统及芯片

"}, bioImg=/gCcIXfqnULENXvqBdI1tA==, bioContent=

1983年生。博士,副研究员。主要研究方向为高性能网络系统及芯片

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])]), Author(id=1325493747080900945, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=3, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1325493747185758547, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747080900945, language=EN, stringName=Mandi Huang, firstName=Mandi, middleName=null, lastName=Huang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1999. Master. Her main research interest includes RDMA

"}, bioImg=null, bioContent=

born in 1999. Master. Her main research interest includes RDMA

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493747273838932, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747080900945, language=CN, stringName=黄曼蒂, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=国防科技大学计算机学院 长沙 410073, bio={"img":"q1ji+p0ZWOExy19IlxdlrA==","content":"

1999年生。硕士。主要研究方向为RDMA

"}, bioImg=q1ji+p0ZWOExy19IlxdlrA==, bioContent=

1999年生。硕士。主要研究方向为RDMA

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])]), Author(id=1325493747345142102, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=4, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1325493747437416792, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747345142102, language=EN, stringName=Tianyun Li, firstName=Tianyun, middleName=null, lastName=Li, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 2000. Master. Her main research interests include data center network, RDMA congestion control, and artificial intelligence

"}, bioImg=null, bioContent=

born in 2000. Master. Her main research interests include data center network, RDMA congestion control, and artificial intelligence

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493747508719961, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747345142102, language=CN, stringName=李天云, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=国防科技大学计算机学院 长沙 410073, bio={"img":"E7Joy2X7kyqZc7wPzZtJEw==","content":"

2000年生。硕士。主要研究方向为数据中心网络、RDMA拥塞控制、人工智能

"}, bioImg=E7Joy2X7kyqZc7wPzZtJEw==, bioContent=

2000年生。硕士。主要研究方向为数据中心网络、RDMA拥塞控制、人工智能

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])]), Author(id=1325493747588411739, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=5, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1325493747659714909, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747588411739, language=EN, stringName=Zhigang Sun, firstName=Zhigang, middleName=null, lastName=Sun, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1973. PhD, professor. His main research interests include software-defined network, time-sensitive networking, network architecture, and network security

"}, bioImg=null, bioContent=

born in 1973. PhD, professor. His main research interests include software-defined network, time-sensitive networking, network architecture, and network security

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493747735212382, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747588411739, language=CN, stringName=孙志刚, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=国防科技大学计算机学院 长沙 410073, bio={"img":"G4TR5c7IqQaR0gG+AqH14w==","content":"

1973年生。博士,研究员。主要研究方向为软件定义网络、时间敏感网络、网络体系结构、网络安全

"}, bioImg=G4TR5c7IqQaR0gG+AqH14w==, bioContent=

1973年生。博士,研究员。主要研究方向为软件定义网络、时间敏感网络、网络体系结构、网络安全

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])])], keywords=[Keyword(id=1325493747827487071, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, orderNo=1, keyword=data center networks (DCN)), Keyword(id=1325493747915567456, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, orderNo=2, keyword=remote direct memory access (RDMA)), Keyword(id=1325493747974287713, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, orderNo=3, keyword=RDMA over converged Ethernet (RoCEv2)), Keyword(id=1325493748033007970, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, orderNo=4, keyword=congestion control (CC)), Keyword(id=1325493748108505443, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, orderNo=5, keyword=congestion awareness), Keyword(id=1325493748179808612, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, orderNo=6, keyword=congestion regulation), Keyword(id=1325493748246917477, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, orderNo=1, keyword=数据中心网络), Keyword(id=1325493748314026342, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, orderNo=2, keyword=远程直接内存访问), Keyword(id=1325493748397912423, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, orderNo=3, keyword=RoCEv2), Keyword(id=1325493748477604200, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, orderNo=4, keyword=拥塞控制), Keyword(id=1325493748599239017, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, orderNo=5, keyword=拥塞感知), Keyword(id=1325493748683125098, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, orderNo=6, keyword=拥塞调整)], refs=[Reference(id=1325493757558272433, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=1, rfOrder=0, authorNames=null, journalName=null, refType=null, unstructuredReference=中国通服数字基建产业研究院. 中国数据中心产业发展白皮书[EB/OL]. [2024-03-24]. https: //aimg8.dlssyht.cn/u/551001/ueditor/file/276/551001/1684888884683143.pdf, articleTitle=null, refAbstract=null), Reference(id=1325493757625381298, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=1, rfOrder=1, authorNames=null, journalName=null, refType=null, unstructuredReference=China Communications Services Digital Infrastructure Industry Research Institute. White paper on the development of China’s data center industry[EB/OL]. [2024-03-24]. https://aimg8.dlssyht.cn/u/551001/ueditor/file/276/551001/1684888884683143.pdf (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1325493757688295859, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=2, rfOrder=2, authorNames=null, journalName=null, refType=null, unstructuredReference=Guo Chuanxiong, Wu Haitao, Deng Zhong, et al. RDMA over commodity ethernet at scale[C]//Proc of the 2016 ACM SIGCOMM Conf. New York: ACM, 2016: 202−215, articleTitle=null, refAbstract=null), Reference(id=1325493757759599028, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=3, rfOrder=3, authorNames=null, journalName=null, refType=null, unstructuredReference=Yan Dingyu, Liu Yaping, Zhang Shuo, at al. PCNP: A RoCEv2 congestion control using precise CNP[J]. Computer Nestworks, 2024, 247(C): 110453, articleTitle=null, refAbstract=null), Reference(id=1325493757847679413, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=4, rfOrder=4, authorNames=null, journalName=null, refType=null, unstructuredReference=Simons J. RDMA in the cloud: Enabling high-bandwidth, low-latency communication in virtual environments for HPC[EB/OL]. [2024-03-24]. https://octo.vmware.com/wp-content/uploads/2014/11/ndm2014-sc14-simons3.pdf, articleTitle=null, refAbstract=null), Reference(id=1325493757918982582, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=5, rfOrder=5, authorNames=null, journalName=null, refType=null, unstructuredReference=Hu Jinbin, Shen Houqiang, Liu Xuchong, et al. RDMA transports in datacenter networks: Survey[J]. IEEE Network, 2024, 38(6): 380−387, articleTitle=null, refAbstract=null), Reference(id=1325493757994480055, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=6, rfOrder=6, authorNames=null, journalName=null, refType=null, unstructuredReference=Mittal R, Shpiner A, Panda A, et al. Revisiting network support for RDMA[C]//Proc of the 2018 Conf of the ACM Special Interest Group on Data Communication. New York: ACM, 2018: 313−326, articleTitle=null, refAbstract=null), Reference(id=1325493758069977528, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=7, rfOrder=7, authorNames=null, journalName=null, refType=null, unstructuredReference=Wang Zilong, Luo Layong, Ning Qingsong, et al. SRNIC: A scalable architecture for RDMA NICs[C]//Proc of the 20th USENIX Symp on Networked Systems Design and Implementation (NSDI’23). Berkeley, CA: USENIX Association, 2023: 1−14, articleTitle=null, refAbstract=null), Reference(id=1325493758141280697, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=8, rfOrder=8, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhu Yibo, Eran H, Firestone D, et al. Congestion control for large-scale RDMA deployments[C]//Proc of the 2015 ACM Conf on Special Interest Group on Data Communication. New York: ACM, 2015: 523−536, articleTitle=null, refAbstract=null), Reference(id=1325493758225166778, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=9, rfOrder=9, authorNames=null, journalName=null, refType=null, unstructuredReference=Li Yuliang, Miao Rui, Liu Hongqiang, et al. HPCC: High precision congestion control[C]//Proc of the 2019 ACM Special Interest Group on Data Communication. New York: ACM, 2019: 44−58, articleTitle=null, refAbstract=null), Reference(id=1325493758288081339, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=10, rfOrder=10, authorNames=null, journalName=null, refType=null, unstructuredReference=Cheng Wenxue, Qian Kun, Jiang Wanchun, et al. Re-architecting congestion management in lossless Ethernet[C]//Proc of the 17th USENIX Symp on Networked Systems Design and Implementation (NSDI’20). Berkeley, CA: USENIX Association, 2020: 131−148, articleTitle=null, refAbstract=null), Reference(id=1325493758355190204, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=11, rfOrder=11, authorNames=null, journalName=null, refType=null, unstructuredReference=Taheri P, Menikkumbura D, Vanini E, et al. RoCC: Robust congestion control for RDMA[C]//Proc of the 16th Int Conf on Emerging Networking Experiments and Technologies. New York: ACM, 2020: 17−30, articleTitle=null, refAbstract=null), Reference(id=1325493758430687677, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=12, rfOrder=12, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Yiran, Liu Yifan, Meng Qingkai, et al. Congestion detection in lossless networks[C]//Proc of the 2021 ACM SIGCOMM 2021 Conf. New York: ACM, 2021: 370−383, articleTitle=null, refAbstract=null), Reference(id=1325493758510379454, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=13, rfOrder=13, authorNames=null, journalName=null, refType=null, unstructuredReference=Dai Huichen, Fu Binzhang, Tan Kun. LDCP: PFC-free low delay control protocol[EB/OL]. [2024-03-30]. https://datatracker.ietf.org/meeting/108/agenda/tsvwg-drafts.pdf, articleTitle=null, refAbstract=null), Reference(id=1325493758569099711, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=14, rfOrder=14, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Yiran, Meng Qingkai, Hu Chaolei, et al. Revisiting congestion control for lossless Ethernet[C]//Proc of the 21st USENIX Symp on Networked Systems Design and Implementation (NSDI’24). Berkeley, CA: USENIX Association, 2024: 131−148, articleTitle=null, refAbstract=null), Reference(id=1325493758648791488, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=15, rfOrder=15, authorNames=null, journalName=null, refType=null, unstructuredReference=IEEE 802.1. 802.1Qbb – Priority-based flow control[EB/OL]. [2024-04-10]. https://1.ieee802.org/dcb/802-1qbb/, articleTitle=null, refAbstract=null), Reference(id=1325493758728483265, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=16, rfOrder=16, authorNames=null, journalName=null, refType=null, unstructuredReference=Alizadeh M, Atikoglu B, Kabbani A, et al. Data center transport mechanisms: Congestion control theory and IEEE standardization[C]//Proc of the 46th Annual Allerton Conf on Communication, Control, and Computing. Piscataway, NJ: IEEE, 2008: 1270−1277, articleTitle=null, refAbstract=null), Reference(id=1325493758803980738, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=17, rfOrder=17, authorNames=null, journalName=null, refType=null, unstructuredReference=Stephens B, Cox A L, Singla A, et al. Practical DCB for improved data center networks[C]//Proc of the 2014 IEEE Conf on Computer Communications. Piscataway, NJ: IEEE, 2014: 1824−1832, articleTitle=null, refAbstract=null), Reference(id=1325493758862700995, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=18, rfOrder=18, authorNames=null, journalName=null, refType=null, unstructuredReference=Hu Shuihai, Zhu Yibo, Cheng Peng, et al. Deadlocks in datacenter networks: Why do they form, and how to avoid them[C]//Proc of the 15th ACM Workshop on Hot Topics in Networks. New York: ACM, 2016: 92−98, articleTitle=null, refAbstract=null), Reference(id=1325493760544616900, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=19, rfOrder=19, authorNames=null, journalName=null, refType=null, unstructuredReference=Shpiner A, Zahavi E, Zdornov V, et al. Unlocking credit loop deadlocks[C]//Proc of the 15th ACM Workshop on Hot Topics in Networks. New York: ACM, 2016: 85−91, articleTitle=null, refAbstract=null), Reference(id=1325493760628502981, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=20, rfOrder=20, authorNames=null, journalName=null, refType=null, unstructuredReference=Le Yanfang, Stephens B, Singhvi A, et al. RoGUE: RDMA over generic unconverged Ethernet[C]//Proc of the 2018 ACM Symp on Cloud Computing. New York: ACM, 2018: 225−236, articleTitle=null, refAbstract=null), Reference(id=1325493760699806150, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=21, rfOrder=21, authorNames=null, journalName=null, refType=null, unstructuredReference=Tang Jian, Xu Tingting, Nguyen C, et al. Tuning target delay for RTT-based congestion control[C]//Proc of the 30th IEEE Int Conf on Network Protocols (ICNP). Piscataway, NJ: IEEE, 2022: 1−11, articleTitle=null, refAbstract=null), Reference(id=1325493760771109319, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=22, rfOrder=22, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Jiao, Shi Jiaming, Zhong Xiaolong, et al. Receiver-driven RDMA congestion control by differentiating congestion types in datacenter networks[C]//Proc of the 29th IEEE Int Conf on Network Protocols (ICNP). Piscataway, NJ: IEEE, 2021: 1−12, articleTitle=null, refAbstract=null), Reference(id=1325493760842412488, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=23, rfOrder=23, authorNames=null, journalName=null, refType=null, unstructuredReference=Huang Shan, Dong Dezun, Bai Wei. Congestion control in high-speed lossless data center networks: A survey[J]. Future Generation Computer Systems, 2018, 89(C): 360−374, articleTitle=null, refAbstract=null), Reference(id=1325493760909521353, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=24, rfOrder=24, authorNames=null, journalName=null, refType=null, unstructuredReference=曾高雄, 胡水海, 张骏雪, 等. 数据中心网络传输协议综述[J]. 计算机研究与发展, 2020, 57(1): 74−84, articleTitle=null, refAbstract=null), Reference(id=1325493760993407434, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=24, rfOrder=25, authorNames=null, journalName=null, refType=null, unstructuredReference=Zeng Gaoxiong, Hu Shuihai, Zhang Junxue, et al. Transport protocols for data center networks: A survey[J]. Journal of Computer Research and Development, 2020, 57(1): 74−84 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1325493761077293515, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=25, rfOrder=26, authorNames=null, journalName=null, refType=null, unstructuredReference=杜鑫乐, 徐恪, 李彤, 等. 数据中心网络的流量控制: 研究现状与趋势[J]. 计算机学报, 2021, 44(7): 1287−1309, articleTitle=null, refAbstract=null), Reference(id=1325493761165373900, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=25, rfOrder=27, authorNames=null, journalName=null, refType=null, unstructuredReference=Du Xinle, Xü Ke, Li Tong, et al. Traffic control for data center network: State of the art and future research[J]. Chinese Journal of Computers, 2021, 44(7): 1287−1309 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1325493761236677069, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=26, rfOrder=28, authorNames=null, journalName=null, refType=null, unstructuredReference=Yan Dingyu, Liu Yaping, Zhang Shuo, at al. A survey of RoCEv2 congestion control[C]//Proc of the 7th Int Conf on Information Science, Communication and Computing. Berlin: Springer, 2023: 42−56, articleTitle=null, refAbstract=null), Reference(id=1325493761295397326, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=27, rfOrder=29, authorNames=null, journalName=null, refType=null, unstructuredReference=Alizadeh M, Greenberg A G, Maltz D A, et al. Data center TCP (DCTCP)[C]//Proc of the ACM SIGCOMM 2010 Conf. New York: ACM, 2010: 63−74, articleTitle=null, refAbstract=null), Reference(id=1325493761358311887, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=28, rfOrder=30, authorNames=null, journalName=null, refType=null, unstructuredReference=Alizadeh M, Yang Shuang, Sharif M, et al. pFabric: Minimal near-optimal datacenter transport[J]. SIGCOMM Computer Communication Review, 2013, 43(4): 435−446, articleTitle=null, refAbstract=null), Reference(id=1325493761429615056, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=29, rfOrder=31, authorNames=null, journalName=null, refType=null, unstructuredReference=Alizadeh M, Kabbani A, Edsall T, et al. Less is more: Trading a little bandwidth for ultra-low latency in the data center[C]//Proc of the 9th USENIX Conf on Networked Systems Design and Implementation. Berkeley, CA: USENIX Association, 2012: 1−14, articleTitle=null, refAbstract=null), Reference(id=1325493761500918225, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=30, rfOrder=32, authorNames=null, journalName=null, refType=null, unstructuredReference=Zats D, Das T, Mohan P, et al. DeTail: Reducing the flow completion time tail in datacenter networks[C]//Proc of the ACM SIGCOMM 2012 Conf on Applications, Technologies, Architectures, and Protocols for Computer Communication. New York: ACM, 2012: 139−150, articleTitle=null, refAbstract=null), Reference(id=1325493761580610002, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=31, rfOrder=33, authorNames=null, journalName=null, refType=null, unstructuredReference=Wilson C, Ballani H, Karagiannis T, et al. Better never than late: Meeting deadlines in datacenter networks[C]//Proc of the ACM SIGCOMM 2011 Conf. New York: ACM, 2011: 50−61, articleTitle=null, refAbstract=null), Reference(id=1325493761660301779, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=32, rfOrder=34, authorNames=null, journalName=null, refType=null, unstructuredReference=Hong Chiyao, Caesar M, Godfrey P B. Finishing flows quickly with preemptive scheduling[J]. SIGCOMM Computer Communication Review, 2012, 42(4): 127−138, articleTitle=null, refAbstract=null), Reference(id=1325493761731604948, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=33, rfOrder=35, authorNames=null, journalName=null, refType=null, unstructuredReference=Dean J, Barroso L A. The tail at scale[J] Communications of the ACM, 2013, 56(2): 74−80, articleTitle=null, refAbstract=null), Reference(id=1325493761828073941, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=34, rfOrder=36, authorNames=null, journalName=null, refType=null, unstructuredReference=Grosvenor M P, Schwarzkopf M, Gog I, et al. Queues don’t matter when you can JUMP them![C]//Proc of the 12th USENIX Symp on Networked Systems Design and Implementation (NSDI’15). Berkeley, CA: USENIX Association, 2015: 1−14, articleTitle=null, refAbstract=null), Reference(id=1325493761903571414, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=35, rfOrder=37, authorNames=null, journalName=null, refType=null, unstructuredReference=Vamanan B, Hasan J, Vijaykumar T N. Deadline-aware datacenter TCP (D2TCP)[C]//Proc of the ACM SIGCOMM 2012 Conf on Applications, Technologies, Architectures, and Protocols for Computer Communication. New York: ACM, 2012: 115−126, articleTitle=null, refAbstract=null), Reference(id=1325493761979068887, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=36, rfOrder=38, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Junxue, Bai Wei, Chen Kai. Enabling ECN for datacenter networks with RTT variations[C]//Proc of the 15th Int Conf on Emerging Networking Experiments and Technologies. New York: ACM, 2019: 233−245, articleTitle=null, refAbstract=null), Reference(id=1325493762046177752, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=37, rfOrder=39, authorNames=null, journalName=null, refType=null, unstructuredReference=Roy A, Zeng Hongyi, Bagga J, et al. Inside the social network’s (datacenter) network[C]//Proc of the 2015 ACM Conf on Special Interest Group on Data Communication. New York: ACM, 2015: 123−137, articleTitle=null, refAbstract=null), Reference(id=1325493762117480921, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=38, rfOrder=40, authorNames=null, journalName=null, refType=null, unstructuredReference=Evangelinos C, Hill C N. Cloud computing for parallel scientific HPC applications: Feasibility of running coupled atmosphere-ocean climate models on Amazon’s EC2[J]. Ratio, 2008, 2(2): 1−6, articleTitle=null, refAbstract=null), Reference(id=1325493762188784090, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=39, rfOrder=41, authorNames=null, journalName=null, refType=null, unstructuredReference=Dean J, Ghemawat S. MapReduce: Simplified data processing on large clusters[J]. Communications of the ACM, 2008, 51(1): 107−113, articleTitle=null, refAbstract=null), Reference(id=1325493762260087259, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=40, rfOrder=42, authorNames=null, journalName=null, refType=null, unstructuredReference=Kohavi R, Longbotham R. Online experiments: Lessons learned[J]. Computer, 2007, 40(9): 103−105, articleTitle=null, refAbstract=null), Reference(id=1325493762335584732, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=41, rfOrder=43, authorNames=null, journalName=null, refType=null, unstructuredReference=Wang Weitao, Moshref M, Li Y, et al. Poseidon: Efficient, robust, and practical datacenter CC via deployable INT[C]//Proc of the 20th USENIX Symp on Networked Systems Design and Implementation (NSDI’23). Berkeley, CA: USENIX Association, 2023: 255−274, articleTitle=null, refAbstract=null), Reference(id=1325493762419470813, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=42, rfOrder=44, authorNames=null, journalName=null, refType=null, unstructuredReference=Dragojevic A, Narayanan D, Hodson O, et al. FaRM: Fast remote memory[C]//Proc of the 11th USENIX Conf on Networked Systems Design and Implementation. Berkeley, CA: USENIX Association, 2014: 401−414, articleTitle=null, refAbstract=null), Reference(id=1325493762507551198, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=43, rfOrder=45, authorNames=null, journalName=null, refType=null, unstructuredReference=Hwang J, Cai Qizhe, Tang Ao, et al. TCP ≈ RDMA: CPU-efficient remote storage access with i10[C]//Proc of the 17th USENIX Symp on Networked Systems Design and Implementation (NSDI’20). Berkeley, CA: USENIX Association, 2020: 127−140, articleTitle=null, refAbstract=null), Reference(id=1325493762578854367, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=44, rfOrder=46, authorNames=null, journalName=null, refType=null, unstructuredReference=Cheng Huang, Simitci H, Xu Yikang, et al. Erasure coding in windows azure storage[C]//Proc of the 2012 USENIX Conf on Annual Technical. Berkeley, CA: USENIX Association, 2012: 1−12, articleTitle=null, refAbstract=null), Reference(id=1325493762650157536, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=45, rfOrder=47, authorNames=null, journalName=null, refType=null, unstructuredReference=Poke M, Hoefler T. DARE: High-performance state machine replication on RDMA networks[C]//Proc of the 24th Int Symp on High-Performance Parallel and Distributed Computing. New York: ACM, 2015: 107−118, articleTitle=null, refAbstract=null), Reference(id=1325493762725655009, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=46, rfOrder=48, authorNames=null, journalName=null, refType=null, unstructuredReference=Mitchell C, Geng Yifeng, Li Jinyang. Using one-sided RDMA reads to build a fast, CPU-efficient key-value store[C]//Proc of the 2013 USENIX Conf on Annual Technical. Berkeley, CA: USENIX Association, 2013: 103−114, articleTitle=null, refAbstract=null), Reference(id=1325493762801152482, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=47, rfOrder=49, authorNames=null, journalName=null, refType=null, unstructuredReference=Wu Haitao, Feng Zhenqian, Guo Chuanxiong, et al. ICTCP: Incast congestion control for TCP in data-center networks[J]. IEEE/ACM Transactions on Networking, 2013, 21(2): 345−358, articleTitle=null, refAbstract=null), Reference(id=1325493762880844259, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=48, rfOrder=50, authorNames=null, journalName=null, refType=null, unstructuredReference=Lao ChonLam, Le Yanfang, Mahajan K, et al. ATP: In-network aggregation for multi-tenant learning[C]//Proc of the 18th USENIX Symp on Networked Systems Design and Implementation (NSDI’21). Berkeley, CA: USENIX Association, 2021: 741−761, articleTitle=null, refAbstract=null), Reference(id=1325493762947953124, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=49, rfOrder=51, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Jiao, Zhang Yali, Guan Zixuan, et al. HierCC: Hierarchical RDMA congestion control[C]//Proc of the 5th Asia-Pacific Workshop on Networking (APNet’2021). New York: ACM, 2021: 29−36, articleTitle=null, refAbstract=null), Reference(id=1325493763036033509, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=50, rfOrder=52, authorNames=null, journalName=null, refType=null, unstructuredReference=Kumar G, Dukkipati N, Jang K, et al. Swift: Delay is simple and effective for congestion control in the datacenter[C]//Proc of the Annual Conf of the ACM Special Interest Group on Data Communication on the Applications, Technologies, Architectures, and Protocols for Computer Communication. New York: ACM, 2020: 514−528, articleTitle=null, refAbstract=null), Reference(id=1325493763111530982, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=51, rfOrder=53, authorNames=null, journalName=null, refType=null, unstructuredReference=Addanki V, Michel O, Schmid S. POWERTCP: Pushing the performance limits of datacenter networks[C]//Proc of the 19th USENIX Symp on Networked Systems Design and Implementation (NSDI’22). Berkeley, CA: USENIX Association, 2022: 51−70, articleTitle=null, refAbstract=null), Reference(id=1325493763174445543, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=52, rfOrder=54, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhong Xiaolong, Zhang Jiao, Zhang Yali, et al. PACC: Proactive and accurate congestion feedback for RDMA congestion control[C]//Proc of the 2022 IEEE Conf on Computer Communications (INFOCOM 2022). Piscataway, NJ: IEEE, 2022: 2228−2237, articleTitle=null, refAbstract=null), Reference(id=1325493763245748712, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=53, rfOrder=55, authorNames=null, journalName=null, refType=null, unstructuredReference=Infiniband Trade Association. InfiniBand architecture volume 1, general specifications, release 1.2.1[EB/OL]. [2024-04-23]. https://www.infinibandta.org/ibta-specification/, articleTitle=null, refAbstract=null), Reference(id=1325493763333829097, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=54, rfOrder=56, authorNames=null, journalName=null, refType=null, unstructuredReference=Recio R J, Culley P R, Garcia D, et al. A remote direct memory access protocol specification[EB/OL]. [2024-04-23]. https://doi.org/10.17487/RFC5040, articleTitle=null, refAbstract=null), Reference(id=1325493764994773482, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=55, rfOrder=57, authorNames=null, journalName=null, refType=null, unstructuredReference=Infiniband Trade Association. Supplement to infiniBand architecture specification volume 1 release 1.2.2 annex A16: RDMA over converged Ethernet (RoCE)[EB/OL]. [2024-04-23]. https://www.infinibandta.org/ibta-specification/, articleTitle=null, refAbstract=null), Reference(id=1325493765070270955, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=56, rfOrder=58, authorNames=null, journalName=null, refType=null, unstructuredReference=Infiniband Trade Association. Supplement to infiniBand architecture specification volume 1 release 1.2.2 annex A17: RoCEv2[EB/OL].[2024-04-23]. https://www.infinibandta.org/ibta-specification/, articleTitle=null, refAbstract=null), Reference(id=1325493765145768428, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=57, rfOrder=59, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhu Yibo, Ghobadi M, Misra V, et al. ECN or delay: Lessons learnt from analysis of DCQCN and TIMELY[C]//Proc of the 12th Int Conf on Emerging Networking Experiments and Technologies. New York: ACM, 2016: 313−327, articleTitle=null, refAbstract=null), Reference(id=1325493765221265901, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=58, rfOrder=60, authorNames=null, journalName=null, refType=null, unstructuredReference=Intel. Intel E810 Ethernet network adapter[EB/OL]. [2024-04-27]. https://www.intel.cn/content/www/cn/zh/products/details/ethernet/800-network-adapters/e810-network-adapters/products.html, articleTitle=null, refAbstract=null), Reference(id=1325493765317734894, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=59, rfOrder=61, authorNames=null, journalName=null, refType=null, unstructuredReference=Broadcom. Broadcom BCM957508-P1200G NIC datasheet[EB/OL]. [2024-04-27]. https://docs.broadcom.com/doc/957508-P1200G-DS1XX, articleTitle=null, refAbstract=null), Reference(id=1325493765397426671, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=60, rfOrder=62, authorNames=null, journalName=null, refType=null, unstructuredReference=NVIDIA. NVIDIA Connectx-7 network adapter datasheet[EB/OL]. [2024-04-27]. https://nvdam.widen.net/s/csf8rmnqwl/infiniband-ethernet-datasheet-connectx-7-ds-nv-us-2544471, articleTitle=null, refAbstract=null), Reference(id=1325493765460341232, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=61, rfOrder=63, authorNames=null, journalName=null, refType=null, unstructuredReference=MARVELL. MARVELL converged Ethernet network adapter[EB/OL]. [2024-04-27]. https://cn.marvell.com/products/ethernet-adapters-andcontrollers/fastlinq-cna-adapters/documents.html, articleTitle=null, refAbstract=null), Reference(id=1325493765540033009, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=62, rfOrder=64, authorNames=null, journalName=null, refType=null, unstructuredReference=华为. 华为SP600网卡[EB/OL]. [2024-04-27]. https: //support.huawei.com/enterprise/zh/doc/EDOC1100309168?idPath=23710424%7C251364409%7C21782478%7C15791, articleTitle=null, refAbstract=null), Reference(id=1325493765640696306, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=62, rfOrder=65, authorNames=null, journalName=null, refType=null, unstructuredReference=Huawei. Huawei SP600 NIC[EB/OL]. [2024-04-27]. https://support.huawei.com/enterprise/zh/doc/EDOC1100309168?idPath=23710424%7C251364409%7C21782478%7C15791 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1325493765707805171, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=63, rfOrder=66, authorNames=null, journalName=null, refType=null, unstructuredReference=Chelsio. Chelsio T7 DPU products[EB/OL]. [2024-04-27]. https://www.chelsio.com/wp-content/uploads/resources/t7-dpu-asic.pdf, articleTitle=null, refAbstract=null), Reference(id=1325493765774914036, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=64, rfOrder=67, authorNames=null, journalName=null, refType=null, unstructuredReference=Lee C, Park C, Jang K, et al. DX: Latency-based congestion control for datacenters[J]. IEEE/ACM Transactions Networking, 2017, 25(1): 335−348, articleTitle=null, refAbstract=null), Reference(id=1325493765842022901, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=65, rfOrder=68, authorNames=null, journalName=null, refType=null, unstructuredReference=Mittal R, Lam V T, Dukkipati N, et al. TIMELY: RTT-based congestion control for the datacenter[C]//Proc of the 2015 ACM Conf on Special Interest Group on Data Communication. New York: ACM, 2015: 537−550, articleTitle=null, refAbstract=null), Reference(id=1325493765909131766, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=66, rfOrder=69, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhou Renjie, Dong Dezun, Huang Shan, et al. FastTune: Timely and precise congestion control in data center network[C]//Proc of the 2021 IEEE Int Conf on Parallel & Distributed Processing with Applications, Big Data & Cloud Computing, Sustainable Computing & Communications, Social Computing & Networking (ISPA/BDCloud/SocialCom/SustainCom). Piscataway, NJ: IEEE, 2021: 238−245, articleTitle=null, refAbstract=null), Reference(id=1325493765976240631, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=67, rfOrder=70, authorNames=null, journalName=null, refType=null, unstructuredReference=Arslan S, Li Yuliang, Kumar G, et al. Bolt: Sub-RTT congestion control for ultra-low latency[C]//Proc of the 20th USENIX Symp on Networked Systems Design and Implementation (NSDI’23). Berkeley, CA: USENIX Association, 2023: 219−236, articleTitle=null, refAbstract=null), Reference(id=1325493766043349496, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=68, rfOrder=71, authorNames=null, journalName=null, refType=null, unstructuredReference=Floyd S, Jacobson V. Random early detection gateways for congestion avoidance[J]. IEEE/ACM Transactions on Networking, 1993, 1(4): 397−413, articleTitle=null, refAbstract=null), Reference(id=1325493766110458361, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=69, rfOrder=72, authorNames=null, journalName=null, refType=null, unstructuredReference=Ramakrishnan K, Floyd S, Black D. The addition of explicit congestion notification (ECN) to IP[EB/OL]. [2024-05-02]. https://doi.org/10.17487/rfc3168, articleTitle=null, refAbstract=null), Reference(id=1325493766169178618, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=70, rfOrder=73, authorNames=null, journalName=null, refType=null, unstructuredReference=IEEE. IEEE 802.1 qau-congestion notification[EB/OL]. [2024-05-02]. http://www.ieee802.org/1/pages/802.1au.html, articleTitle=null, refAbstract=null), Reference(id=1325493766227898875, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=71, rfOrder=74, authorNames=null, journalName=null, refType=null, unstructuredReference=Wu Haitao, Ju Jiabo, Lu Guohan, et al. Tuning ECN for data center networks[C]//Proc of the 8th Int Conf on Emerging Networking Experiments and Technologies. New York: ACM, 2012: 25−36, articleTitle=null, refAbstract=null), Reference(id=1325493766286619132, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=72, rfOrder=75, authorNames=null, journalName=null, refType=null, unstructuredReference=Shan Danfeng, Ren Fengyuan. ECN marking with micro-burst traffic: Problem, analysis, and improvement[J] IEEE/ACM Transactions on Networking, 2018, 26(4): 1533−1546, articleTitle=null, refAbstract=null), Reference(id=1325493766345339389, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=73, rfOrder=76, authorNames=null, journalName=null, refType=null, unstructuredReference=Qian Kun, Cheng Wenxue, Zhang Tong, et al. Gentle flow control: Avoiding deadlock in lossless networks[C]//Proc of the ACM Special Interest Group on Data Communication. New York: ACM, 2019: 75−89, articleTitle=null, refAbstract=null), Reference(id=1325493766416642558, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=74, rfOrder=77, authorNames=null, journalName=null, refType=null, unstructuredReference=The P4. org Applications Working Group. In-band network telemetry (INT) dataplane specification[EB/OL]. [2024-05-12]. https://p4.org/p4-spec/docs/INT_v2_1.pdf, articleTitle=null, refAbstract=null), Reference(id=1325493766487945727, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=75, rfOrder=78, authorNames=null, journalName=null, refType=null, unstructuredReference=Ben Basat R, Ramanathan S, Li Yuliang, et al. PINT: Probabilistic in-band network telemetry[C]//Proc of the Annual Conf of the ACM Special Interest Group on Data Communication on the Applications, Technologies, Architectures, and Protocols for Computer Communication. New York: ACM, 2020: 662−680, articleTitle=null, refAbstract=null), Reference(id=1325493766567637504, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=76, rfOrder=79, authorNames=null, journalName=null, refType=null, unstructuredReference=Chelsio. Chelsio T5 packet rate performance report[EB/OL]. [2024-05-13]. https://www.chelsio.com/wp-content/uploads/2013/08/Chelsio-T5-Packet-Rate-Performance-Report.pdf, articleTitle=null, refAbstract=null), Reference(id=1325493766638940673, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=77, rfOrder=80, authorNames=null, journalName=null, refType=null, unstructuredReference=Silicom. Dual port 10 gigabit server adapter with precision time stamping[EB/OL]. [2024-05-13]. https://www.silicom-usa.com/wp-content/uploads/2016/08/PE210G2TSI9-10G-Precision-Time-Stamping-Server-Adapter.pdf, articleTitle=null, refAbstract=null), Reference(id=1325493766706049538, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=78, rfOrder=81, authorNames=null, journalName=null, refType=null, unstructuredReference=ntop. Using hardware timestamps with PF RING[EB/OL]. [2024-05-13]. https://www.ntop.org/pf_ring/hardware-timestamps-with-pf_ring/#:~:text=PF_RING%20has%20been%20enhanced%20to%20support%20hw%20timestamps%2C,the%20NIC%20that%20received%20packets%20must%20be%20timestamped, articleTitle=null, refAbstract=null), Reference(id=1325493766773158403, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=79, rfOrder=82, authorNames=null, journalName=null, refType=null, unstructuredReference=ntop. Who (really) needs sub-microsecond packet simestamps[EB/OL]. [2024-05-13]. https://www.ntop.org/pf_ring/who-really-needs-sub-microsecond-packet-timestamps/, articleTitle=null, refAbstract=null), Reference(id=1325493766848655876, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=80, rfOrder=83, authorNames=null, journalName=null, refType=null, unstructuredReference=NVIDIA Mellanox. Highly accurate time synchronization with connectX-3 and timeKeeper[EB/OL]. [2024-05-13]. https://enterprise-support.nvidia.com/s/article/highly-accurate-time-synchronization-with-connectx--174--3-and-timekeeper--174-x, articleTitle=null, refAbstract=null), Reference(id=1325493766924153349, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=81, rfOrder=84, authorNames=null, journalName=null, refType=null, unstructuredReference=Brakmo L S, O’Malley S W, Peterson L L. TCP Vegas: New techniques for congestion detection and avoidance[J]. SIGCOMM Computer Communication Review, 1994, 24(4): 24−35, articleTitle=null, refAbstract=null), Reference(id=1325493767138062854, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=82, rfOrder=85, authorNames=null, journalName=null, refType=null, unstructuredReference=Wei D X, Cheng Jin, Low S H, et al. FAST TCP: Motivation, architecture, algorithms, performance[J]. IEEE/ACM Transactions on Networking, 2006, 14(6): 1246−1259, articleTitle=null, refAbstract=null), Reference(id=1325493767226143239, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=83, rfOrder=86, authorNames=null, journalName=null, refType=null, unstructuredReference=Tan Kun, Song Jingmin, Zhang Qian, et al. A compound TCP approach for high-speed and long distance networks[C]//Proc of the 25th IEEE Int Conf on Computer Communications (INFOCOM’2006). Piscataway, NJ: IEEE, 2006: 1−12, articleTitle=null, refAbstract=null), Reference(id=1325493767284863496, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=84, rfOrder=87, authorNames=null, journalName=null, refType=null, unstructuredReference=Jiang Huiling, Li Qing, Jiang Yong, et al. When machine learning meets congestion control: A survey and comparison[J]. Computer Networks, 2021, 192: 108033, articleTitle=null, refAbstract=null), Reference(id=1325493767381332489, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=85, rfOrder=88, authorNames=null, journalName=null, refType=null, unstructuredReference=Fuhrer B, Shpigelman Y, Tessler C, et al. Implementing reinforcement learning datacenter congestion control in NVIDIA NICs[C]//Proc of the 23rd IEEE/ACM Int Symp on Cluster, Cloud and Internet Computing (CCGrid). New York: ACM, 2022: 331−343, articleTitle=null, refAbstract=null), Reference(id=1325493767444247050, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=86, rfOrder=89, authorNames=null, journalName=null, refType=null, unstructuredReference=Dong Mo, Meng Tong, Zarchy D, et al. PCC vivace: Online-learning congestion control[C]//Proc of the 15th USENIX Conf on Networked Systems Design and Implementation. Berkeley, CA: USENIX Association, 2018: 343−356, articleTitle=null, refAbstract=null), Reference(id=1325493767507161611, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=87, rfOrder=90, authorNames=null, journalName=null, refType=null, unstructuredReference=Sutton R S, Barto A G. Reinforcement Learning: An Introduction[M]. Cambridge, MA: A Bradford Book, 2018, articleTitle=null, refAbstract=null), Reference(id=1325493767574270476, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=88, rfOrder=91, authorNames=null, journalName=null, refType=null, unstructuredReference=Tessler C, Shpigelman Y, Dalal G, et al. Reinforcement learning for datacenter congestion control[J]. SIGMETRICS Performance Evaluation Review, 2022, 49(2): 43−46, articleTitle=null, refAbstract=null), Reference(id=1325493767632990733, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=89, rfOrder=92, authorNames=null, journalName=null, refType=null, unstructuredReference=Callon R. RFC1925: The twelve networking[EB/OL]. [2024-05-02]. https://doi.org/10.17487/RFC1925, articleTitle=null, refAbstract=null)], funds=null, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])], figs=[ArticleFig(id=1325493748808954219, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 1, caption=Comparison of RDMA and TCP communication methods, figureFileSmall=lTAOnf69fBJWIHrWAUObKw==, figureFileBig=kw0eqh1wf9bcmXdCFI1XWQ==, tableContent=null), ArticleFig(id=1325493748876063084, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图1, caption=RDMA与TCP的通信方式对比, figureFileSmall=lTAOnf69fBJWIHrWAUObKw==, figureFileBig=kw0eqh1wf9bcmXdCFI1XWQ==, tableContent=null), ArticleFig(id=1325493749073195373, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 2, caption=Three implementations of RDMA, figureFileSmall=Ulo/Z2s4/4DLmvndjNLxUw==, figureFileBig=gN8ROLohe1zBqfkEYLKUBA==, tableContent=null), ArticleFig(id=1325493749127721326, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图2, caption=RDMA的3种实现方式, figureFileSmall=Ulo/Z2s4/4DLmvndjNLxUw==, figureFileBig=gN8ROLohe1zBqfkEYLKUBA==, tableContent=null), ArticleFig(id=1325493749194830191, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 3, caption=Working principle and trigger mechanism of PFC, figureFileSmall=1OckAkGYOI/d0uZYg2PFkQ==, figureFileBig=juRJfwgu6NOnZz6MW1Y0gA==, tableContent=null), ArticleFig(id=1325493749270327664, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图3, caption=PFC的工作原理和触发机制, figureFileSmall=1OckAkGYOI/d0uZYg2PFkQ==, figureFileBig=juRJfwgu6NOnZz6MW1Y0gA==, tableContent=null), ArticleFig(id=1325493749341630833, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 4, caption=Victimization flow problems caused by incast, figureFileSmall=JLXBU9xPAYchVtqhGRA98A==, figureFileBig=bqqybhrDlwMymS+OXHuP8g==, tableContent=null), ArticleFig(id=1325493749425516914, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图4, caption=incast造成的受害流问题, figureFileSmall=JLXBU9xPAYchVtqhGRA98A==, figureFileBig=bqqybhrDlwMymS+OXHuP8g==, tableContent=null), ArticleFig(id=1325493749480042867, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 5, caption=RED mechanism, figureFileSmall=GELfLTSzTpe3L3+5zxEsDQ==, figureFileBig=jeI2TyHTHkQ76bczZ1ZGlA==, tableContent=null), ArticleFig(id=1325493749635232116, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图5, caption=RED机制, figureFileSmall=GELfLTSzTpe3L3+5zxEsDQ==, figureFileBig=jeI2TyHTHkQ76bczZ1ZGlA==, tableContent=null), ArticleFig(id=1325493749698146677, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 6, caption=Congestion control mechanism of QCN, figureFileSmall=y7fj/eIytgdxuzK6sxwF9w==, figureFileBig=BLI6KnCvJ7uYZFKmv0pTvw==, tableContent=null), ArticleFig(id=1325493749756866934, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图6, caption=QCN的拥塞控制机制, figureFileSmall=y7fj/eIytgdxuzK6sxwF9w==, figureFileBig=BLI6KnCvJ7uYZFKmv0pTvw==, tableContent=null), ArticleFig(id=1325493749811392887, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 7, caption=Packet format of CNP, figureFileSmall=MXZ5F+B/1rWJZLYHDeeoZw==, figureFileBig=h8XDwSEy5KaqOTBH7XCgrw==, tableContent=null), ArticleFig(id=1325493749870113144, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图7, caption=CNP数据包格式, figureFileSmall=MXZ5F+B/1rWJZLYHDeeoZw==, figureFileBig=h8XDwSEy5KaqOTBH7XCgrw==, tableContent=null), ArticleFig(id=1325493751526863225, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 8, caption=Forwarding mode of INT-MD, figureFileSmall=rLe96CdGyjz1TFq+BbGA9g==, figureFileBig=Jj1e2GFYmBd0FfdccL9mJA==, tableContent=null), ArticleFig(id=1325493751623332218, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图8, caption=INT-MD转发模式, figureFileSmall=rLe96CdGyjz1TFq+BbGA9g==, figureFileBig=Jj1e2GFYmBd0FfdccL9mJA==, tableContent=null), ArticleFig(id=1325493751707218299, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 9, caption=Switch processing logic of FastTune, figureFileSmall=92Mnuw3TpWbxwBJie+H6GA==, figureFileBig=W0EoHapSUHZxCgWTSxK8XQ==, tableContent=null), ArticleFig(id=1325493751778521468, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图9, caption=FastTune的交换机处理逻辑, figureFileSmall=92Mnuw3TpWbxwBJie+H6GA==, figureFileBig=W0EoHapSUHZxCgWTSxK8XQ==, tableContent=null), ArticleFig(id=1325493751849824637, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 10, caption=Switch processing logic of Bolt, figureFileSmall=ahmdvChIkf4WCdiV69dGkg==, figureFileBig=cKaUxrMJqX+o0WpD8TlCBQ==, tableContent=null), ArticleFig(id=1325493751929516414, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图10, caption=Bolt的交换机处理逻辑, figureFileSmall=ahmdvChIkf4WCdiV69dGkg==, figureFileBig=cKaUxrMJqX+o0WpD8TlCBQ==, tableContent=null), ArticleFig(id=1325493752025985407, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 11, caption=Switch framework of RoCC, figureFileSmall=1QT3wluGVKWQ7nhvbzs8wg==, figureFileBig=XUW8p6MFScS8dK6zqO84Uw==, tableContent=null), ArticleFig(id=1325493752105677184, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图11, caption=RoCC交换机框架, figureFileSmall=1QT3wluGVKWQ7nhvbzs8wg==, figureFileBig=XUW8p6MFScS8dK6zqO84Uw==, tableContent=null), ArticleFig(id=1325493752185368961, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 12, caption=Rate calculation process of RoCC, figureFileSmall=SCbEGP2Ws5o+VSuKHTgB2Q==, figureFileBig=kmyo9GtqiBMBIUXGyHzCMg==, tableContent=null), ArticleFig(id=1325493752256672130, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图12, caption=RoCC的速率计算过程, figureFileSmall=SCbEGP2Ws5o+VSuKHTgB2Q==, figureFileBig=kmyo9GtqiBMBIUXGyHzCMg==, tableContent=null), ArticleFig(id=1325493752332169603, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 13, caption=Overall framework of HierCC, figureFileSmall=RJ5a2Rl3TVRX3imUqLTwDA==, figureFileBig=rk8o6+ZMHTCQecMcYn88iw==, tableContent=null), ArticleFig(id=1325493752437027204, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图13, caption=HierCC整体框架, figureFileSmall=RJ5a2Rl3TVRX3imUqLTwDA==, figureFileBig=rk8o6+ZMHTCQecMcYn88iw==, tableContent=null), ArticleFig(id=1325493752516718981, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 14, caption=Processing logic of PACC switch, figureFileSmall=pNm1HzrzAtgmfs/c7mG0Tw==, figureFileBig=2GVaVtSpqKsHX8FTzt85RA==, tableContent=null), ArticleFig(id=1325493752583827846, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图14, caption=PACC交换机的处理逻辑, figureFileSmall=pNm1HzrzAtgmfs/c7mG0Tw==, figureFileBig=2GVaVtSpqKsHX8FTzt85RA==, tableContent=null), ArticleFig(id=1325493752663519623, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 15, caption=Data segment sending process and RTT calculation of TIMELY, figureFileSmall=1RrwODl8t0Sf2HGLqF3x0A==, figureFileBig=9bv0eDI+MedQDO1V71JmcQ==, tableContent=null), ArticleFig(id=1325493752759988616, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图15, caption=TIMELY的数据段发送流程与RTT计算, figureFileSmall=1RrwODl8t0Sf2HGLqF3x0A==, figureFileBig=9bv0eDI+MedQDO1V71JmcQ==, tableContent=null), ArticleFig(id=1325493752818708873, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 16, caption=Complete data sending and receiving process and RTT components of Swift, figureFileSmall=xo0roZ4oLxQlQI0L50Yg+w==, figureFileBig=LlSQKuZMvGiR3snpb+cLHg==, tableContent=null), ArticleFig(id=1325493752885817738, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图16, caption=Swift完整的数据收发过程以及RTT组成部分, figureFileSmall=xo0roZ4oLxQlQI0L50Yg+w==, figureFileBig=LlSQKuZMvGiR3snpb+cLHg==, tableContent=null), ArticleFig(id=1325493752952926603, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 17, caption=RTT calculation process of Swift, figureFileSmall=xKVFu8Ul73fdHSbv22ZCFw==, figureFileBig=sJ8pmywIN/UEXf/pS6bi8A==, tableContent=null), ArticleFig(id=1325493753036812684, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图17, caption=Swift 的RTT计算过程, figureFileSmall=xKVFu8Ul73fdHSbv22ZCFw==, figureFileBig=sJ8pmywIN/UEXf/pS6bi8A==, tableContent=null), ArticleFig(id=1325493753103921549, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 18, caption=The process of calculating the target delay for RET, figureFileSmall=xK/8igc9JoawEXbbxyIC7g==, figureFileBig=E/t72t4tl8Rcb8fn2Bf3OA==, tableContent=null), ArticleFig(id=1325493753171030414, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图18, caption=RET的目标延迟计算过程, figureFileSmall=xK/8igc9JoawEXbbxyIC7g==, figureFileBig=E/t72t4tl8Rcb8fn2Bf3OA==, tableContent=null), ArticleFig(id=1325493753250722191, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 19, caption=Rate-adjusted state machine of DCQCN, figureFileSmall=2OuMAOAxMqUrvFdOqT7zzg==, figureFileBig=M5sQTGAsEj35gogjhwMFEg==, tableContent=null), ArticleFig(id=1325493753317831056, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图19, caption=DCQCN的速率调整状态机, figureFileSmall=2OuMAOAxMqUrvFdOqT7zzg==, figureFileBig=M5sQTGAsEj35gogjhwMFEg==, tableContent=null), ArticleFig(id=1325493753397522833, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 20, caption=Rate adjustment process of TIMELY, figureFileSmall=eWMU5vptPMzmy3A/yMC8kg==, figureFileBig=SHuaUoqP3M2WIalN8OjIbw==, tableContent=null), ArticleFig(id=1325493753456243090, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图20, caption=TIMELY的速率调整过程, figureFileSmall=eWMU5vptPMzmy3A/yMC8kg==, figureFileBig=SHuaUoqP3M2WIalN8OjIbw==, tableContent=null), ArticleFig(id=1325493753523351955, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 21, caption=Window PID control process of RCC, figureFileSmall=1JgyrEd7g37MHyyvZMpzbQ==, figureFileBig=vsKihO4SSCUx2w268QAwIA==, tableContent=null), ArticleFig(id=1325493753586266516, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图21, caption=RCC的窗口PID控制过程, figureFileSmall=1JgyrEd7g37MHyyvZMpzbQ==, figureFileBig=vsKihO4SSCUx2w268QAwIA==, tableContent=null), ArticleFig(id=1325493753665958293, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 22, caption=Window adjustment process for RET, figureFileSmall=DBBj4w//O3tgtia5qfs4RQ==, figureFileBig=uGag7wJTCm1H+DhQE2YfSA==, tableContent=null), ArticleFig(id=1325493753749844374, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图22, caption=RET的窗口调整过程, figureFileSmall=DBBj4w//O3tgtia5qfs4RQ==, figureFileBig=uGag7wJTCm1H+DhQE2YfSA==, tableContent=null), ArticleFig(id=1325493753816953239, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 23, caption=Comparison of adjustment strategies for TIMELY and Patched TIMELY, figureFileSmall=ogR+mcOBksaoTRXqSkp1nw==, figureFileBig=nFndekVsIo761ZsenQ/iwA==, tableContent=null), ArticleFig(id=1325493753892450712, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图23, caption=TIMELY和Patched TIMELY的调整策略对比, figureFileSmall=ogR+mcOBksaoTRXqSkp1nw==, figureFileBig=nFndekVsIo761ZsenQ/iwA==, tableContent=null), ArticleFig(id=1325493754014085529, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Figure 24, caption=Labeling strategies for CEDM, figureFileSmall=PWu7wyG0wQA15ZsIf/mbZQ==, figureFileBig=zowQ+vITg/ik0pjYuvwzAg==, tableContent=null), ArticleFig(id=1325493754081194394, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=图24, caption=CEDM的标记策略, figureFileSmall=PWu7wyG0wQA15ZsIf/mbZQ==, figureFileBig=zowQ+vITg/ik0pjYuvwzAg==, tableContent=null), ArticleFig(id=1325493754160886171, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Table 1, caption=Congestion Control Algorithms of Data Center RDMA, figureFileSmall=null, figureFileBig=null, tableContent=
算法年份拥塞感知信号拥塞调整方式
DCQCN[8]2015ECN速率
TIMELY[65]2015RTT速率
Patched TIMELY[57]2016RTT速率
RoGUE[20]2018RTT速率+窗口
HPCC[9]2019INT速率+窗口
PCN[10]2020ECN速率
Swift[50]2020RTT速率+窗口
RoCC[11]2020网内反馈速率
FastTune[66]2021INT速率+窗口
HierCC[49]2021网内反馈速率
LDCP[13]2021ECN窗
RCC[22]2021RTT窗口
PowerTCP[51]2022INT速率+窗口
RET[21]2022RTT速率
PACC[52]2022网内反馈速率
Bolt[67]2023INT窗口
ACC[14]2024ECN速率
), ArticleFig(id=1325493754248966556, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=表1, caption=

数据中心RDMA拥塞控制算法

, figureFileSmall=null, figureFileBig=null, tableContent=
算法年份拥塞感知信号拥塞调整方式
DCQCN[8]2015ECN速率
TIMELY[65]2015RTT速率
Patched TIMELY[57]2016RTT速率
RoGUE[20]2018RTT速率+窗口
HPCC[9]2019INT速率+窗口
PCN[10]2020ECN速率
Swift[50]2020RTT速率+窗口
RoCC[11]2020网内反馈速率
FastTune[66]2021INT速率+窗口
HierCC[49]2021网内反馈速率
LDCP[13]2021ECN窗
RCC[22]2021RTT窗口
PowerTCP[51]2022INT速率+窗口
RET[21]2022RTT速率
PACC[52]2022网内反馈速率
Bolt[67]2023INT窗口
ACC[14]2024ECN速率
), ArticleFig(id=1325493754332852637, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Table 2, caption=ECN Fields and Their Meanings as Defined by RFC 3168, figureFileSmall=null, figureFileBig=null, tableContent=
ECN字段含义
ECT位CE位
00Not-ECT:不支持ECN标记
01ECT字段:支持ECN标记
10ECT字段:支持ECN标记
11CE:经历拥塞
), ArticleFig(id=1325493754412544414, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=表2, caption=

RFC 3168定义的ECN字段及其含义

, figureFileSmall=null, figureFileBig=null, tableContent=
ECN字段含义
ECT位CE位
00Not-ECT:不支持ECN标记
01ECT字段:支持ECN标记
10ECT字段:支持ECN标记
11CE:经历拥塞
), ArticleFig(id=1325493756090266015, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Table 3, caption=Port State Transition Process Defined by TCD Algorithm, figureFileSmall=null, figureFileBig=null, tableContent=
判断条件状态转换
Ton < max(Ton)LAST_STATE == 拥塞拥塞→未确定
LAST_STATE == 未拥塞未拥塞→未确定
LAST_STATE == 未确定未确定→未确定
Ton > max(Ton)LAST_STATE == 拥塞QueueLength > threshold拥塞→拥塞
QueueLength < threshold拥塞→未拥塞
LAST_STATE == 未拥塞QueueLength > threshold未拥塞→拥塞
QueueLength < threshold未拥塞→未拥塞
LAST_STATE == 未确定QueueLength > threshold队列未减小未确定→拥塞
队列减小未确定→未拥塞
QueueLength < threshold未确定→未拥塞
), ArticleFig(id=1325493756178346400, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=表3, caption=

TCD算法定义的端口状态转换过程

, figureFileSmall=null, figureFileBig=null, tableContent=
判断条件状态转换
Ton < max(Ton)LAST_STATE == 拥塞拥塞→未确定
LAST_STATE == 未拥塞未拥塞→未确定
LAST_STATE == 未确定未确定→未确定
Ton > max(Ton)LAST_STATE == 拥塞QueueLength > threshold拥塞→拥塞
QueueLength < threshold拥塞→未拥塞
LAST_STATE == 未拥塞QueueLength > threshold未拥塞→拥塞
QueueLength < threshold未拥塞→未拥塞
LAST_STATE == 未确定QueueLength > threshold队列未减小未确定→拥塞
队列减小未确定→未拥塞
QueueLength < threshold未确定→未拥塞
), ArticleFig(id=1325493756270621089, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Table 4, caption=ECN Fields and Their Meanings Defined by TCD, figureFileSmall=null, figureFileBig=null, tableContent=
ECN字段含义
ECT位CE位
00不具备TCD传输功能
01具备TCD传输功能
10未确定状态
11经历拥塞
), ArticleFig(id=1325493756354507170, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=表4, caption=

TCD定义的ECN字段及其含义

, figureFileSmall=null, figureFileBig=null, tableContent=
ECN字段含义
ECT位CE位
00不具备TCD传输功能
01具备TCD传输功能
10未确定状态
11经历拥塞
), ArticleFig(id=1325493756434198947, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Table 5, caption=Header Information of HPCC, figureFileSmall=null, figureFileBig=null, tableContent=
HPCC标头字段字段大小/b字段含义
nHop4路径上的交换机跳数
pathID12路径上所有交换机ID异或
B4出端口的速率(枚举值)
TS24出端口发出的时间戳
txBytes20出端口发送的累计字节
qlen16出端口当前队列长度
), ArticleFig(id=1325493756526473636, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=表5, caption=

HPCC的标头信息

, figureFileSmall=null, figureFileBig=null, tableContent=
HPCC标头字段字段大小/b字段含义
nHop4路径上的交换机跳数
pathID12路径上所有交换机ID异或
B4出端口的速率(枚举值)
TS24出端口发出的时间戳
txBytes20出端口发送的累计字节
qlen16出端口当前队列长度
), ArticleFig(id=1325493756597776805, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Table 6, caption=Header Information of Bolt, figureFileSmall=null, figureFileBig=null, tableContent=
Bolt标头字段字段大小/b字段含义
q_size24入端口的队列大小
link_rate8入端口的链路速率
data1普通数据包
ack1ACK数据包
src1SRC数据包
last1属于最后一个发送窗口
first1属于第一个发送窗口
inc1发送方应增大窗口
dec1发送方应减小窗口
reserved1保留字段
t_data_tx32原始数据包的发送时间戳
), ArticleFig(id=1325493756669079974, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=表6, caption=

Bolt的标头信息

, figureFileSmall=null, figureFileBig=null, tableContent=
Bolt标头字段字段大小/b字段含义
q_size24入端口的队列大小
link_rate8入端口的链路速率
data1普通数据包
ack1ACK数据包
src1SRC数据包
last1属于最后一个发送窗口
first1属于第一个发送窗口
inc1发送方应增大窗口
dec1发送方应减小窗口
reserved1保留字段
t_data_tx32原始数据包的发送时间戳
), ArticleFig(id=1325493756748771751, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Table 7, caption=Variables and Parameters Adjusted by DCQCN Related Rates, figureFileSmall=null, figureFileBig=null, tableContent=
名称含义
变量RT目标速率,用于快速恢复
RC当前速率
α拥塞程度
T记录timer被触发的次数
BC记录bytes_counter被触发的次数
参数RAI速率增量,用于主动增加
RHAI超速率增量,用于超主动增加
timer相邻速率提升事件之间的最小间隔
bytes_counter相邻速率提升事件之间的最小字节数量
F控制3个速率提升阶段的跳转
), ArticleFig(id=1325493756820074920, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=表7, caption=

DCQCN相关速率调整的变量与参数

, figureFileSmall=null, figureFileBig=null, tableContent=
名称含义
变量RT目标速率,用于快速恢复
RC当前速率
α拥塞程度
T记录timer被触发的次数
BC记录bytes_counter被触发的次数
参数RAI速率增量,用于主动增加
RHAI超速率增量,用于超主动增加
timer相邻速率提升事件之间的最小间隔
bytes_counter相邻速率提升事件之间的最小字节数量
F控制3个速率提升阶段的跳转
), ArticleFig(id=1325493756891378089, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Table 8, caption=Description of Each Phases of RTT in ACC, figureFileSmall=null, figureFileBig=null, tableContent=
RTT阶段数据发
送速率
数据包
发送总量
ACK数据包
接收速率
ACK数据包
接收总量
注:Null表示无效。
RTT1RSΔSNullNull
RTT2RSΔSRΔA
RTT3(开始暂停)NullNullRΔS×$ \dfrac{R}{{R}_{\mathrm{S}}} $
RTTn(结束暂停)RΔANullNull
), ArticleFig(id=1325493756987847082, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=表8, caption=

ACC各RTT阶段的说明

, figureFileSmall=null, figureFileBig=null, tableContent=
RTT阶段数据发
送速率
数据包
发送总量
ACK数据包
接收速率
ACK数据包
接收总量
注:Null表示无效。
RTT1RSΔSNullNull
RTT2RSΔSRΔA
RTT3(开始暂停)NullNullRΔS×$ \dfrac{R}{{R}_{\mathrm{S}}} $
RTTn(结束暂停)RΔANullNull
), ArticleFig(id=1325493757054955947, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Table 9, caption=Explanation of Variables and Parameters in RET Algorithm, figureFileSmall=null, figureFileBig=null, tableContent=
分类具体名称含义
变量RttDiff加权后的当前延迟梯度
Delay当前延迟
DelayPrev上阶段延迟
RttDiffPrev上阶段加权后的延迟梯度
QDelay排队延迟
QTarget目标延迟
NumberAcked延迟ACK响应的数据包数量
cwnd窗口大小
参数BaseRTT基本RTT
α加权延迟梯度中新旧延迟值的权重
βg延迟梯度MD因子
βr目标延迟MD因子
Mdft最大MD因子
δr加性增量
), ArticleFig(id=1325493757134647724, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=表9, caption=

RET算法的变量和参数说明

, figureFileSmall=null, figureFileBig=null, tableContent=
分类具体名称含义
变量RttDiff加权后的当前延迟梯度
Delay当前延迟
DelayPrev上阶段延迟
RttDiffPrev上阶段加权后的延迟梯度
QDelay排队延迟
QTarget目标延迟
NumberAcked延迟ACK响应的数据包数量
cwnd窗口大小
参数BaseRTT基本RTT
α加权延迟梯度中新旧延迟值的权重
βg延迟梯度MD因子
βr目标延迟MD因子
Mdft最大MD因子
δr加性增量
), ArticleFig(id=1325493757214339501, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Table 10, caption=codepoints Defined by CDEM and Their Meanings, figureFileSmall=null, figureFileBig=null, tableContent=
ECN字段含义
ECT位CE位
00不具备ECN传输功能
01未确定数据包
10未标记数据包
11经历拥塞
), ArticleFig(id=1325493757289836974, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=表10, caption=

CDEM定义的codepoints及其含义

, figureFileSmall=null, figureFileBig=null, tableContent=
ECN字段含义
ECT位CE位
00不具备ECN传输功能
01未确定数据包
10未标记数据包
11经历拥塞
), ArticleFig(id=1325493757382111663, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=EN, label=Table 11, caption=Connection of RL with Congestion Control, figureFileSmall=null, figureFileBig=null, tableContent=
强化学习的实体与函数拥塞控制的实体与函数
代理发送方
环境交换机或接收方
策略函数拥塞调整函数
价值函数网络性能整体表现
奖励函数网络性能变化
观测函数拥塞感知函数
), ArticleFig(id=1325493757453414832, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, language=CN, label=表11, caption=

强化学习与拥塞控制的联系

, figureFileSmall=null, figureFileBig=null, tableContent=
强化学习的实体与函数拥塞控制的实体与函数
代理发送方
环境交换机或接收方
策略函数拥塞调整函数
价值函数网络性能整体表现
奖励函数网络性能变化
观测函数拥塞感知函数
)], attaches=null, journal=Journal(id=1324759606947315908, delFlag=0, nameCn=计算机研究与发展, nameEn=Journal of Computer Research and Development, nameHistory1=null, nameHistory2=null, issn=1000-1239, eissn=null, cn=11-1777/TP, coden=null, periodic=0, language=CN, oaType=null, ccby=null, superviseOffice=null, ownerOffice=null, pubOffice=null, editorOffice=null, officeType=null, aims=null, clcCode=null, officeProv=null, officeCity=null, officeAddr=null, officeZip=null, officeEmail=null, officePhone=null, editDirector=null, officeDirector=null, officeDirectorPhone=null, officeStaffNum=null, officeEmpNum=null, coverPicUrl=k2l8zgqwstuxxevFX2qxVA==, journalPrice=null, startedYear=null, abbrevIsoEn=Journal of Computer Research and Development, journalRemark=null, publicationField=null, createdTime=1793776903141, updatedTime=1793850698612, createdBy=18614031015, updatedBy=13701087609, firstLetterCn=J, firstLetterEn=J, subjectCode=Electronic Information, subjectName=null, subjectCodeEn=Electronic Information, subjectNameEn=null, picCn=k2l8zgqwstuxxevFX2qxVA==, picEn=r6k0Ry+RD+c/VXj5QarmDw==, jcr=null, cjcr=null, exts=[JournalExt(id=1325069127710286688, language=CN, name=计算机研究与发展, nameHistory1=null, nameHistory2=null, managedBy=, sponsoredBy=, publishedBy=, editorOffice=, officeProv=null, officeCity=null, officeAddr=, officeZip=, editDirector=, officeDirector=null, officePhone=null, coverPicUrl=null, journalRemark=, submitArticleUrl=null, websiteUrl=, createdTime=1793850698640, updatedTime=1793850698640, createdBy=13701087609, updatedBy=13701087609, submissionGuidelinesUrl=, submissionAuthorUrl=https://crad.ict.ac.cn/journalx/authorLogOn.action, submissionEditorUrl=https://crad.ict.ac.cn/journalx/editorLogOn.action, submissionReviewUrl=https://crad.ict.ac.cn/journalx/expertLogOn.action, submissionCeEditorUrl=https://crad.ict.ac.cn/journalx/editorInChiefLogOn.action, submissionAeEditorUrl=, option={"copyright":""}), JournalExt(id=1325069127777395553, language=EN, name=Journal of Computer Research and Development, nameHistory1=null, nameHistory2=null, managedBy=, sponsoredBy=, publishedBy=, editorOffice=, officeProv=null, officeCity=null, officeAddr=, officeZip=, editDirector=, officeDirector=null, officePhone=null, coverPicUrl=null, journalRemark=, submitArticleUrl=null, websiteUrl=, createdTime=1793850698656, updatedTime=1793850698656, createdBy=13701087609, updatedBy=13701087609, submissionGuidelinesUrl=, submissionAuthorUrl=https://crad.ict.ac.cn/journalx/authorLogOn.action, submissionEditorUrl=https://crad.ict.ac.cn/journalx/editorLogOn.action, submissionReviewUrl=https://crad.ict.ac.cn/journalx/expertLogOn.action, submissionCeEditorUrl=https://crad.ict.ac.cn/journalx/editorInChiefLogOn.action, submissionAeEditorUrl=, option={"copyright":""})], databaseList=null, tenantJournalId=1324759914146529478, websiteList=[Website(id=1325070300139892771, webName=null, webTitle=null, webDomain=null, webCopyrigh=null, webIpcNo=null, seoTitle=null, seoKeywords=null, seoDescription=null, tenantJournalId=null, journalId=1324759914146529478, journalNameCn=null, journalNameEn=null, grayFlag=null, tenantId=1146029695717560320, platformId=null, journalGroupId=null, journalGroupNameCn=null, journalGroupNameEn=null, type=1, domain=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN, language=CN, createTime=1793850978169, createBy=18614031015, updateTime=1793851040259, updateBy=18614031015, name=计算机研究与发展-中文, tplId=1146099689490845704, title=计算机研究与发展, delFlag=0, indexPage=/home, props=[WebsiteProps(id=1325070893751349360, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=articleTextType, value=kx, createTime=1793851119697, updateTime=1793851119697, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893730377837, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=banner, value=null, createTime=1793851119692, updateTime=1793851119692, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893772320883, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=grayFlag, value=0, createTime=1793851119702, updateTime=1793851119702, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893721989228, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=logo, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/file/pic?fileId=U27tDsy717YF7XSPA+Vwdw==, createTime=1793851119690, updateTime=1793851119690, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893784903797, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=minRunFlag, value=0, createTime=1793851119705, updateTime=1793851119705, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893742960751, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=picServerUrl, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/file/pic, createTime=1793851119695, updateTime=1793851119695, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893780709492, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=silenceFlag, value=0, createTime=1793851119704, updateTime=1793851119704, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893738766446, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=staticResourcePath, value=https://castjournals.cast.org.cn/joweb/cast_kjdb_cn_619/, createTime=1793851119694, updateTime=1793851119694, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893755543665, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=themeColor, value=null, createTime=1793851119698, updateTime=1793851119698, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893763932274, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=themeStyle, value=null, createTime=1793851119700, updateTime=1793851119700, creator=18614031015, updator=18614031015)]), Website(id=1325070300580294698, webName=null, webTitle=null, webDomain=null, webCopyrigh=null, webIpcNo=null, seoTitle=null, seoKeywords=null, seoDescription=null, tenantJournalId=null, journalId=1324759914146529478, journalNameCn=null, journalNameEn=null, grayFlag=null, tenantId=1146029695717560320, platformId=null, journalGroupId=null, journalGroupNameCn=null, journalGroupNameEn=null, type=1, domain=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN, language=EN, createTime=1793850978274, createBy=18614031015, updateTime=1793851036638, updateBy=18614031015, name=计算机研究与发展-英文, tplId=1146101810881728533, title=Journal of Computer Research and Development, delFlag=0, indexPage=/home, props=[WebsiteProps(id=1325070925288321148, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=articleTextType, value=kx, createTime=1793851127216, updateTime=1793851127216, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925263155321, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=banner, value=null, createTime=1793851127210, updateTime=1793851127210, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925309292671, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=grayFlag, value=0, createTime=1793851127221, updateTime=1793851127221, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925254766712, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=logo, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/file/pic?fileId=U27tDsy717YF7XSPA+Vwdw==, createTime=1793851127208, updateTime=1793851127208, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925326069889, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=minRunFlag, value=0, createTime=1793851127225, updateTime=1793851127225, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925279932539, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=picServerUrl, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/file/pic, createTime=1793851127214, updateTime=1793851127214, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925317681280, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=silenceFlag, value=0, createTime=1793851127223, updateTime=1793851127223, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925271543930, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=staticResourcePath, value=https://castjournals.cast.org.cn/joweb/cast_kjdb_en_623/, createTime=1793851127212, updateTime=1793851127212, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925292515453, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=themeColor, value=null, createTime=1793851127217, updateTime=1793851127217, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925300904062, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=themeStyle, value=null, createTime=1793851127219, updateTime=1793851127219, creator=18614031015, updator=18614031015)])], journalTitle=计算机研究与发展, weixinUrl=null, journalUrl=https://crad.ict.ac.cn/, iacademicId=null, status=1, seqNo=null, journalTitleEn=Journal of Computer Research and Development, journalPhotoCn=k2l8zgqwstuxxevFX2qxVA==, journalPhotoEn=r6k0Ry+RD+c/VXj5QarmDw==, journalFirstLetter=J, journalRecommend=null, journalNew=null, journalCollection=null, jcrJf=null, cjcrJf=null, jcrJfStr=null, cjcrJfStr=null, submissionFirstDecision=null, sciSubjectClassification=null, casSubjectClassification=null, citeScore=null, totalCitationFrequency=null, icpCode=null, psCode=null, advertisingLicenseCode=null, copyrightInformation=null, country=null, option=, provinceCode=null, provinceName=null, collectFlag=false, interPubPlatform=, interPubPlatformUrl=null), detailUrlCn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/10.7544/issn1000-1239.202440623, detailUrlEn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/10.7544/issn1000-1239.202440623, pdfUrlCn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/PDF/10.7544/issn1000-1239.202440623, pdfUrlEn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/PDF/10.7544/issn1000-1239.202440623, aliStartDate=0, aliEndDate=0, collectionFlag=false, citedCount=null, citedUrl=null, previewStatus=0, delFlag=0, hasFullText=1, orderTime=1769875200000, fullTextJson=null, articleText=null, reference=null)
收藏切换
数据中心网络RDMA拥塞控制技术综述
张毓涛 , 杨惠 , 李韬 * , 黄曼蒂 , 李天云 , 孙志刚
计算机研究与发展 | 网络与信息安全 2026,63(2): 477-504
收起
收藏切换
计算机研究与发展 |网络与信息安全 2026 , 63 (2) : 477 -504
数据中心网络RDMA拥塞控制技术综述
全屏
[Author(id=1325493744815976770, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=0, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=zhang_yt@nudt.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1325493744887279940, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493744815976770, language=EN, stringName=Yutao Zhang, firstName=Yutao, middleName=null, lastName=Zhang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 2000. Master. His main research interests include RDMA and data center congestion control

"}, bioImg=null, bioContent=

born in 2000. Master. His main research interests include RDMA and data center congestion control

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493744954388805, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493744815976770, language=CN, stringName=张毓涛, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=国防科技大学计算机学院 长沙 410073, bio={"img":"MRsYvq4PBHe8RA7p4dYueA==","content":"

2000年生。硕士。主要研究方向为RDMA、数据中心拥塞控制

"}, bioImg=MRsYvq4PBHe8RA7p4dYueA==, bioContent=

2000年生。硕士。主要研究方向为RDMA、数据中心拥塞控制

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])]), Author(id=1325493745017303367, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=1, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1325493745088606537, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493745017303367, language=EN, stringName=Hui Yang, firstName=Hui, middleName=null, lastName=Yang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1986. PhD, associate professor. Her main research interests include RDMA and network processor

"}, bioImg=null, bioContent=

born in 1986. PhD, associate professor. Her main research interests include RDMA and network processor

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493745164104010, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493745017303367, language=CN, stringName=杨惠, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=国防科技大学计算机学院 长沙 410073, bio={"img":"AwoySCYfASLKXOQILp89Ag==","content":"

1986年生。博士,副研究员。主要研究方向为RDMA、网络处理器

"}, bioImg=AwoySCYfASLKXOQILp89Ag==, bioContent=

1986年生。博士,副研究员。主要研究方向为RDMA、网络处理器

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])]), Author(id=1325493745256378700, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=2, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=litao@nudt.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=1, authorType=1, ext={EN=AuthorExt(id=1325493745336070478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493745256378700, language=EN, stringName=Tao Li, firstName=Tao, middleName=null, lastName=Li, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=*, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1983. PhD, associate professor. His main research interest includes high-performance network system and chip

"}, bioImg=null, bioContent=

born in 1983. PhD, associate professor. His main research interest includes high-performance network system and chip

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493745407373647, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493745256378700, language=CN, stringName=李韬, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=*, address=国防科技大学计算机学院 长沙 410073, bio={"img":"/gCcIXfqnULENXvqBdI1tA==","content":"

1983年生。博士,副研究员。主要研究方向为高性能网络系统及芯片

"}, bioImg=/gCcIXfqnULENXvqBdI1tA==, bioContent=

1983年生。博士,副研究员。主要研究方向为高性能网络系统及芯片

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])]), Author(id=1325493747080900945, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=3, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1325493747185758547, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747080900945, language=EN, stringName=Mandi Huang, firstName=Mandi, middleName=null, lastName=Huang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1999. Master. Her main research interest includes RDMA

"}, bioImg=null, bioContent=

born in 1999. Master. Her main research interest includes RDMA

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493747273838932, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747080900945, language=CN, stringName=黄曼蒂, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=国防科技大学计算机学院 长沙 410073, bio={"img":"q1ji+p0ZWOExy19IlxdlrA==","content":"

1999年生。硕士。主要研究方向为RDMA

"}, bioImg=q1ji+p0ZWOExy19IlxdlrA==, bioContent=

1999年生。硕士。主要研究方向为RDMA

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])]), Author(id=1325493747345142102, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=4, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1325493747437416792, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747345142102, language=EN, stringName=Tianyun Li, firstName=Tianyun, middleName=null, lastName=Li, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 2000. Master. Her main research interests include data center network, RDMA congestion control, and artificial intelligence

"}, bioImg=null, bioContent=

born in 2000. Master. Her main research interests include data center network, RDMA congestion control, and artificial intelligence

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493747508719961, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747345142102, language=CN, stringName=李天云, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=国防科技大学计算机学院 长沙 410073, bio={"img":"E7Joy2X7kyqZc7wPzZtJEw==","content":"

2000年生。硕士。主要研究方向为数据中心网络、RDMA拥塞控制、人工智能

"}, bioImg=E7Joy2X7kyqZc7wPzZtJEw==, bioContent=

2000年生。硕士。主要研究方向为数据中心网络、RDMA拥塞控制、人工智能

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])]), Author(id=1325493747588411739, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, orderNo=5, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1325493747659714909, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747588411739, language=EN, stringName=Zhigang Sun, firstName=Zhigang, middleName=null, lastName=Sun, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1973. PhD, professor. His main research interests include software-defined network, time-sensitive networking, network architecture, and network security

"}, bioImg=null, bioContent=

born in 1973. PhD, professor. His main research interests include software-defined network, time-sensitive networking, network architecture, and network security

, aboutCorrespAuthor=null), CN=AuthorExt(id=1325493747735212382, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, authorId=1325493747588411739, language=CN, stringName=孙志刚, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=null, address=国防科技大学计算机学院 长沙 410073, bio={"img":"G4TR5c7IqQaR0gG+AqH14w==","content":"

1973年生。博士,研究员。主要研究方向为软件定义网络、时间敏感网络、网络体系结构、网络安全

"}, bioImg=G4TR5c7IqQaR0gG+AqH14w==, bioContent=

1973年生。博士,研究员。主要研究方向为软件定义网络、时间敏感网络、网络体系结构、网络安全

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1325493744618844478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, xref=null, ext=[AuthorCompanyExt(id=1325493744627233087, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1325493744635621696, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1325493736079241471, companyId=1325493744618844478, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=国防科技大学计算机学院 长沙 410073)])])]
张毓涛 , 杨惠, 李韬* , 黄曼蒂, 李天云, 孙志刚
作者信息
  • 国防科技大学计算机学院 长沙 410073
通讯作者:
李韬()
作者简介:

2000年生。硕士。主要研究方向为RDMA、数据中心拥塞控制

1986年生。博士,副研究员。主要研究方向为RDMA、网络处理器

1983年生。博士,副研究员。主要研究方向为高性能网络系统及芯片

1999年生。硕士。主要研究方向为RDMA

2000年生。硕士。主要研究方向为数据中心网络、RDMA拥塞控制、人工智能

1973年生。博士,研究员。主要研究方向为软件定义网络、时间敏感网络、网络体系结构、网络安全

Survey of RDMA Congestion Control Techniques for Data Center Networks
Yutao Zhang , Hui Yang, Tao Li* , Mandi Huang, Tianyun Li, Zhigang Sun
Affiliations
  • College of Computer Science and Technology, National University of Defense Technology, Changsha 410073
  • born in 2000. Master. His main research interests include RDMA and data center congestion control

    born in 1986. PhD, associate professor. Her main research interests include RDMA and network processor

    born in 1983. PhD, associate professor. His main research interest includes high-performance network system and chip

    born in 1999. Master. Her main research interest includes RDMA

    born in 2000. Master. Her main research interests include data center network, RDMA congestion control, and artificial intelligence

    born in 1973. PhD, professor. His main research interests include software-defined network, time-sensitive networking, network architecture, and network security

出版时间: 2026-02-01 doi: 10.7544/issn1000-1239.202440623
文章导航
收藏切换

拥塞控制是实现高性能数据中心网络的关键技术之一,影响吞吐量、延迟、丢包率等重要网络性能指标。过去20年间,随着数据中心规模不断扩大,上层应用对网络性能的要求不断提高,基于无损底层网络的远程直接内存访问(remote direct memory access,RDMA)技术在数据中心的部署受到了业内广泛关注。然而,基于优先级的流控(priority-based flow control,PFC)机制在维护无损网络的同时会引入头阻塞等问题,导致网络性能下降甚至网络瘫痪。作为实现无损网络的关键辅助手段,如何设计实用的RDMA拥塞控制机制成为了热点问题。通过将拥塞控制过程划分为拥塞感知与拥塞调整,全面综述了该领域的研究成果:首先从显式反馈与延迟的角度详细阐述并总结了不同的拥塞感知代表算法;其次从速率和窗口的维度对拥塞调整代表算法进行了详细介绍并对其优缺点进行了总结;而后补充了部分算法的优化工作以及基于强化学习方法的拥塞控制算法;最后总结并讨论了该领域存在的挑战。

数据中心网络  /  远程直接内存访问  /  RoCEv2  /  拥塞控制  /  拥塞感知  /  拥塞调整

Congestion control is one of the key technologies for realizing high-performance data center networks, and it affects important network performance indicators such as throughput, latency, and packet loss rate. Over the past 20 years, with the continuous expansion of the scale of data centers and the increasing requirements of upper-layer applications for network performance, the deployment of remote direct memory access (RDMA) technology based on lossless underlying networks has received widespread attention within the industry. However, the priority-based flow control (PFC) mechanism, while maintaining a lossless network, introduces problems such as head-of-line blocking, leading to a decline in network performance or even network paralysis. As a crucial auxiliary means for achieving a lossless network, how to design a practical RDMA congestion control mechanism has become a hot issue. By dividing the congestion control process into congestion awareness and congestion regulation, this paper comprehensively reviews the research achievements in this field: Firstly, from the perspectives of explicit feedback and latency, different representative algorithms for congestion awareness are elaborated and summarized in detail; Secondly, representative algorithms for congestion regulation are introduced in detail from the dimensions of rate and window, and their advantages and disadvantages are summarized; Some optimization work of algorithms and congestion control algorithms based on reinforcement learning methods are supplemented; Finally, the existing challenges in this field are summarized and discussed.

data center networks (DCN)  /  remote direct memory access (RDMA)  /  RDMA over converged Ethernet (RoCEv2)  /  congestion control (CC)  /  congestion awareness  /  congestion regulation
张毓涛, 杨惠, 李韬, 黄曼蒂, 李天云, 孙志刚. 数据中心网络RDMA拥塞控制技术综述. 计算机研究与发展, 2026 , 63 (2) : 477 -504 . DOI: 10.7544/issn1000-1239.202440623
Yutao Zhang, Hui Yang, Tao Li, Mandi Huang, Tianyun Li, Zhigang Sun. Survey of RDMA Congestion Control Techniques for Data Center Networks[J]. Journal of Computer Research and Development, 2026 , 63 (2) : 477 -504 . DOI: 10.7544/issn1000-1239.202440623
近20年来,数据中心在全球范围内蓬勃发展。随着数据中心从最初的计算中心阶段发展到如今的算力中心阶段,网络规模不断扩大,对网络品质的要求不断提高[1]。尽管目前传输控制协议/网际协议(transmission control protocol/internet protocol,TCP/IP)协议栈在数据中心仍占统治地位[2-3],但是该协议栈的弊端逐渐显露,业内正积极尝试将以远程直接内存访问(remote direct memory access,RDMA)技术为代表的新技术部署于数据中心[2,4]。
RDMA实现对硬件有较强的依赖,有限的网卡片上资源使得RDMA重传采取了资源占用与消耗较小的回退n步(go-back-n,GBN)重传策略[5-7],即使在丢包率极低的情况下,该策略也会导致性能的急剧下降[8],因此RDMA的部署需要数据中心提供一个底层无损网络[2,8-14],该网络不会因内部问题发生丢包,除非链路受到外界干扰或网络存在错误配置。
数据中心利用基于优先级的流量控制(priority-based flow control,PFC)[15-16]增强的聚合以太网(converged Ethernet,CE)来实现无损网络,然而逐跳流量控制方法会带来一系列问题,例如头阻塞、PFC暂停帧风暴、死锁等[2-3,8,10-14,17-22]。为了降低触发PFC的频率,需要对RDMA拥塞控制进行精细地设计与研究[8,17,21]。
当前国内外的综述文章缺少对近年来RDMA拥塞控制机制的详细分析,例如:Huang等人[23]将无损数据中心网络的拥塞控制方案划分为主动式拥塞控制和响应式拥塞控制,对早期的RDMA拥塞控制算法进行了介绍;Hu等人[5]将数据中心RDMA传输设计划分为单路径拥塞控制、多路径传输和其他解决方案,其中单路径拥塞控制方案对部分主流算法进行了简单的阐述;曾高雄等人[24]将RDMA拥塞控制作为一种发展趋势简要地引出;杜鑫乐等人[25]则将RDMA拥塞控制、流调度和负载均衡作为数据中心流量控制的组成部分,剖析了部分拥塞控制算法;Yan等人[26]将RDMA拥塞控制算法按照发送方驱动、交换机驱动和接收方驱动3个维度划分,其侧重点与本文并不相同。本文将尝试较为全面、系统地对当前的数据中心RDMA拥塞控制机制进行分析与总结。同时,考虑到一些早期的TCP拥塞控制算法启发了RDMA的相关工作,本文尝试在RDMA拥塞控制算法之后补充相关的TCP拥塞控制算法。
近20年来,数据中心相继经历了计算中心阶段、信息中心阶段、云中心阶段和算力中心阶段。随着云计算、大数据、人工智能、物联网等产业的加速发展,数据中心的规模将进一步扩大。据统计,我国的数据中心市场规模在2021年为1 800亿元,在2022年达到了2 200亿元[1]。可以预见的是,数据中心规模在当前及未来一段时间内将继续保持增长趋势。
伴随着规模的不断扩大,数据中心应用程序对网络性能的要求也随之提高。例如,网络搜索、在线零售和广告推荐等软实时应用对网络延迟十分敏感[13,21,27-40];云存储、云互联和云安全等云服务业务需要网络提供高带宽的同时不能产生过高的CPU开销[4,8-9,20,36,41-45];分布式机器学习、分布式数据库和分布式计算等分布式应用要求网络同时提供高带宽和低延迟的特性[3,8-9,14,22,41-42,46-52]。尽管TCP/IP协议栈在数据中心网络仍占据统治地位,但显然已无法满足上述要求[3,8,11,13],主要原因在于以操作系统为中心的TCP/IP协议栈执行一次完整的数据收发流程需要进行数据复制、系统调用、中断处理等操作,导致了CPU开销大、延迟高,并对吞吐量等其他性能产生影响。
具有内核旁路功能的RDMA是一项高度依赖硬件实现的高性能数据传输技术,数据收发的大部分操作流程绕过了操作系统内核,转而由硬件执行。双端的CPU仅在控制通路参与,在数据通路中,网卡能够直接访问远端内存,避免了数据在用户空间与内核空间的多次复制,报文的封装与解封操作均由网卡硬件完成,无需经过内核协议栈的逐层处理。得益于零拷贝技术、操作系统内核旁路和CPU卸载技术,RDMA能够同时满足高吞吐量、低延迟以及低CPU开销的目标。图1展示了以内核为中心的传统TCP通信方式和内核旁路的RDMA通信方式。
RDMA最初在无线带宽(infiniband,IB)架构上以IB协议的形式实现[53],该架构要求传输层、网络层和数据链路层均使用IB协议,尽管能够提供较优的性能,但是使用IB协议进行组网通常需要定制的IB交换机和IB网卡,这与数据中心普遍使用的以太网并不兼容且成本昂贵。互联网广域远程直接内存访问协议(internet wide area RDMA protocol,iWARP)[54]同样能够实现RDMA数据传输。该协议在以太网数据链路层、TCP/IP实现的网络层基础上,将传输层由上至下划分为3层子协议。尽管该技术可以利用以太网交换机和iWARP网卡实现,但是TCP/IP协议栈本身对延迟和吞吐量并不友好,因此相较于IB以及基于融合以太网的远程直接内存访问(RDMA over converged Ethernet,RoCE)而言,iWARP的性能表现并不出众。
RoCE协议当前有2个版本,如图2所示。RoCEv1[55]的架构与IB架构的不同之处在于前者利用以太网数据链路层代替了后者的链路层,因此RoCEv1的报文结构与IB报文相比添加了一个2层以太网报文头,这使得RoCEv1只能部署在2层以太网之上;RoCEv2[56]除了使用以太网链路层外,其网络层基于UDP/IP协议实现,因此RoCEv2的报文结构增加了UDP、IP和2层以太网报文头,可部署在3层以太网之上。RoCEv2协议实现了基于以太网基础设施的组网方案,通常仅需RoCEv2网卡的支持。综上,RoCEv2因其成本低廉、性能优越且兼容于以太网的特性,已成为数据中心高性能数据传输的主要选择。
拥塞这一概念反映的是网络状况:在分组交换网络中,当数据包数量超过了存储转发节点的处理能力时,网络性能将大幅下降。尽管此时的带宽利用率仍保持在较高水平,但延迟将急剧增大,丢包的可能性也将急剧升高。解决这一问题的根本途径是调整数据包发送速率或发送数量以缓解网络设备的处理压力。
RoCEv2拥塞控制算法的实现主要依赖于网卡设计,然而网卡有限的硬件资源难以支持RoCEv2实现高效且复杂的重传策略[3,5-7,22,52]。GBN策略的低效性使RoCEv2对丢包较为敏感,因此需要一个无损网络以实现高性能数据传输。数据中心网络利用PFC协议来实现这一目标[3,15]。
如图3所示,为了向上层用户提供更好的服务质量(quality of service,QoS),交换机或网卡的入/出端口缓冲区被划分为8个优先级队列,同一优先级队列可能存在多条数据流[15]。PFC规定,当交换机或网卡的入口缓冲区某队列达到缓冲区阈值时,向上游网络设备发送PFC暂停帧(PFC PAUSE frame),使数据源暂停发送,避免因持续发送数据包造成网络丢包。然而PFC策略也会带来一些严重的性能问题,例如头阻塞(head-of-line,HOL)、死锁、PFC暂停帧风暴等[2-3,8,10-14,17-22,52]。造成上述性能问题的一个重要原因是当入端口队列中的数据包数量超过阈值时,PFC向上游端口队列泛洪PFC暂停帧,导致拥塞树扩散或错误循环依赖。
图4举例说明了incast流量模式引起的PFC受害流问题。1号流和2号流为长流/大象流,其源和汇分别为SH1,RH1和SH2,RH2,假设二者被映射到P3和P4的同一优先级队列中。源为SH3~SHn,汇为RH2的短流/老鼠流同时启动并造成P6拥塞。在此场景下,有2个原因导致受害流问题:1)PFC暂停帧以队列为单位反压。端口P4的队列因下游端口拥塞而逐渐增大,当其大小达到PFC阈值时,向数据源(SH1和SH2)发送PFC暂停帧,导致数据源暂停发送数据包。2)端到端拥塞控制协议无法识别拥塞流。端口P3收到下游的PFC暂停帧直到数据源停止发送,期间P3的队列仍会增长,一旦队列大小超过标记阈值,1号流的数据包在传输重启后同样会被标记。综上,尽管1号流未对P6端口发生的拥塞情况有任何影响,但是因为与2号流在同一队列中排队而成为受害流。
一个有效的拥塞控制机制在一定程度上有利于避免触发PFC策略,在实现无损网络的同时尽可能规避上述严重的性能问题[2-3,8-9,12,14,17,21,50,57]。近年来,随着技术与工艺的发展,网络设备实现了更多的功能,例如显式拥塞通知(explicit congestion notification,ECN)、高精度时间戳测量往返时延(round-trip time,RTT)和带内网络遥测(in-band network telemetry,INT)等[58-63],结合P4交换机与FPGA可编程网卡的可编程性,拥塞控制机制在此基础上得到了较大的创新。
拥塞控制过程可以视为拥塞感知和拥塞调整的组合。拥塞感知为拥塞程度的度量提供了直接途径,是端到端拥塞控制的基石[12,64],其精确性以及感知反馈的速度将从根本上决定一个拥塞控制算法的优劣。为了兼顾吞吐量、延迟、稳定性、公平性以及收敛速度等重要指标,拥塞调整的过程需要参考网络的拥塞程度。表1按相关论文发表时间展示了部分数据中心RDMA拥塞控制算法。
根据拥塞感知信号来源的不同,本节将拥塞感知划分为基于显式反馈的拥塞感知和基于延迟的拥塞感知。其中,基于显式反馈的拥塞感知根据当前主流算法的显式反馈信号,可再进一步细分为ECN、INT以及网内反馈。本节先对上述感知信号的采集原理进行说明,再描述相关主流算法,最后对各类型的优缺点进行总结。
显式反馈信号在拥塞感知中得到了广泛的应用,按照端侧获得的信息,可以将显式反馈划分为ECN标记、INT标记和网内反馈。本节将尝试对上述显式反馈原理以及基于各种反馈信息的感知算法进行详细阐释。
1993年Floyd等人[68]在TON期刊上提出了RED(random early detection)方法。RED通过比较队列的平均长度与阈值区间来决定对数据包采取的动作,如图5所示,当队列的平均大小在最小阈值Kmin以下时,RED不会标记数据包;高于最大阈值Kmax时,RED将以1的概率标记数据包;若在Kmin和Kmax之间,则以一个关于平均队列大小的概率函数对数据包进行标记。与传统的队列管理方法相比,RED能够实现较短的队列长度,允许网络存在一定的突发,避免流的全局同步现象。
2001年由Ramakrishnan等人[69]撰写的RFC 3168标准将RED-ECN方法纳入到TCP/IP网络中。作为一种主动队列管理(active queue management,AQM)机制,RED为交换机ECN功能的实现提供了途径:结合上述RED方法,交换机对IP报文头中的ECN字段进行标记,将网络中的拥塞情况通告给接收方。ECN字段及其含义如表2所示。
若数据包携带ECN使能传输(ECN-capable transport,ECT)标记,则交换机在拥塞时会将该字段标记为经历拥塞(congestion experienced,CE);若数据包携带非ECN使能传输(Not-ECN-capable transport,Not-ECT)标记,则不会对数据包进行标记。ECT和Not-ECT均由发送方指定。
微软、Nvidia Mellanox公司和加利福尼亚大学圣芭芭拉分校的研究人员在2015年SIGCOMM会议上提出的DCQCN算法[8]是利用ECN标记进行拥塞感知的经典算法,该算法使得数据中心大规模部署RDMA成为了可能。它参考了QCN[16,70]和DCTCP[27]的相关工作。
以太网标准QCN[16,70]是一个作用于网络第2层的拥塞控制算法,由2个子算法组成:拥塞点(congestion point,CP)算法和响应点(reaction point,RP)算法。其中CP是指发生拥塞的网络设备,一般是指数据中心的交换机,也可以是接收方主机或网卡;RP是指调整拥塞的网络设备,一般是指发送方主机或网卡。图6展示了QCN的拥塞控制机制。
CP算法支持对交换机出口端口队列大小进行周期性采样以计算当前的拥塞程度Fb,根据Fb的正负决定是否向某抽样数据包的源返回拥塞通知消息(congestion notification message,CNM),数据包抽样的概率与|Fb|有关。RP算法将根据CNM以及维护的定时器和字节计数器来对发送方的发送速率进行调整。
DCTCP算法由Alizadeh等人[27]在2010年的SIGCOMM会议上提出,是数据中心网络的早期经典拥塞控制算法之一。DCTCP是一种增强TCP对ECN反应的算法,TCP窗口的减小程度取决于ECN标记数据包的比例[29],尽管不支持RDMA在数据中心的部署,但其关键思想仍然值得借鉴。
DCTCP算法的拥塞感知部分对RED-ECN方法进行了修改:将Kmin与Kmax设为相同值K,并以瞬时队列大小采样代替平均队列大小,这样做的好处在于部署简单且发送方反应迅速。具体来说,若瞬时队列长度大于阈值K,则数据包将被标记为CE,否则不对数据包做任何处理。接收方为每一个数据包返回一个ACK数据包(per-packet ACK),若收到携带CE标记的数据包,则在对应的ACK数据包头部字段标记ECN响应(ECN-echo,ECE)。
DCQCN的拥塞感知借鉴了DCTCP的做法,拥塞调整则同时借鉴了QCN和DCTCP的做法。DCQCN的拥塞感知包括2个子算法:CP算法和通知点(notification point,NP)算法,NP是指将拥塞感知信号返回给RP的网络设备,可以是交换机,也可以是接收方。CP算法定义为当交换机某出端口队列的瞬时大小超过阈值K时,交换机将对ECT(0)或ECT(1)状态的数据包以概率1进行标记。NP算法则指定接收方网卡在N微秒内最多生成1个拥塞通知数据包(congestion notification packet,CNP),即N微秒内仅对1个带有CE标记的数据包做出反应。为了避免CNP在反向路径上排队,DCQCN将其设置为高优先级传输。
在2020年的NSDI会议上,清华大学等高校与研究机构提出了PCN算法[10]。相较于以前的工作,该算法提出:1)定义交换机的出端口为真拥塞、无拥塞和准拥塞3种状态,在此基础上提出了新的CP算法。2)将拥塞感知细分为拥塞检测与拥塞识别,其中拥塞检测阶段进行拥塞信息的采集,由CP算法实现;拥塞识别阶段利用上述信息对真正的拥塞流进行判断,由NP算法实现。
PCN将真拥塞定义为缓冲区队列积累了大量数据包且出口链路过载的状态;无拥塞则是缓冲区队列无数据包积累,出口链路正常工作的状态;准拥塞端口的缓冲区队列有数据包积累,出口链路由于PFC暂停帧和PFC重启帧(PFC RESUME frame)的交替作用而处于间歇性过载。以图4为例,P6为真拥塞状态,P5为无拥塞状态,P3因为受到P4发送的PFC暂停帧和PFC重启帧的交替影响而处于准拥塞状态。
PCN定义的RP算法负责调整发送速率,CP算法负责标记数据包,NP算法负责判断真正的拥塞流、计算接收速率以及返回CNP。本节仅对为拥塞感知阶段设计的CP算法和NP算法进行介绍。
CP算法实现了名为NP-ECN(none-PAUSE ECN)的标记方法。该方法根据出端口是否处于PFC暂停帧抑制状态将数据包分为非暂停状态数据包和暂停状态数据包,前者即为真拥塞端口、无拥塞端口中的数据包以及PFC重启帧作用下进入准拥塞端口中的数据包;后者则是PFC暂停帧作用下在准拥塞端口中排队的数据包,包含了在PFC暂停帧作用期间由于其生效的滞后性导致的新入队数据包以及原本就处于队列中的数据包。对非暂停状态数据包,当交换机出口队列大于0时,以1的概率标记,否则不标记;对于暂停状态数据包,交换机维护了一个计数器,参数PN表示受PFC暂停帧作用而排队的数据包数量。当端口收到PFC重启帧并再次发送数据包时,PN随着暂停数据包的发送递减,PCN不对其标记;当PN递减为0后,PCN采用与非暂停状态数据包一致的标记方法。PCN的字段定义与ECN相同。
NP算法逐流记录了N微秒内收到的数据包总数recNum以及带ECN标记的数据包总数ECNNum。前者随数据包递增,后者随携带CE标记的数据包递增,通过计算$ \dfrac{1}{recNum}{ECNNum} $是否大于0.95来做出拥塞判断:若大于0.95,则将对应数据流判断为拥塞流,标记CNP报文IP头部ECN字段的CE位为1;否则即为未拥塞流,此时的CE位标记为0。
PCN证明了当数据流发生拥塞时,其最佳发送速率为其接收速率,即接收方在每秒内接收的数据量大小。为了计算接收速率,PCN接收方除了维护recNum以及ECNNum外,还会逐流记录N微秒内收到的数据包总大小recData,该变量随数据包大小pSize递增。对于拥塞流,其接收速率为$ \dfrac{1}{N}{recData} $,对于非拥塞流,其接收速率为0。
不同于DCQCN只在接收到CE数据包时才返回CNP的做法,PCN在N微秒结束时无论是否收到CE数据包均返回CNP、CNP数据包格式如图7所示,发送方通过CNP携带的接收速率对其发送速率进行快速而准确的调整,有效地避免了多轮启发式迭代过程。
清华大学和北京国家信息科学与技术研究中心在2021年的SIGCOMM会议上提出的TCD标记算法[12]同样注重拥塞流的准确识别。与PCN类似,TCD指出正是无损网络所依赖的PFC机制混淆了真正拥塞的数据流。为了能够区分拥塞流与受害流,TCD同样定义了3种端口状态:拥塞状态、未拥塞状态和未确定状态。其中未确定状态是指端口处于ON-OFF模式(即受PFC暂停帧和PFC重启帧交替作用的模式)且队列中的数据包排队原因并不明确的状态(即无法确定排队原因是拥塞或PFC暂停帧的作用)。
TCD认为拥塞感知的关键是正确判断3种状态的端口之间的相互转换,特别是未确定状态向其他状态的转换。在每个数据包从出端口离开队列时,TCD都会计算ON模式的持续时间Ton并将其与max(Ton)进行比较,结合队列长度QueueLength的演变特征来决定3种端口状态之间的转换,表3解释了转换过程。其中,ON模式是指受PFC重启帧作用的模式,LAST_STATE用于记录上一个数据包离队时的端口状态,QueueLength表示队列长度,threshold是队列长度阈值。
TCD根据端口状态来对数据包进行标记,若端口处于拥塞状态,则对其中的数据包进行CE标记;若端口处于未拥塞状态,则不对数据包进行标记;若端口处于未确定状态,则对其中的数据包进行UE标记。CE标记可以覆盖UE标记,而UE标记则无法覆盖CE标记。TCD认为网络中通过拥塞端口的数据流均是拥塞流,而仅通过未确定端口的数据流则有可能是受害流。TCD对字段定义的修改如表4所示。
华为公司称在2021年提出的互联网草案LDCP[13]与RoCEv2协议已在华为公有云安全运行了1年时间。由于能够维持恒定的低队列长度,几乎不发生丢包,因此该草案放弃了作为“第2道防线”的PFC机制。
LDCP依据数据包所处的不同阶段被划分为:首RTT阶段的数据包(first-RTT-packets),即在第1个RTT内发送的数据包;稳定阶段的数据包(stable-stage-packets),即第1个RTT时间之后发送的数据包。前者的ECN字段为00,LDCP赋予其低优先级,并规定在队列大小超过阈值K时主动将其丢弃;后者的ECN字段为10或01,LDCP赋予其高优先级并采用与DCQCN相同的标记方法。按阶段划分数据包的主要目的是通过first-RTT-packets为数据流的启动探测合适的窗口,避免窗口过大造成网络拥塞或窗口过小造成带宽利用率过低。
接收方网卡对稳定状态数据包有2种反应方式:第1种是per-packet ACK,即为每一个数据包生成并返回ACK数据包;第2种是延迟ACK(delay ACK),即最多为m个数据包生成并返回一个ACK数据包,但是如果第n个数据包带有CE标记,则会立即发送ACK数据包(n≤m)。对未标记数据包而言,其ACK数据包中ECE标记为0;对标记数据包而言,ECE标记为1。发送方根据ACK数据包中ECE标记的不同值采取不同的拥塞调整策略。
除了上述RDMA拥塞控制算法外,数据中心早期大量面向TCP流量模式的拥塞控制算法也利用ECN进行拥塞感知。例如:Vamanan等人[35]在2012年的SIGCOMM会议上提出的D2TCP算法在DCTCP的基础上引入流量截止时间(deadline)作为除ECN标记之外的拥塞信号,二者协同计算得到表示网络拥塞程度与流量截止时间紧迫性的值。Zats等人[30]也在2012年的SIGCOMM会议上提出了Detail。Detail是跨层网络堆栈,该方案为交换机的每一个出口队列按优先级维护了一个耗尽字节计数器,该计数器的值表示新到达的数据包前面的同等或更高优先级的字节计数。当计数器的值超过其关联阈值时,交换机将对排队的低优先级数据包进行ECN标记。同年,Alizadeh等人[29]在NSDI会议上提出的HULL架构利用串联在交换机队列之后的幽灵队列(phantom queue,PQ)来模拟链路占用率,PQ利用基于链路利用率的ECN标记方法提前发出拥塞信号,通过牺牲一部分吞吐量达到显著降低平均延迟和尾部延迟的效果。Wu等人[71]在2012年的CoNEXT会议上提出的ECN*算法则借鉴了DCTCP的交换机ECN标记方法和标准ECN的发送方窗口调整方法,并指出了ECN*阈值的上下限,实现了与DCTCP同样的性能。Stephens等人[17]于2014年在INFOCOM会议上提出的TCP-Bolt利用与DCTCP相同的ECN标记方法,抛弃了慢启动过程,有效降低了流量完成时间(flow completion time,FCT)。
相比于丢包等传统的拥塞信号,基于ECN的拥塞感知能够将队列大小维持在一个较低的水平,有利于实现低延迟和较高的吞吐量;相比于RTT,由于ECN标记在出口队列处发生,因此能够分离控制环路延迟与排队延迟,使其免受逐渐增大的队列的影响,有利于保持稳定[57]。同时,ECN也具有以下局限性:
1)ECN标记是一种粗粒度的拥塞信号,它仅能反映网络是否存在拥塞,并不能说明拥塞的严重程度[3,6,9,20-21,50,65-66]。这导致该类算法需采用启发式的拥塞调整,即经过多轮次的调整才能收敛到合适的速率,不仅会消耗较长的时间,也无法解决突发传输问题。另外该类算法还需要设置大量的参数用于适配不同的运行环境,参数的调整往往是一个复杂且困难的工作[9,66,71]。
2)ECN标记阈值的恰当与否对性能的影响很大[21,27,36,71-72]:阈值过低有助于保持较短队列并减小延迟,但吞吐量会受到影响;阈值过高有利于提高吞吐量,但会损害延迟。在动态环境下,固定的阈值可能无法适应环境的变化[72]。
3)PFC触发可能会“污染”队列大小,从而影响ECN标记的正确性,即无法识别真正的受害流[10,12]。近年来研究人员已经关注到了该问题并在拥塞控制层面提出了解决方法,例如PCN算法和TCD算法。此外,还有工作通过抛弃PFC机制[7]或替换PFC机制[73]同样能够缓解该问题。
INT是一种基于网络数据平面对网络状态进行监控的框架[74],其被划分为3种应用模式:INT-XD(export data),INT-MX(embed instruct(x)ion),INT-MD(embed data)。由于当前的RDMA拥塞控制解决方案大多采用INT-MD模式实现,因此本文仅对INT-MD模式的拥塞信息采集原理进行阐释。
如图8所示,在INT-MD模式下,发送方(INT Source)往数据包嵌入INT标头;交换机(INT Transit)逐跳向数据包嵌入元数据;交换机(INT Sink)剥离INT标头,汇总元数据并将其发送回INT Source。其中,INT Source是在数据包中创建并插入INT标头的实体,INT Transit是按照INT标头中的指令向数据平面收集元数据的实体,INT Sink则是提取INT标头和相关元数据的实体。
用户可以通过INT指令来指定需要收集哪些INT元数据,常见的元数据有节点ID、入端口标识符、入口时间戳、出端口标识符、出口时间戳、跳转延迟、出端口TX链路利用率、队列占用率和缓冲区占用率等。相较于ECN方法而言,INT能够反馈的信息更加丰富。
阿里巴巴公司、哈佛大学、剑桥大学和麻省理工学院在2019年的SIGCOMM会议上提出的HPCC算法[9]指出ECN标记或RTT无法提供精细的拥塞信息,导致性能之间的权衡,因此HPCC主张利用INT机制获取精确的链路信息,实现精准的拥塞控制。HPCC利用INT-MD模式感知拥塞的过程如下:INT Source将数据包发送到网络中,INT Transit逐跳向数据包中插入元数据,最后由INT Sink汇总所有元数据并将其插入ACK数据包中,INT Source根据ACK数据包中携带的详细信息进行拥塞调整。
HPCC在具有P4可编程功能的交换机ASIC上实现了INT功能,需测量网络状态主要有:1)链路带宽容量B(4 b),表示当前出端口的速率(40 Gbps, 100 Gbps等),是一个枚举类型字段。2)时间戳TS(24 b),表示数据包离开当前出端口的时间。3)传输字节txBytes(20 b),表示从当前出端口发送的累计字节(128 B),4)队列长度qlen(16 b),表示当前出端口的队列长度(80 B)。由于数据中心规模有限, HPCC规定使用40 B的数据包标头对至多5跳的元数据进行记录。HPCC的标头信息如表5所示。
国防科技大学在2021年提出的FastTune算法[66]关注了数据中心流量在时间和空间维度上的动态变化。通过INT机制感知细粒度拥塞信号以解决数据流在空间维度上动态变化的问题,FastTune采集的信息与HPCC相同。通过在第1个RTT内由交换机生成并返回携带INT元数据的ACK数据包(S-ACK机制),第1个RTT之后由交换机向来自接收方的ACK数据包插入INT元数据(H-ACK机制)来解决时间维度动态变化的问题。
S-ACK机制的目的是使发送方主机在第1个RTT内感知到网络拥塞。如图9所示,当数据流第1个数据包到达交换机时,交换机生成一个S-ACK数据包,并在出口处将INT元数据插入到S-ACK数据包并返回发送方。交换机维护一个名为NewACK的变量,该变量在交换机产生S-ACK数据包或转发H-ACK数据包后被设置为False,表示交换机不再为后续到来的数据包做出响应。
H-ACK机制的目的则是确保发送方获得最新的拥塞信息。其做法是,接收方主机生成H-ACK数据包并返回,交换机采集反向链路入端口,即正向链路出端口处的拥塞信息,并把INT元数据插入H-ACK数据包中。值得注意的是,实现上述机制需要ACK数据包与其对应的数据包路径互相对称,同时需要交换机支持跨端口采集信息。
维也纳大学、柏林工业大学和普林斯顿大学在2022年的NSDI会议上发表了POWERTCP算法[51],该算法将网络类比为电路,网络的绝对属性类比为电压(例如队列大小、RTT、ECN等),网络的变化属性类比为电流(例如队列大小变化、RTT梯度等)。POWERTCP证明了单纯基于电压的方案具有唯一收敛点,稳定性较好,但是处理抖动的能力较弱、收敛速度慢;与此相反,单纯基于电流的方案收敛较快,抖动较小,但是不存在唯一的收敛点。因此,POWERTCP提出基于功率的方案,功率定义为电压与电流之积,从网络的角度看,该方案旨在同时对网络的绝对属性和变化属性做出响应,从而获取二者各自的优点。
在POWERTCP方案中,电压的定义是瓶颈链路的带宽延迟积(bandwidth delay product,BDP)与队列字节数之和,电流的定义是瓶颈链路的总发送速率与队列变化率之和,POWERTCP证明在该定义下,功率等于瓶颈链路带宽与总发送窗口的乘积,即发送窗口与功率之间存在直接联系,因此可以通过计算功率来调节发送窗口。
POWERTCP利用与HPCC相同的INT-MD机制来为功率计算提供必要的链路信息,采集的元数据以及整体流程与HPCC相同,即二者的拥塞感知相同。在第3节将阐述POWERTCP如何有效利用上述元数据来设计拥塞调整策略。
斯坦福大学和谷歌公司的研究人员在2023年的NSDI会议上提出的Bolt算法[67]关注到了3个问题,分别是控制环路过长的问题、未完成数据流无法及时抢占完成数据流释放的带宽导致的吞吐量降低问题以及每个RTT仅进行1次的窗口调整导致的收敛缓慢问题。本节通过在交换机入端口处生成并返回带INT信息的次RTT控制(Sub-RTT control,SRC)数据包来解决第1个问题;通过基于信令的主动速率提升(proactive ramp up,PRU)以及带宽供应匹配(supply matching,SM)来解决第2个问题。
PRU的目的是帮助其余流量快速抢占由完成流量释放的带宽:last字段为“1”的数据包表示其所属流即将结束,当交换机收到该类型数据包时,将递增pru_token值;SM的目的是帮助流量充分利用链路剩余带宽,交换机会记录相邻数据包之间的间隔,供应量为带宽与间隔的乘积,需求量为数据包大小,sm_token的增量即为二者的差值。表6展示了Bolt需要维护的标头信息。
图10为Bolt定义的交换机处理逻辑:收到数据包后交换机入端口首先计算供应令牌sm_token,其次判断当前队列大小cur_q_size与阈值CCTHRESH大小,若超过了阈值且pkt.dec标记为0,则对数据包的q_size,link_rate,t_data_tx字段进行更新(对应图10的①),否则不进行任何操作。若未超过阈值且pkt.last=1,当first字段为0时,将pru_token加1(对应图10的②),否则也不进行任何操作。最后,当pkt.inc=1且pru_token>0,sm_token>MTU,则保持pkt.inc字段不变,分别消耗一个pru_token和sm_token(对应图10的③),否则将pkt.inc字段置为0。
综上所述,Bolt利用控制环路较短的SRC数据包进行窗口递减,利用控制环路较长的ACK数据包进行窗口递增。
相比于ECN方法,INT能够提供更加丰富的信息,具有以下优势:1)相较于多轮迭代的启发式算法,基于INT的算法有利于发送速率或发送窗口快速收敛,避免反应过度或反应不足[67];2)减少了参数的数量,提高了算法在不同环境下的适应能力[67];3)可以利用INT信号将正反向路径上的拥塞信息分离,避免了反向路径延迟可能带来的影响[41]。
INT也有以下不足之处:1)依赖硬件支持[3,9,67,75],有可能导致因交换机对INT的支持不同而使网络异构;2)在不修改终端主机与交换机的情况下可能与RoCEv2不兼容[3];2)向数据包插入元数据会导致交换机花费更长的时间处理数据包[75];3)数据包携带大量的元数据信息会对网络造成更大的负担[75];4)不同的路径使得数据包携带的元数据数量不同,导致数据包头大小不同,这与当前某些处理固定大小数据包头的主机堆栈互相矛盾[75]。
针对INT机制开销过大的问题,近年来学术界提出了概率网络带内遥测(probabilistic in-band network telemetry,PINT)机制,这部分内容将在第4节介绍。
除了ECN标记和INT标记外,交换机主动计算的速率、主动分配的信令或其他信号同样可视为显式反馈的一种,本节对其中的部分方案进行简要介绍,并总结其优缺点。
针对发送方驱动的方案反馈延迟较长,无法快速获得拥塞信号的问题,2020年,由思科公司、普渡大学、提契诺大学在CoNEXT会议上提出的RoCC算法[11]把速率计算任务由端侧转移到了交换机,发送方直接利用交换机CNP携带的速率进行调整。
如图11所示,在RoCC设计中,交换机额外包括速率计算器、CNP生成器和流量表3个组件。速率计算器通过定期采样队列大小Qcur,结合最大队列大小Qmax、参考队列大小Qref等参数,对流速率进行计算。CNP生成器将速率计算结果封装成CNP报文,并通过查询流量表,将其发送回发送方。
RoCC的速率计算过程包括乘法递减(multiplicative decrease,MD)和比例积分(proportional integral,PI),相关变量及计算逻辑如图12所示。
北京邮电大学、华为公司和紫金山实验室在2021年的APNet会议上提出的HierCC算法[49]认为数据中心的流量不确定性是大多数拥塞控制算法无法同时满足高吞吐量和确定性延迟的根本原因,在此基础上,HierCC指出拥塞控制算法应具有聚合流量、短控制环和积极主动的速率分配3个特征。
HierCC的框架如图13所示,ToR(top of rack)交换机实现了虚拟输出队列(virtual output queues,VOQ),其数量与网络中主机的总数N相同;除了VOQ外,ToR交换机维护下游输出队列(downstream output queue,DOQ),其数量与该ToR交换机连接的主机数量k相同;ToR交换机内部实现了1个VOQ调度器和k个DOQ调度器,VOQ调度器用于向网络发送数据包,DOQ调度器负责向VOQ发送信用(credits)。
聚合流量是指将来自相同ToR交换机、去往相同接收方的数据包聚合到同一VOQ中排队,有利于缓解流量的不确定性;短控制环是指发送主机的速率由第1跳ToR交换机来控制;主动速率分配是指ToR交换机之间利用信令实现聚合流的速率控制,发送方的速率由第1跳ToR交换机的聚合流速率计算得到。综合上述3种机制,HierCC实现了更快速更准确的拥塞控制,有效降低了流量完成时间。
北京邮电大学、华为公司和紫金山实验室在2022年的INFOCOM会议上提出的PACC算法[52]认为控制环路过长以及启发式调整算法阻碍了数据中心RDMA的大规模部署,在借鉴RoCC算法的基础上,PACC格外关注部署问题。为此,PACC仅对交换机逻辑进行修改,而终端则与DCQCN兼容。
PACC的交换机算法主要由3个后台周期性进程和1个函数组成:首先是周期性流表更新,当交换机收到上游数据包时,根据当前队列长度Qcur和队列阈值Qth对流表的字段进行更新;其次是周期性CNP数量计算,基于PI控制器对所有拥塞流共享的CNP总数Nall进行计算;再次是周期性平均队列大小计算,通过对比平均队列大小Qavg和突发性队列阈值Qburst区分并处理突发情况;最后是基于拥塞程度的CNP分配与生成函数,通过3个后台进程得到的数据实现基于权重的CNP分配,达到精确控制拥塞的效果。图14展示了上述3个过程及1个函数的联系。
基于网内反馈的拥塞感知方法由交换机直接计算并反馈速率等信息,既缩短了控制环路,又能避免多轮启发式调整带来的诸多问题[3];然而其复杂性从端侧转移到交换机实现使得部署难度更大,成本更高[3];另外,交换机对速率等网内反馈的计算通常也是采取启发式算法,尽管不需要多轮迭代,但同样需要仔细设计参数以满足不同情况下的目标[51]。
作为一个简单且有效的信号,延迟在数据中心RDMA拥塞控制中得到了较好的应用。其简单性体现在延迟作为衡量网络性能的基本要素,几乎所有类型的网络都提供了测量条件,不需要交换机的支持;其有效性体现在目前的大部分网卡都支持高精度硬件时间戳功能,延迟测量达到了微秒级别[76-80],通过避免软件时间戳引入的额外耗时,延迟能够更加准确地反映网络状况[50]。目前大部分基于延迟的拥塞控制研究中,大多采用测量RTT的方式,因此下文将以RTT代替延迟进行论述。
谷歌公司于2015年在SIGCOMM会议上提出的TIMELY算法[65]是第1个利用RTT进行数据中心RDMA拥塞控制的经典算法,这主要得益于网卡支持高精度时间戳以及硬件生成ACK数据包,避免了因测量误差以及主机延迟带来的影响。
如图15所示,TIMELY以数据段为单位进行突发传输,一个数据段本质上是一个固定大小的字节序列,接收方网卡为一个数据段而非单个数据包返回一个ACK数据包。
TIMELY对RTT的定义不同于以数据包为传输单位产生的定义:发送方主机将数据段第1个字节的发送时刻记录为tsend,将对应于该数据段的ACK数据包的到达时间记录为tcompletion,二者都是被主机读取的硬件记录的时间。假设数据段大小为seg.size且该数据段以网卡线速率发送,RTT的定义如式(1)所示。
$ {RTT}_{\mathrm{T}\mathrm{I}\mathrm{M}\mathrm{E}\mathrm{L}\mathrm{Y}}={t}_{\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}-{t}_{\mathrm{s}\mathrm{e}\mathrm{n}\mathrm{d}}-\dfrac{seg.size}{\mathrm{N}\mathrm{I}\mathrm{C}\;\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{e}\;\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{e}} \;。 $
发送方根据ACK数据包携带的RTT信息来计算RTT变化梯度并基于梯度指导拥塞调整,拥塞调整部分将在第3节中进行介绍。
威斯康星大学麦迪逊分校和伊利诺伊大学芝加哥分校在2018年的SoCC会议上提出的RoGUE算法[20]意图通过重新设计软件实现的拥塞控制和丢包恢复机制来摆脱对PFC的依赖。其拥塞感知部分与TIMELY类似,传输单位都是数据段而非数据包,因此RTT的计算与式(1)相同,在此不予赘述。
2020年,谷歌公司在TIMELY的基础上提出Swift算法[50],该算法以RTT作为拥塞感知信号,在具有内核旁路功能的Pony Express协议栈中实现。与TIMELY相比:1)Swift的发送单位是数据包而不是数据段;2)Swift利用主机软件时间戳和网卡硬件时间戳对多个时刻进行记录;3)Swift不仅关注网络的拥塞情况,还格外关注主机内部的拥塞情况。
图16展示了Swift数据包收发过程以及划分的不同延迟组成部分:①是发送方网卡发送延迟,表示数据包在网卡中等待发送的时间,可忽略;②是正向链路网络延迟,表示正向链路上序列化、排队和传播的时间;③是接收方网卡接收延迟,表示数据包被处理前在网卡中的等待时间;④是处理延迟,表示协议栈处理数据包以及生成ACK数据包的时间;⑤是接收方网卡发送延迟,表示ACK数据包等待发送的时间;⑥是反向链路的网络延迟,正反链路通常不对称;⑦是发送方网卡接收延迟,表示ACK数据包被发送方协议栈处理前的等待时间。综上可知:②与⑥反映网络延迟、③与④反映远程主机延迟。
结合上述定义,Swift测量RTT及划分延迟的方法如图17所示,1)发送方主机记录数据包的发送时间t1(当网卡就绪时才将数据包交付网卡,即交付时刻等于发送时刻);2)接收方网卡记录数据包的到达时间t2;3)接收方主机记录数据包开始被主机堆栈处理的时间t3;4)接收方主机记录ACK数据包的发送时间t4(ACK数据包的交付时刻等于发送时刻);5)发送方网卡记录ACK数据包的到达时间t5;6)发送方主机记录ACK数据包开始被主机堆栈处理的时间t6。
RTT定义为t6$ - $t1,包括网络中的序列化、排队、传播时延,以及主机内的排队与处理时延。为了准确反映网络状况,Swift定义了端内延迟和网络延迟。端内延迟为接收方的排队及处理时间(t4$ - $t2)与本地排队时间(t6$ - $t5)之和,即t6+t4$ - $t5$ - $t2(接收方会把t4$ - $t2插入ACK数据包中返回);网络延迟为RTT减去端点延迟,即t5+t2$ - $t4$ - $t1。
北京邮电大学和紫金山实验室在2021年的ICNP会议上提出的RCC算法[22]利用单向延迟(single trip time,STT)信号将数据中心拥塞划分为最后1跳拥塞以及网内拥塞,并分别利用接收方显式窗口分配机制和比例积分微分(proportion integral differential,PID)窗口控制机制解决2种拥塞情况。RCC将拥塞感知细分为拥塞判断机制和拥塞划分机制,前者利用STT对是否发生拥塞进行判断,后者利用STT内收到的总字节数对拥塞类别进行划分。
1)拥塞判断机制。RCC在每个数据包头部插入发送时间戳,接收方根据该时间戳计算得到STT。如式(2)所示,若连续n个STT均大于基本单向延迟(baseSTT)的1+δ倍,则可判断为发生拥塞。
$ {STT}_{i} > {(1+\delta )\times baseSTT}_{i}\;。 $
2)拥塞划分机制。接收方将每一轮STT内收到的总字节数记录为B,若B大于最后1跳的单程带宽延迟积的η倍(η表示预期链路利用率),说明最后1跳带宽被充分利用,此时的拥塞类型为最后1跳拥塞,否则为网内拥塞。
南京大学于2022年在ICNP会议上提出的RET算法[21]证明了固定的目标延迟不利于延迟和吞吐量的博弈以及队列稳定性,因此RET尝试利用交换机提供的ECN标记来估计队列大小(标记阈值为0),并实时更新目标延迟。RET的ECN标记并不代表网络存在拥塞,而是表明队列是否为空:若数据包被标记,说明队列不为空,即存在排队延迟;反之则不存在排队延迟。
通过实时更新目标延迟,RET将存在排队延迟的数据包(即被标记数据包)比例维持在η附近,根据经验,当η=0.99时,目标延迟达到最佳状态。RET的接收方算法如图18所示,收到带ECN标记的数据包表示路径存在排队情况,此时需要降低目标延迟,其中δt表示更新步长;收到未携带ECN标记的数据包表示路径不存在排队情况,此时需要提高目标延迟,其中ConN_ECN用于限制提高次数,避免过度反应。
面向互联网的TCP拥塞控制方案在发展早期便把延迟应用到拥塞感知中,由于互联网规模相较数据中心而言较大,延迟较高,因此对时间戳的精度要求较低,且能够容忍端系统内部因中断、调度等机制带来的额外延迟影响。相较而言,数据中心因规模小、高带宽、流量动态变化等特点对时间测量的精度和终端的处理速度具有很高的要求。鉴于互联网的TCP拥塞控制机制在一定程度上对数据中心的方案有所启发,本节在下面对早期的互联网和数据中心部分基于延迟的TCP解决方案进行简单介绍。
Brakmo等人[81]于1994年提出的TCP Vegas算法是最早将延迟引入互联网TCP拥塞控制的方案,通过粗粒度测量RTT,结合baseRTT与当前窗口,可以分别计算得到实际吞吐量与期望吞吐量,并根据二者的差值决定如何修改发送窗口。Wei等人[82]于2006年提出了FAST TCP算法,该算法指出互联网中丢包等二进制拥塞信号表示拥塞程度的能力有限,通常会导致拥塞窗口和瓶颈队列振荡,即稳定性较差。而利用多比特的拥塞信号则有利于实现高利用率、公平性和稳定性等目标。该算法同时利用丢包与RTT来调整发送窗口。Tan等人[83]于2006年在INFOCOM会议上提出的CTCP算法专门为高速长距网络设计,在基于丢包的标准TCP拥塞控制算法中添加测量延迟的组件,借鉴TCP vegas的思想,根据测量得到的延迟计算预期吞吐量与实际吞吐量的差值,不仅实现了带宽的有效利用,还能够实现RTT公平性和TCP公平性。
相对于需要交换机支持的显式反馈信号,RTT优势体现在:1)RTT的测量并不需要交换机的支持,对于算法的部署而言,该优势所带来的好处十分明显[50,64-65];2)RTT能够直接反映链路上所有交换机中数据包因排队带来的延迟[64],即RTT能够直接反映拥塞情况,相比之下,ECN等显式反馈信号可能会因为 QoS等机制的影响无法反映真实的拥塞情况[65];3)利用多比特信息对RTT进行编码不仅能够表示拥塞的存在,还可以以较细的粒度反映拥塞严重程度[50,64-65];4)在RTT的基础上计算得到的RTT梯度能够较好地体现拥塞的发展趋势,基于RTT梯度的拥塞控制有利于快速收敛且抖动较小[51,57];5)RTT有助于区分网络拥塞和主机拥塞[11,50]。
RTT也有5个缺陷:1)尽管RTT能够反映网络和主机的拥塞程度,但是与INT等信号相比,其丰富性并不足以驱动发送方将速率或窗口直接调整到目标值,因此基于RTT的拥塞调整大多也是采用加性增大乘性减小(addictive increase and multiplicative decrease,AIMD)或乘性增大乘性减小(multiplicative increase and multiplicative decrease,MIMD)启发式算法[3,50,64-65]。2)RTT通常会将反向链路的延迟包含在内,对发送方而言反向链路延迟是无关紧要的,甚至会带来一定的影响,因此需要采取额外的补救措施,例如赋予ACK数据包更高的传输优先级[50]。这并不是RTT特有的缺点,所有端到端的解决方案都会面临该问题。3)基于RTT的拥塞感知依赖于RTT的测量精度,这通常会对主机网络协议栈和网卡硬件功能提出额外的要求,需要同步主机时间与硬件时间。4)利用RTT进行拥塞调整的算法通常需要与RTT阈值进行比较,与ECN标记阈值类似,恰当的RTT阈值选择通常是一项困难的工作[65]。5)简单地基于RTT绝对值与RTT阈值的差值来对拥塞程度进行评估不利于算法适应抖动较大的网络环境,收敛速率较慢;而简单地基于RTT梯度进行拥塞感知则可能导致算法不存在单一且固定的收敛点[51,57]。
拥塞感知为拥塞调整提供了依据,拥塞调整则直接决定了拥塞控制算法的效果。缓解拥塞的方法通常有3种:1)降低发送方的发送速率;2)减小发送方的发送窗口;3)同时降低发送速率和减小发送窗口。本节将根据上述3种划分方式阐述现有的拥塞控制算法,并分析各自的优缺点。
为了实现高吞吐、低延迟和快速的公平收敛性,DCQCN[8]的RP算法借鉴了DCTCP和QCN的思路,设计了一个基于参数α的“折半”降速过程和3个逐渐激进的提速过程,并维护了表7中的变量和参数用于控制速率调整。其中:参数timer表示若在该参数规定时间内未收到CNP,则触发一次速率提升事件;bytes_counter表示当发送字节达到该参数规定的数量且未收到CNP时触发一次速率提升。DCQCN将速率提升分为3个阶段,分别是快速恢复、主动增加和超增量增加。参数F和变量T,BC用于控制不同阶段之间的跳转。
如图19所示,①当收到CNP时,DCQCN进入速率下降阶段,分别对RT,RC,α进行计算。其中,g是0~1之间的恒定值,用于对α进行指数加权移动平均(exponential weighted moving average,EWMA)运算。②当timer或bytes_counter被触发且T,BC均小于F,则进入快速恢复阶段,根据保存的RT值将当前速率快速恢复到速率下降前的水平。③timer或bytes_counter被触发且F位于T与BC区间内,此时处于主动增加阶段,对速率进行缓慢提升。④timer或bytes_counter被触发且T,BC均大于F,此时超主动增加以快速抢占可用带宽,其中i表示触发此次超主动增加的T值或BC值。
DCQCN通过建立流体模型,理论上证明了数据包的实际标记概率接近于0,这说明在Kmin附近存在固定队列长度点,即DCQCN能够有效地将队列长度控制在较低水平;此外,DCQCN存在所有流量获得公平带宽的固定点,即DCQCN能够实现公平性。DCQCN重点对吞吐量进行研究,实验结果表明DCQCN与PFC相结合的方案能够有效提高拥塞情况下的吞吐量;另外,由于将队列长度维持在较低水平,DCQCN也能降低通信延迟。
TIMELY[65]除了关注高带宽和低延迟的性能表现,还格外强调RTT在数据中心拥塞控制领域的价值,利用低时间阈值Tlow和高时间阈值Thigh将拥塞的调整划分为3个阶段,分别是加性增大(addictive increase,AI)阶段、基于RTT梯度的增大/减小阶段和MD阶段。过程如图20所示。
当数据段传输事件完成并收到来自接收方的ACK数据包时,发送方利用新测量的RTT值(new_rtt)减去旧的RTT值(prev_rtt)得到RTT差值(new_rtt_diff),并对new_rtt_diff进行EWMA处理,随后除以最小RTT(minRTT)得到归一化RTT梯度(normalized_gradient)。其中,minRTT是指理想网络环境下的RTT值。
发送方通过对比new_rtt与Tlow和Thigh的大小来决定进入不同的速率调整阶段。若new_rtt<Tlow,则进入AI阶段,其中δ表示增量常数;若new_rtt>Thigh,则进入MD阶段,其中β表示MD因子。若new_rtt在Tlow和Thigh之间,则进入基于RTT梯度的增大/减小阶段。在该阶段中,若normalized_gradient<0,说明此时队列具有减小的趋势,进入AI阶段;反之,若normalized_gradient>0,说明队列具有增大的趋势并进入MD阶段。
TIMELY证明了与单一且固定的目标延迟法相比,RTT梯度方法能够同时实现高吞吐量与低延迟目标,且队列大小的波动更为平缓。实验结果表明,在数百台服务器组成的Clos网络拓扑中,TIMELY的平均RTT和尾延迟相较于DCTCP算法分别降低了1000%和1300%,且并未观察到数据包丢失和PFC触发的情况。
PCN算法[10]的核心是提高算法在突发流量下的表现,尽可能避免触发PFC机制,为此PCN设计了新颖的拥塞识别和速率计算策略并将其转移到接收方执行,因此RP算法比较简洁:在收到CNP后,发送方首先检查其ECN字段的CE位。若CE为1,此时发送速率以式(3)(4)进行调整。其中权重参数w用于控制PCN速率增长的激进程度。在发生拥塞时,w应设置为最小值wmin,代表此后的一段时间内的速率以温和的方式增长,该做法有利于发生拥塞后对速率的更新不会触发PFC机制。
$ w={w}_{\mathrm{m}\mathrm{i}\mathrm{n}}\;, $
$ sendRate=\mathrm{m}\mathrm{i}\mathrm{n}\left\{sendRate,recRate\times \left(1-w\right)\right\}\;。 $
若CE为0,发送速率以式(5)(6)进行调整。权重参数w随着非拥塞调整次数的增多而逐渐增大,表示速率调整由温和逐渐过渡到激进状态,该做法有利于发送方快速抢占可用带宽,避免资源浪费。
$ sendRate=sendRate\times (1-w)+lineRate\times w\;, $
$ w=w\times (1-w)+{w\times w}_{\mathrm{m}\mathrm{a}\mathrm{x}} \;。 $
得益于接收方对拥塞流的准确识别和接收方驱动的速率调整,PCN能够迅速稳定队列大小,为流量分配公平带宽并降低突发情况下的PFC触发概率。在由3对服务器和1台交换机组成的哑铃拓扑环境中,PCN在抑制PFC暂停消息方面比DCQCN和TIMELY分别减少了47%和90%;在快速收敛方面则比DCQCN和TIMELY快25倍和45倍;在降低延迟方面表现最佳,最多可相应减少1130%和1320%的延迟。
作为交换机驱动的拥塞控制算法,RoCC算法[11]通过快速接收并利用拥塞信号,其设计目标更为广泛,除高带宽和低延迟外还能够满足公平性、稳定性和快速收敛的特性。
RoCC的发送方并不需要进行复杂的操作。其拥塞调整主要包括2个部分:1)参考CNP携带的速率进行直接调整。当RoCC发送方收到来自交换机的CNP报文时,它首先检查当前CNP的源交换机是否与上一次进行速率调整参考CNP的源交换机相同,同时检查当前CNP携带的速率值是否小于当前的发送速率,若二者满足其一,则将发送速率替换为CNP携带的速率,该做法能够保证RoCC每次都对瓶颈链路做出响应。2)进行快速恢复操作。执行该操作的前提是长时间未收到满足上述条件的CNP,发送方判定网络状况良好并以指数级增长当前的发送速率,直到发送速率达到最大值。利用上述机制,RoCC能有效处理多瓶颈场景下的拥塞,同时有利于快速收敛和达到高带宽利用率。
在N个源节点通过一个交换机连接到单个目标节点的网络拓扑环境中,RoCC与DCQCN相比降低了多达700%的PFC帧生成,实现了比DCQCN低800%的尾部延迟,并且在公平性、稳定性和收敛性方面均优于DCQCN,TIMELY,HPCC。
清华大学等院校在2024年NSDI会议上发表的ACC算法[14]基于TCD的拥塞感知算法,针对无损网络拥塞控制提出2个关键设计原则:1)在无损网络中,ACK数据包序列和到达速率是对排队数据包的数量和发送速率进行推断的工具;2)对拥塞流进行时间充足的暂停是抑制头阻塞的首要手段,同时受害流应根据拥塞程度调整发送速率。在这2个设计原则指导下,ACC分别对拥塞流、受害流和未拥塞流采取不同的调整策略。
1)对拥塞流采取先暂停后启动的措施的关键在于暂停时间和启动速率的计算。如表8所示,为了计算暂停时间,ACC首先要知道RTT1和RTT2积累的未确认数据包数量,前者是RTT1发送的数据包总数ΔS与RTT2接收的ACK总数ΔA之差;后者是RTT2发送的数据包总数ΔS与RTT3即将接收的ACK总数$ \mathrm{\Delta }S\times \dfrac{R}{{R}_{\mathrm{S}}} $之差,其中R是ACK数据包接收速率,RS是普通数据包发送速率。
对于启动速率,ACC认为ACK数据包的到达速率在一定程度上可以反映接收方的接收速率,因此将暂停后的发送速率设置成暂停期间的ACK数据包到达速率。
2)对受害流采取自适应速率调整措施。ACC认为带UE标记的ACK数据包的持续时间可以反映拥塞程度。若拥塞不严重且不扩散,则ACC发送方将很少收到带UE标记的ACK数据包;若拥塞严重导致扩散,ACC发送方将在多个周期内都接收到带UE标记的ACK数据包。因此,ACC规定,若接收带UE标记的ACK数据包的持续周期数大于阈值Pthre,则将受害流的发送速率减半,否则不对其进行调整。
3)速率较小的未拥塞流应快速提高发送速率,速率较大的未拥塞流应缓慢提高发送速率。ACC通过式(7)来对未拥塞流的速率进行调整,其中lineRate是指网卡的线速率,RS是指当前的发送速率,RI是单周期内的最大增加量。在前2个周期内ACC将RI设置为RIlow,之后将其设置为RIhigh,从而实现从温和到激进的增长策略。
$ R=R+\dfrac{ lineRate-{R}_{\mathrm{S}} }{lineRate}\times RI\;。 $
由于能够准确识别拥塞流并对不同类型的流量施加不同的调整策略,ACC极大地减少了PFC暂停的数量,抑制了拥塞扩散,使得速率收敛速度和稳定性得到了一定的提高。ACC算法在包含了320台服务器、20台聚合交换机和16台核心交换机的fat-tree拓扑中进行了验证,结果表明,与HPCC相比,ACC降低了29%的小流量平均FCT以及降低了40%的99百分位FCT;与DCQCN和TIMELY相比能够降低79.2%和96.7%的死锁发生概率。
基于速率的拥塞调整算法可以由软件步调器或硬件步调器来实现,该方法具有以下优势:1)相比于基于窗口的策略,控制速率能够有效地避免数据突发,对数据包传输的控制更加精确[8,65];2)基于速率的方法比基于窗口的方法更加容易实现[8]。然而也有以下缺点:1)若反馈信号延迟,单纯地利用速率进行拥塞控制会导致控制环路期间发送过多的数据包,从而对链路造成负担[9,11];2)对数据包进行速率控制可能会造成主机内拥塞[29];3)相比于窗口方案,基于速率的调整需要消耗更多的资源[84]。
Swift[50]通过与目标延迟比较来调整窗口,目标延迟大小随着拓扑的变化和竞争流数量N的变化而扩展。与谷歌公司早期的工作相比,Swift不仅考虑网络的拥塞,还关注到了主机拥塞的影响。由于是基于绝对目标延迟进行拥塞调整,因此Swift在降低平均延迟和尾部延迟方面表现出色,算法稳定性也有所提高。
Swift利用网络拥塞窗口(fabric congestion window,FCWND)来控制网络拥塞,利用终端拥塞窗口(endpoint congestion window,ECWND)来控制主机内(包括发送方和接收方)的拥塞,二者利用同一个算法对窗口进行调整,前者的算法输入是网络延迟,后者的算法输入是端点延迟,实际发送窗口取二者较小值。
Swift认为,尽管可以设置一个足够大的目标延迟以保证全网拓扑中的所有流量达到较高的吞吐率,但是这样“不负责任”的做法会导致路径较短的流量形成更大的队列。因此,Swift根据拓扑和负载对目标延迟进行扩展,具体做法是:1)根据拓扑扩展。数据包头中的TTL(time-to-live)字段反映了流量的路径长短,可以将路径的长短转化为目标延迟的大小。2)根据负载扩展。Swift通过模拟实验获得竞争流量数量N与使瓶颈链路达到饱和所需的目标延迟之间的关系,结果表明,目标延迟的变化与$ \sqrt{N} $成正比。
Swift用于调整窗口的算法如下:若实际延迟高于目标延迟,则利用MD方法减小窗口,考虑到一个RTT周期内不能对同一个拥塞事件做出多次反应,该算法通过比较RTT与当前时间和上一次窗口减小时间的差值来决定是否执行MD。Swift利用式(8)对拥塞窗口(cwnd)进行调整。其中β是MD因子,max_mdf是最大MD因子。
$ \begin{split} &cwnd=\mathrm{m}\mathrm{a}\mathrm{x}\Bigg\{ 1- \\&\dfrac{\beta \times (delay-target\_delay)}{delay},1-max\_mdf \Bigg\} \times cwnd\;。\end{split} $
若实际延迟低于目标延迟,则进入AI阶段。Swift允许在一个RTT周期内对多个ACK做出反应,若cwnd>1,增量大小为ai/cwnd;若cwnd<1,增量ai/cwnd被修改为ai,用于快速恢复。若更新后的cwnd′<1,Swift将根据RTT值与窗口值计算数据包步进延迟,并利用速率控制组件在相邻数据包之间插入步进延迟来控制发送;若cwnd′>1,步进延迟为0,此时数据包的发送仅由窗口控制。如式(9)所示:
$ cwnd'=cwnd+\dfrac{ai}{cwnd}\times num\_acked\;。 $
变量num_acked表示接收方的延迟ACK数据包对num_acked个数据包进行反馈。
Swift在60台配备了50 Gbps网卡的主机组成的小型测试平台和500台配备了100 Gbps网卡的主机组成的大型测试平台中进行了实验,结果显示,Swift能够维持小于50 μs的99.9百分位RTT,比DCTCP低400%;在incast场景下实现了线速吞吐量、低延迟和几乎零丢包,其公平性表现也要优于DCTCP。
LDCP[13]的窗口调整算法PAWA(per-ACK window adjustment)对每一个ACK数据包做出反应,其关键见解是利用ACK数据包向发送方授予或撤销“信用”,模仿接收方驱动的拥塞控制。对应于2种不同的ACK数据包生成方式,其拥塞调整部分也可以被划分为2种不同的算法。
第1种调整算法对应于LDCP的per-packet ACK策略,即如果发送方接收到ECE字段为0的ACK数据包,它将判定此时网络状态良好,并以α/cwnd为单位增大发送窗口;如果收到ECE字段为1的ACK数据包,它将判定此时网络中存在拥塞,将以β为单位减小窗口。α和β都是0~1的参数。
第2种调整算法对应于LDCP的delay ACK策略,即此时的调整算法在上述基础上将增量或减量扩大了n倍,n表示接收方的延迟ACK数据包对n个数据包进行集中反馈。
考虑到突发流量导致的极度拥塞使得发送窗口小于1的情况:在ECE字段为0时,LDCP将增量设置为γ,γ表示最小窗口大小,通常为1/4,1/8,1/16等;在ECE字段为1时,窗口取γ与$ eta\times cwnd $之间的较大值,eta通常设置为1/2。
在32个节点的2层Clos网络中,LDCP的快速启动算法能够在不引起网络拥塞的情况下最大限度地利用空闲带宽,稳定阶段算法能够维持稳定的低队列大小,整体性能表现优于DCQCN。
RCC算法[22]利用接收方驱动的显式窗口分配机制缓解最后1跳拥塞,且利用PID窗口控制机制缓解网内拥塞。对于前者,RCC接收方根据数据包的基本传输报文头(base transport header,BTH)统计连接的数量N,并将每个连接的窗口大小设置为最后1跳带宽C与连接数量N之比;对于后者,RCC接收方利用PID控制器根据估计的目标STT(targetSTT)调整窗口。
如图21所示,PID控制过程可以分为2个阶段,分别是计算控制因子u和更新发送窗口cwnd。在第1阶段,误差e定义为测量得到的STT与targetSTT之差,该值同时作为比例项参与控制因子的计算,误差的导数$ {\mathrm{d}e}/{\mathrm{d}t} $则作为微分项参与计算,积分项为0;第2阶段则利用u值和tanh函数对发送窗口进行调整。
得益于RCC能够对最后1跳拥塞做出快速反应,同时对网内拥塞采取了参考比例项和微分项的PID控制机制,该算法能够实现快速收敛和较好的稳定性,有利于减小延迟和流量完成时间。
在320台主机、20台ToR交换机、20台聚合交换机和16台核心交换机构成的fat-tree拓扑中,RCC算法在网络搜索、数据挖掘、Web服务器和缓存跟随者4种流量负载下与HPCC,DCQCN,TIMELY等算法进行了仿真比较,结果表明,RCC在网络搜索工作负载下的平均FCT比上述3个算法分别低9%,30%,45%,在数据挖掘工作负载下的平均FCT则低了7%,18%,19%,在其余2类流量模式下的性能表现也整体优于HPCC等4个算法。
RET算法[21]借鉴了TIMELY的延迟梯度法和Swift的目标延迟法。通过对比延迟梯度和目标延迟的大小,RET不仅能够克服TIMELY不公平等缺点,还能在拥塞发生前做出反应。与TIMELY和Swift一样,RET同样利用AIMD对窗口进行调整。
表9对RET拥塞调整算法涉及的变量进行了解释,图22表示RET的窗口调整过程。在一个包含了288台主机、12台Leaf交换机和6台Spine交换机的Leaf-Spine拓扑结构中,RET在不到2 ms的时间内快速稳定了公平速率,显示出良好的公平性;在不同的流量竞争情况下,RET能够保持比RSWIFT(Swift算法的变体)短300%的队列长度;在500个并发流到达时,RET能够在不到2 ms的时间内排出约4 MB的积压数据,而仅依赖目标调整的RSWIFT需要的时间是RET的2倍;在其他工作负载下,RET在FCT方面的表现均要优于RSWIFT和DCQCN。
Bolt算法[67]在主机侧分别对SRC数据包和ACK数据包做出不同反应。对SRC数据包,Bolt认为每窗口一次(once-per-window)的调整策略显得比较笨拙,容易造成利用率不足和利用率振荡。于是,Bolt将流量窗口的最小可调整间隔设置为SRC数据包控制环路延迟rttsrc与目标队列大小targetq之比,其中rttsrc为当前时间now与数据包出口时间戳标记t_data_tx的差值,targetq利用流速率(flow.rate),SRC携带的瓶颈链路速率标记(link_rate)和队列大小(q_size)由式(10)计算得到。若相邻SRC数据包间隔大于该比值,Bolt将对窗口进行递减。
$ {target}_{\mathrm{q}}=\dfrac{flow.rate}{link\_rate}\times q\_size \;。 $
响应ACK数据包的方式则相对简单,当ACK数据包携带的inc标记为1时,说明链路中存在可用带宽,此时Bolt对窗口进行递增;否则保持窗口大小不变。通过细粒度拥塞信号、短控制环路延迟以及快速占领可用带宽3个机制,Bolt能够更加迅速地做出拥塞调整,加快流速率收敛和缩短流量完成时间。
微基准测试实验结果表明,Bolt的SRC机制能够在不到2个RTT的时间内完全排空拥塞队列;PRU机制能够在流完成时立即捕获可用带宽,几乎没有排队或欠利用现象;SM机制能够在链路故障或路由变更等不可预测事件后快速恢复利用带宽。
基于窗口的拥塞控制能够保证进入网络的数据包数量是固定的,提高了对反馈延迟的容忍度[9,22],其实施难度较低,但是无法避免数据包突发[84]。
RoGUE算法[20]借鉴TCP Vegas算法来对窗口进行调整。具体而言,每当完成一个数据段传输并获得相应的RTT值curr_rtt后,RoGUE将结合基本RTT值base_rtt计算实际窗口与预期窗口之差diff,如式(11)所示。若diff>β,则对窗口进行加法减小;若diff<α,则对窗口进行加法增大。
$ diff=cwnd\times \dfrac{curr\_rtt-base\_rtt}{curr\_rtt} \;。 $
为了避免单纯依靠窗口所带来的突发,RoGUE还利用硬件速率限制器对发送速率进行控制,发送速率rate_limit为cwnd与base_rtt之比。
RoGUE旨在为不需要PFC机制的RoCE网络提供有效的拥塞控制和恢复机制。实验结果表明,RoGUE在丢包率为0.025%时性能良好,能够处理几乎所有的丢包情况;在没有PFC机制的情况下,RoGUE的收敛性和公平性优于DCQCN和TIMELY,平均吞吐量接近公平份额。
HPCC算法[9]利用较为精确的INT信号实现了MIMD算法。该算法具有2个关键思想:其一是利用窗口和速率对发送方的行为进行限制以避免突发,同时避免因反馈延迟导致的网络负载过重;其二是在1个RTT周期内仅对窗口首个数据包对应的ACK数据包做出反应,避免反应过度造成带宽浪费。
HPCC通过窗口来对链路j上的inflight字节数进行控制,初始窗口大小cwndinit设置为发送方网卡带宽BNIC与BaseRTT的乘积。HPCC认为如果能将每条链路的inflight字节总数Ij保持在略小于该链路带宽Bj与BaseRTT乘积的水平,则一定不会出现排队与拥塞,因此HPCC把目的inflight字节总数设置为0.95$ \times $Bj$ \times $BaseRTT,窗口的大小以式(12)进行变化,其中kj是实际inflight字节数与目的inflight字节数之比,cwndai是确保公平性而设置的窗口增量。
$ cwnd=\dfrac{cwnd}{\max\{{k}_{j}\}}+{cwnd}_{\mathrm{a}\mathrm{i}}\;。 $
HPCC认为,如果对每一个到来的ACK都做出响应将导致窗口的过度调整,原因在于1个RTT周期内的大部分ACK数据包所反映的拥塞情况是相同送窗口${cwnd}^{\,\mathrm{c}} $并“记住”每次窗口调整后发送的第1个数据包只有收到该数据包对应的ACK数据包时才会更新$ cwnd\mathrm{^c} $值。此时的发送窗口调整函数修改为式(13)。
$ cwnd=\dfrac{{cwnd}^{\mathrm{c}}}{\max\{{k}_{j}\}}+{cwnd}_{\mathrm{a}\mathrm{i}}\;。 $
为了避免流量突发,HPCC在窗口的基础上还利用速率对数据包进行控制,发送速率设置为cwnd与BaseRTT之比。综上所述,HPCC利用MIMD加快了流量的收敛速度,引入AI提高了流量公平性,同时综合利用了窗口与速率控制提高了稳定性,通过少量带宽换取零排队实现了超低延迟。
FastTune算法[66]对窗口的调节在很大程度上参考了HPCC的思路,即通过txBytes,qLen,ts等信息计算链路利用率、通过设定预期链路利用率作为参考值、通过参考窗口cwndc实现per-RTT调整、通过将发送窗口转换为发送速率以控制突发等,本文不再赘述。
POWERTCP算法[51]的优化目标集中于流完成时间,同时不会影响吞吐量和延迟等基本性能要求。其拥塞调整部分由3个函数组成,分别为:1)NEWACK。由ACK数据包到达事件触发功率计算和窗口更新函数。2)NORMPOWER。利用ACK数据包携带的INT信息计算功率Γsmooth为更新窗口提供依据。3)UPDATEWINDOW。利用Γsmooth和前一阶段的窗口大小cwndold来调整当前的窗口大小cwnd。
具体来说,NORMPOWER函数利用ACK携带的队列大小qsize、传输字节数txBytes、链路带宽B和时间戳TS,结合事先商定的BaseRTT以及历史信息来计算瓶颈链路BDP、队列大小、传输速率以及队列变化率。其中,链路BDP和队列大小之和为电压,传输速率和队列变化率之和为电流,功率等于电压与电流的乘积。
UPDATEWINDOW函数利用计算出的归一化功率Γsmooth和cwndold来计算新窗口值,如式(14)所示。其中$\gamma $是EWMA参数,β是加性增量。
$ cwnd=\gamma \times \left(\dfrac{{cwnd}_{\mathrm{o}\mathrm{l}\mathrm{d}}}{{\varGamma }_{\mathrm{s}\mathrm{m}\mathrm{o}\mathrm{o}\mathrm{t}\mathrm{h}}}+\beta \right)+(1-\gamma )\times cwnd\;。 $
实验结果表明,POWERTCP在不同程度网络负载环境下的长流、短流完成时间均要优于HPCC、DCQCN和TIMELY算法;在incast流量模式下,POWERTCP能够快速减轻拥塞,保持稳定的吞吐量,而TIMELY和HPCC在拥塞发生后吞吐量会出现不稳定的情况。
基于窗口与速率的调整算法显然同时具备二者的优点,既能避免数据突发,又可以限制在控制环路期间发送的数据包数量,提高网络稳定性。代价是需要更多的硬件组件或逻辑组件,给终端主机的资源造成负担。
网络场景多样性和网络内在动态性使得通用拥塞控制算法的设计极为困难[84],例如大象流和老鼠流共存的混合环境、多对1传输的incast环境和多对多传输的多瓶颈环境等。另外,性能之间的掣肘使得拥塞控制算法难以实现全面的性能表现,例如吞吐量与延迟、吞吐量与稳定性之间存在着一定的权衡[9]。为了克服或缓解上述问题带来的影响,已有工作对现有算法进行优化,本文将对部分优化工作进行介绍。
微软公司和哥伦比亚大学在2016年的CoNEXT会议上发表的文章[57]就DCQCN和TIMELY算法以及它们所代表的ECN标记感知方法和延迟梯度感知方法进行了讨论,并提出了Patched TIMELY算法来对TIMELY公平性不足进行补救。
1)DCQCN和TIMELY的对比。DCQCN具有唯一实现流量公平性的固定点;若采用推荐的参数值,DCQCN将以指数级的速度收敛;若控制环路的延迟较低,DCQCN在固定点附近能保持较好的稳定性,然而其稳定性易受流量数量的影响。相较而言,TIMELY理论上没有固定点,即TIMELY无法实现公平性;队列长度和流量速率不会收敛;系统将支持处于振荡状态。
2)TIMELY与Patched TIMELY的对比。TIMELY在实际使用中通过数据段突发实现了公平性和收敛性,但其收敛速度仍然不理想。另外,由于数据段突发会引起队列波动,因此其稳定性较差,链路利用率较低。为此,本文提出了Patched TIMELY算法用于解决上述问题。如图23所示,Patched TIMELY利用加权函数w将RTT梯度(normalized_gradient)转换为权重(weight),结合归一化RTT测量值(new_rtt)与RTT参考值(rtt_ref)之差(error)将速率调整函数修改为基于weight和error的函数,使得Patched TIMELY具有唯一的固定点,并保证其收敛速度和稳定性。
3)ECN与延迟的对比。出端口的ECN标记方法将排队延迟和反馈延迟分离开来,而延迟却无法将二者解耦,因此随着队列增大,延迟携带的拥塞信息滞后;此外,基于延迟的方案难以同时实现流量公平性和固定队列大小;最后,ECN携带的拥塞信号只会随着反向链路的拥塞而被推迟,而延迟携带的拥塞信号将会被反向链路的拥塞污染,因此ECN对反向链路拥塞的容忍度更高。综上所述,本文认为基于ECN的方案保持的队列长度更短且更稳定,有利于减小FCT。
西安交通大学和清华大学2018年在TON上提出的CEDM标记方法[72]针对突发流量对ECN进行了适当优化:在数据包进入出口队列和离开出口队列时通过瞬时队列长度queue_length、缓冲区阈值K1和K2以及队列变化率avgs对数据包进行标记,引入2个阈值以及队列变化率的主要目的是排除突发流量引起的短暂队列变化对标记产生影响,避免拥塞反应过度以及队列长度振荡。
根据IP报文头中的ECN字段,CEDM的数据包可以划分为4种类型,如表10所示,入队前和出队后的数据包要么是未标记数据包,要么是拥塞数据包,而未确定数据包只会在排队时出现。
图24展示了CEDM在出口队列处入队和出队的标记逻辑:①入队长度enqueue_length>K1,队列斜率avgs>0,则将未标记状态修改为未确定;②出队长度dequeue_length>K1,avgs>0,则将未标记状态修改为拥塞状态;③enqueue_length>K2,则将未标记状态修改为拥塞状态;④dequeue_length>K2,则将未标记/未确定状态修改为拥塞状态。除上述4种情况外均不会修改数据包字段。CEDM还对K1和K2的值进行了细致的推理,使队列具备防止缓冲区溢出以及容纳突发流量的能力。
香港科技大学和微软公司在2019年的CoNEXT会议上提出的ECN#标记方法[36]认为在计算ECN标记阈值时应考虑RTT的变化,否则会对吞吐量和延迟的权衡产生不良影响。ECN#做出该判断的前提是将网络组件对数据包的处理延迟包含在RTT内(类似于Swift的做法),通过实验ECN#证明了上层网络组件的数量、类型和负载程度会导致RTT成倍数变化,因此有必要在确定标记阈值时将RTT变化考虑在内。
考虑到实施难度,ECN#并未实现动态标记阈值,而是利用较高百分位的RTT来计算阈值以保证较高吞吐量和突发容忍(ECN标记阈值K的计算公式:$ K=\lambda \times C\times BaseRTT $,其中λ是0~1的参数,C是瓶颈链路带宽,BaseRTT是基本往返时间),同时在长期队列堆积时进一步标记数据包以保证较低延迟。在具体实现中,ECN#将队列长度转换为排队时间sojourn_time参与运算。
ECN#标记算法可总结为3个函数,分别是:瞬时标记函数;持久标记函数;持久队列堆积检测函数。当数据包离队时,首先进入瞬时标记函数,若数据包排队时间(sojourn_time)大于瞬时阈值(ins_target),则进行标记;否则进入持久标记函数,首先调用检测函数来判断当前队列是否处于持久堆积状态,根据其返回结果detected以及其他参数来保证一个持久标记间隔(pst_interval)内仅标记1个数据包,其中持久队列堆积检测函数认为只有当sojourn_time在大于持久阈值(pst_target)的状态保持pst_interval的时间才可将队列视为持久堆积状态。
谷歌公司和米兰大学2023年提出的Poseidon算法[41]底层拥塞感知机制是INT机制,用于指导拥塞控制的信号是由INT机制采集的最大每跳延迟(maximum per-hop delay,MPD)。其原理可以概括为:利用交换机的INT技术,Poseidon能够获取网络中瓶颈跳的延迟(MPD),将MPD与基于流速率调整的最大每跳延迟目标(maximum per-hop delay target,MPT)进行比较,以自适应的方式决定增加或减少拥塞窗口大小。
HPCC,POWERTCP,Bolt在各节点采集的大量元数据随着节点累加,Poseidon则仅采集每跳的延迟信息并实时替换,最终保留MPD。其优势在于数据头空间大大减小,减轻了交换机和网络的负担;同时由于将拥塞信息准确定位到瓶颈跳,有利于实现数据流的最大最小公平性。
Poseidon作为谷歌TIMELY和Swift的后继方案,其底层拥塞感知机制由早期算法的时间戳标记更改为更加先进的INT标记,有能力直接采集数据包在某节点的排队时间,端主机无需对时间信息进行复杂运算;三者相同之处为均使用延迟作为拥塞感知信号,TIMELY和Swift基于端到端延迟测量,拥塞信号包含了所有跳的拥塞信息,Poseidon则基于瓶颈跳的排队时间,端主机仅响应瓶颈跳拥塞。
哈佛大学、南加州大学和伦敦玛丽大学在2020年的SIGCOMM会议上提出的PINT算法[75]旨在限制每个数据包中添加的信息量的同时,提供与INT类似的网络可见性。其基本原理为:1)通过在多个数据包上编码请求的数据,允许每个数据包的开销限制在用户设定的范围内,最低可至每个数据包1位;2)使用概率编码技术,通过近似来满足用户设定的开销预算,对于上层用户而言,遥测数据的近似值已足够。
PINT在拥塞控制领域的应用具有以下优点:1)PINT不仅将收集统计数据相关的开销降至最低,而且能够获取近似甚至更好的性能;2)由于降低了开销,PINT机制有助于提高算法的可扩展性;3)由于PINT允许网络在不同的网络负载和拓扑变化下灵活地收集遥测数据,使得拥塞控制算法能够更好地适应网络条件的变化。
然而,由于本身的概率特性,PINT也存在以下缺陷:1)概率编码不确定。基于PINT的拥塞控制机制意味着不是每个数据包都会携带完整的遥测信息,可能导致在某些情况下无法获得足够的信息来准确判断网络状态。2)对短流不友好。对于以短流为主的网络环境,PINT可能无法有效地收集足够的遥测数据,因为它们没有足够的数据包来分散编码遥测信息。3)实际部署较为复杂。PINT需要核心交换设备的支持,而当前支持PINT的交换机型号有限,在一定程度上不利于PINT机制在拥塞控制领域的大规模应用。
基于规则的拥塞控制算法在特定环境、特定任务下有助于实现良好性能,但在环境或任务发生变化时需要网络管理员花费数周的时间重新设置规则与参数[84-86]。鉴于机器学习(machine learning,ML)较高的灵活性和适应性,网络领域正探讨将ML应用于拥塞控制,即基于学习的拥塞控制。
文献[84]将基于学习的拥塞控制算法分为面向性能的算法和数据驱动的算法,前者通常需要提前设定效用函数并手动修改其中的参数,后者则依赖于数据集进行实时学习。数据驱动的算法根据其采用的技术可以划分为监督学习、无监督学习和强化学习(reinforcement learning,RL)三类,该文指出基于学习的拥塞控制大部分研究集中在强化学习这一类。然而,由于数据中心网络通常是单一管理阈下的有线网络环境,拓扑结构与流量状态相对稳定,因此相较于互联网等高带宽延迟积的网络,面向数据中心的强化学习拥塞控制研究并不是主流方向[84]。
强化学习本质上是一类闭环问题[87],强化学习系统由代理与环境组成,其中代理是做出行为决策的实体,环境是向代理反馈行为代价的实体。代理实体的策略函数用于指导其具体行动,价值函数用于定义强化学习的目标;环境实体的奖励函数用于反馈每次行动的收益,观测函数向代理反馈环境的变化信息。其具体过程可概括为:代理在状态St下执行行动At,环境受到At的影响并向代理反馈一个奖励信号Rt以及观测信号Ot,代理利用Rt和Ot决定下一阶段的状态St+1以及所要做出的行动At+1,在此过程中代理要保证价值函数达到最大。该过程可以被建模为马尔可夫决策过程(Markov decision processes,MDP)[87-88]。表11所示为强化学习问题与拥塞控制问题的实体与函数对应关系。
以色列理工学院、英伟达研究所和巴伊兰大学在2021年提出的PCC-RL算法[88]指出基于强化学习的拥塞控制算法在实际应用中面临多代理(即多个发送方的行动可能互相影响)、部分可观测(即并不是所有的环境变量都可以反馈)和多目标(即多个网络性能目标导致价值函数相对复杂)三个问题,因此MDP并不完全适用于拥塞控制机制。为此,ADPG将拥塞控制视为部分可观测的多代理多目标强化学习任务,并建模为部分可观测马尔可夫过程(partially observable Markov decision process,POMDP)。
PCC-RL规定每个代理只观测与自身相关的部分环境变量而不是全局信息;每个代理拥有一个价值函数;所有代理共享相同的策略函数。在上述规定的制约下,PCC-RL创造性地引入了策略梯度方案,通过精心设计奖励函数的结构来推测能够获取最大奖励的策略。同年,该团队在CCGrid会议上发表的文章[85]讨论了在网卡上部署基于强化学习的拥塞控制算法面对的计算能力有限问题,通过将神经网络架构精简为决策树大大降低了推理时间,并对推理得到的策略的质量影响微乎其微。鉴于奖励函数的设计与推理以及精简神经网络的内容并不是本文的重点,对此不再赘述。
近年来,大模型的迅速发展对计算、网络与存储构成的分布式系统提出了更高要求。为了匹配计算与存储的指数增长趋势,数据中心网络必须为分布式系统间通信提供高带宽、低延迟、零丢包与高可靠等特性。作为一种高性能网络实现技术,RDMA在数据中心的部署逐渐得到了更广泛的关注。
设计良好的拥塞控制机制是维持RDMA性能优势的重要手段之一。本文通过将拥塞控制划分为感知与拥塞调整对国内外的研究现状进行了详述,并对不同感知方法和调整方法的优劣进行总结,最后阐述了算法优化和基于学习的拥塞控制工作。
根据发送速率或窗口的计算实体进行算法划分是比较常见的划分方法。在发送方驱动的算法中,发送方接收来自交换机或接收方的拥塞信号,启发式地对速率或窗口进行调整,上文提到的DCQCN,TIMELY,RoGUE,HPCC,Swift,FastTune,LDCP,PowerTCP,RET,Bolt,ACC属于该范畴。交换机驱动的算法充分利用了现代交换机的计算能力和可编程能力,通过直接向数据源返回更新后的速率或窗口有效地缩短了控制环长度,但是难以大规模部署,RoCC,HierCC,PACC算法属于该范畴。接收方驱动的算法相较于发送方能够更早地利用拥塞信息进行速率或窗口计算,同时可以将接收方的接收能力考虑在内,PCN,RCC算法属于该范畴。
数据中心RDMA拥塞控制已获得10余年的关注与发展。结合当前研究成果,该领域面临的挑战可以系统地总结为“快、准、智、广、全”。“快”表现为难以快速识别并响应拥塞;“准”则是指难以精准识别拥塞流与受害流,难以精确调整发送速率或窗口;“智”揭示了人工智能方法当前难以在拥塞控制领域展开广泛应用;“广”体现在难以大规模部署新算法;“全”意味着难以兼顾网络拥塞与主机拥塞,难以兼顾各种流量模式。对其详细阐述如下:
1)第1个RTT内感知拥塞。作为支撑各种上层应用的基础设施,数据中心存在形式丰富的流量模式,其中老鼠流突发造成的首个RTT内拥塞是一个亟待解决的问题。端到端拥塞控制机制需要大于一个RTT的时间才能将拥塞感知信号或拥塞调整信号返回到发送方,因此在发送方对拥塞做出反应之前老鼠流就已停止,即端到端拥塞控制算法无法解决老鼠流突发造成的拥塞问题。当前的解决方案通常是利用交换机的可编程能力,由交换机返回拥塞信号,使发送方在第1个RTT内对网络拥塞做出反应。
2)拥塞流的精准识别与控制。交换机共享缓冲区的有限队列划分及排队机制,以及PFC机制在队列粒度上的识别共同造成了受害流问题。当受害流与拥塞流在同一出端口的相同队列中排队且超过标记阈值时,受害流将被无辜地标记,最终导致其发送速率或窗口被减小;若下游触发了PFC机制,受害流甚至会被暂停。解决该问题的根本途径是精准识别并控制拥塞流,这通常需要自定义交换机的标记策略以及终端的识别策略。
3)速率或窗口的精确调整。拥塞信号的丰富度一定程度上限制了拥塞调整的准确性。纵观网络设备及算法的发展历史,早期的TCP Reno算法以丢包为拥塞信号,以“慢启动、拥塞避免、快速重传和快速恢复”为拥塞调整手段,即丢包携带的二进制信息决定了拥塞调整不得不基于一定的经验法则;当前较为先进的算法则基于交换机的INT功能,或是基于绝对延迟和延迟梯度的PID控制器,有望实现在少量迭代次数后将速率或窗口调整到精确值。
4)人工智能具有通过自适应学习和决策优化网络传输性能的优势,在网络拥塞控制领域的应用越来越广泛,然而也面临着诸多挑战。首先,模型的训练与推理通常需要大量的计算资源,这要求充足的硬件支持,如高速处理器和专用加速器,使得相关算法在网卡上的部署变得困难;其次,模型在网络环境快速变化时可能会出现灾难性遗忘,无法有效适应新的任务;此外,模型的可解释性也是一大挑战,因为黑盒模型难以诊断和修正潜在的决策偏差。为了应对上述挑战,首先可以通过模型蒸馏和网络结构优化减少对硬件资源的需求,以降低部署成本;其次可采用多任务学习和迁移学习等技术,提高模型对新环境的适应能力,减轻灾难性遗忘的影响;最后,结合可解释性技术,增强模型的透明度和可控性,以便更好地理解和改进拥塞控制策略。这些方法为人工智能在拥塞控制领域的应用提供了新的解决思路。
5)新算法的大规模部署。定制网卡与定制交换机能否批量生产决定了算法能否在短期内大规模部署。学术界的许多先进算法通常需要异构硬件的支持,部署面临困难;工业界则更加注重算法的落地应用,因此微软、谷歌、阿里巴巴等大型科技公司在面对该挑战时具有更大的优势。
6)网络拥塞与主机拥塞。网络拥塞是指网络设备存储转发能力有限导致的数据包在网络中大量积聚并扩散的情况。主机拥塞是指端主机存储处理能力有限导致数据包在主机排队甚至反压到网络中的情况。前者通常是当前拥塞控制方案的主要设计目标,后者在某种意义上可类比为流量控制,近年来许多算法已尝试解决该问题。上述2种类型拥塞需要软硬件协同设计解决,通常来说底层硬件提供快速缓解网络拥塞的处理逻辑,上层软件则需要为缓解主机拥塞提供评估依据。
7)面向多流量模式的通用算法设计。数据中心流量在时间和空间尺度表现出了一定程度的动态特征,例如在时间上表现为不可预测性以及在空间上表现为局部聚集性。一些拥塞控制算法在特定流量模式下表现出色,然而当流量模式改变后其性能也将随之下降。如何利用有限的资源,通过算法设计来实现一个较为通用的拥塞控制算法是一项兼具挑战与价值的工作。
正如RFC 1925所说:“It is always something. Good, Fast, Cheap: Pick any two (you can’t have all three)[89]”。在网络领域,“Good”通常指高质量、高性能和高可靠的网络,“Fast”则是指快速的响应时间或完成时间,“Cheap”意味着较高的成本经济性,这三者通常无法兼得。实现“快、准、智、全”的根本目的是提高底层网络的性能与可靠性,加速数据流完成时间,提高上层应用的表现,即实现“Good”与“Fast”,这通常需要增加网卡或交换机设计的复杂度。然而,对硬件的修改不可避免会引入更高的成本,不利于算法的大规模部署,即无法满足“广”的需求,实现“Cheap”目标。实际上,拥塞控制算法的设计需要综合考虑网络拓扑、流量模式和应用需求等多方面因素,并对设计目标进行优先级排序,以达到性能与成本以及性能的不同方面之间的权衡。
随着人工智能技术与应用的快速演进,RDMA拥塞控制作为更高带宽、更低延迟数据传输的基础保障技术,其创新发展与迭代部署势在必行。本文预测,学术界作为开展创新研究最活跃的群体,利用可编程网络设备解决快速拥塞感知、拥塞流的精确识别以及准确的拥塞调整等难点问题将是未来的主流研究方向;以微软、谷歌、英伟达和阿里巴巴等大型科技企业为代表的工业界则会更加关注算法创新与商品设备成本之间的权衡,利用其规模优势推动形成新的事实标准。
作者贡献声明:张毓涛和杨惠负责相关工作调研和整理、论文架构设计、算法的仿真测试以及论文的撰写和修改;李韬指出了该领域的未来发展方向,并对论文进行了修改;黄曼蒂参与了论文架构设计与修改;李天云补充了基于强化学习的拥塞控制方法;孙志刚负责论文审阅并提出指导意见。张毓涛和杨惠为共同第一作者。

参考文献 引证文献
排序方式:
1
中国通服数字基建产业研究院. 中国数据中心产业发展白皮书[EB/OL]. [2024-03-24]. https: //aimg8.dlssyht.cn/u/551001/ueditor/file/276/551001/1684888884683143.pdf
China Communications Services Digital Infrastructure Industry Research Institute. White paper on the development of China’s data center industry[EB/OL]. [2024-03-24]. https://aimg8.dlssyht.cn/u/551001/ueditor/file/276/551001/1684888884683143.pdf (in Chinese)
2
Guo Chuanxiong, Wu Haitao, Deng Zhong, et al. RDMA over commodity ethernet at scale[C]//Proc of the 2016 ACM SIGCOMM Conf. New York: ACM, 2016: 202−215
3
Yan Dingyu, Liu Yaping, Zhang Shuo, at al. PCNP: A RoCEv2 congestion control using precise CNP[J]. Computer Nestworks, 2024, 247(C): 110453
4
Simons J. RDMA in the cloud: Enabling high-bandwidth, low-latency communication in virtual environments for HPC[EB/OL]. [2024-03-24]. https://octo.vmware.com/wp-content/uploads/2014/11/ndm2014-sc14-simons3.pdf
5
Hu Jinbin, Shen Houqiang, Liu Xuchong, et al. RDMA transports in datacenter networks: Survey[J]. IEEE Network, 2024, 38(6): 380−387
6
Mittal R, Shpiner A, Panda A, et al. Revisiting network support for RDMA[C]//Proc of the 2018 Conf of the ACM Special Interest Group on Data Communication. New York: ACM, 2018: 313−326
7
Wang Zilong, Luo Layong, Ning Qingsong, et al. SRNIC: A scalable architecture for RDMA NICs[C]//Proc of the 20th USENIX Symp on Networked Systems Design and Implementation (NSDI’23). Berkeley, CA: USENIX Association, 2023: 1−14
8
Zhu Yibo, Eran H, Firestone D, et al. Congestion control for large-scale RDMA deployments[C]//Proc of the 2015 ACM Conf on Special Interest Group on Data Communication. New York: ACM, 2015: 523−536
9
Li Yuliang, Miao Rui, Liu Hongqiang, et al. HPCC: High precision congestion control[C]//Proc of the 2019 ACM Special Interest Group on Data Communication. New York: ACM, 2019: 44−58
10
Cheng Wenxue, Qian Kun, Jiang Wanchun, et al. Re-architecting congestion management in lossless Ethernet[C]//Proc of the 17th USENIX Symp on Networked Systems Design and Implementation (NSDI’20). Berkeley, CA: USENIX Association, 2020: 131−148
11
Taheri P, Menikkumbura D, Vanini E, et al. RoCC: Robust congestion control for RDMA[C]//Proc of the 16th Int Conf on Emerging Networking Experiments and Technologies. New York: ACM, 2020: 17−30
12
Zhang Yiran, Liu Yifan, Meng Qingkai, et al. Congestion detection in lossless networks[C]//Proc of the 2021 ACM SIGCOMM 2021 Conf. New York: ACM, 2021: 370−383
13
Dai Huichen, Fu Binzhang, Tan Kun. LDCP: PFC-free low delay control protocol[EB/OL]. [2024-03-30]. https://datatracker.ietf.org/meeting/108/agenda/tsvwg-drafts.pdf
14
Zhang Yiran, Meng Qingkai, Hu Chaolei, et al. Revisiting congestion control for lossless Ethernet[C]//Proc of the 21st USENIX Symp on Networked Systems Design and Implementation (NSDI’24). Berkeley, CA: USENIX Association, 2024: 131−148
15
IEEE 802.1. 802.1Qbb – Priority-based flow control[EB/OL]. [2024-04-10]. https://1.ieee802.org/dcb/802-1qbb/
16
Alizadeh M, Atikoglu B, Kabbani A, et al. Data center transport mechanisms: Congestion control theory and IEEE standardization[C]//Proc of the 46th Annual Allerton Conf on Communication, Control, and Computing. Piscataway, NJ: IEEE, 2008: 1270−1277
17
Stephens B, Cox A L, Singla A, et al. Practical DCB for improved data center networks[C]//Proc of the 2014 IEEE Conf on Computer Communications. Piscataway, NJ: IEEE, 2014: 1824−1832
18
Hu Shuihai, Zhu Yibo, Cheng Peng, et al. Deadlocks in datacenter networks: Why do they form, and how to avoid them[C]//Proc of the 15th ACM Workshop on Hot Topics in Networks. New York: ACM, 2016: 92−98
19
Shpiner A, Zahavi E, Zdornov V, et al. Unlocking credit loop deadlocks[C]//Proc of the 15th ACM Workshop on Hot Topics in Networks. New York: ACM, 2016: 85−91
20
Le Yanfang, Stephens B, Singhvi A, et al. RoGUE: RDMA over generic unconverged Ethernet[C]//Proc of the 2018 ACM Symp on Cloud Computing. New York: ACM, 2018: 225−236
21
Tang Jian, Xu Tingting, Nguyen C, et al. Tuning target delay for RTT-based congestion control[C]//Proc of the 30th IEEE Int Conf on Network Protocols (ICNP). Piscataway, NJ: IEEE, 2022: 1−11
22
Zhang Jiao, Shi Jiaming, Zhong Xiaolong, et al. Receiver-driven RDMA congestion control by differentiating congestion types in datacenter networks[C]//Proc of the 29th IEEE Int Conf on Network Protocols (ICNP). Piscataway, NJ: IEEE, 2021: 1−12
23
Huang Shan, Dong Dezun, Bai Wei. Congestion control in high-speed lossless data center networks: A survey[J]. Future Generation Computer Systems, 2018, 89(C): 360−374
24
曾高雄, 胡水海, 张骏雪, 等. 数据中心网络传输协议综述[J]. 计算机研究与发展, 2020, 57(1): 74−84
Zeng Gaoxiong, Hu Shuihai, Zhang Junxue, et al. Transport protocols for data center networks: A survey[J]. Journal of Computer Research and Development, 2020, 57(1): 74−84 (in Chinese)
25
杜鑫乐, 徐恪, 李彤, 等. 数据中心网络的流量控制: 研究现状与趋势[J]. 计算机学报, 2021, 44(7): 1287−1309
Du Xinle, Xü Ke, Li Tong, et al. Traffic control for data center network: State of the art and future research[J]. Chinese Journal of Computers, 2021, 44(7): 1287−1309 (in Chinese)
26
Yan Dingyu, Liu Yaping, Zhang Shuo, at al. A survey of RoCEv2 congestion control[C]//Proc of the 7th Int Conf on Information Science, Communication and Computing. Berlin: Springer, 2023: 42−56
27
Alizadeh M, Greenberg A G, Maltz D A, et al. Data center TCP (DCTCP)[C]//Proc of the ACM SIGCOMM 2010 Conf. New York: ACM, 2010: 63−74
28
Alizadeh M, Yang Shuang, Sharif M, et al. pFabric: Minimal near-optimal datacenter transport[J]. SIGCOMM Computer Communication Review, 2013, 43(4): 435−446
29
Alizadeh M, Kabbani A, Edsall T, et al. Less is more: Trading a little bandwidth for ultra-low latency in the data center[C]//Proc of the 9th USENIX Conf on Networked Systems Design and Implementation. Berkeley, CA: USENIX Association, 2012: 1−14
30
Zats D, Das T, Mohan P, et al. DeTail: Reducing the flow completion time tail in datacenter networks[C]//Proc of the ACM SIGCOMM 2012 Conf on Applications, Technologies, Architectures, and Protocols for Computer Communication. New York: ACM, 2012: 139−150
31
Wilson C, Ballani H, Karagiannis T, et al. Better never than late: Meeting deadlines in datacenter networks[C]//Proc of the ACM SIGCOMM 2011 Conf. New York: ACM, 2011: 50−61
32
Hong Chiyao, Caesar M, Godfrey P B. Finishing flows quickly with preemptive scheduling[J]. SIGCOMM Computer Communication Review, 2012, 42(4): 127−138
33
Dean J, Barroso L A. The tail at scale[J] Communications of the ACM, 2013, 56(2): 74−80
34
Grosvenor M P, Schwarzkopf M, Gog I, et al. Queues don’t matter when you can JUMP them![C]//Proc of the 12th USENIX Symp on Networked Systems Design and Implementation (NSDI’15). Berkeley, CA: USENIX Association, 2015: 1−14
35
Vamanan B, Hasan J, Vijaykumar T N. Deadline-aware datacenter TCP (D2TCP)[C]//Proc of the ACM SIGCOMM 2012 Conf on Applications, Technologies, Architectures, and Protocols for Computer Communication. New York: ACM, 2012: 115−126
36
Zhang Junxue, Bai Wei, Chen Kai. Enabling ECN for datacenter networks with RTT variations[C]//Proc of the 15th Int Conf on Emerging Networking Experiments and Technologies. New York: ACM, 2019: 233−245
37
Roy A, Zeng Hongyi, Bagga J, et al. Inside the social network’s (datacenter) network[C]//Proc of the 2015 ACM Conf on Special Interest Group on Data Communication. New York: ACM, 2015: 123−137
38
Evangelinos C, Hill C N. Cloud computing for parallel scientific HPC applications: Feasibility of running coupled atmosphere-ocean climate models on Amazon’s EC2[J]. Ratio, 2008, 2(2): 1−6
39
Dean J, Ghemawat S. MapReduce: Simplified data processing on large clusters[J]. Communications of the ACM, 2008, 51(1): 107−113
40
Kohavi R, Longbotham R. Online experiments: Lessons learned[J]. Computer, 2007, 40(9): 103−105
41
Wang Weitao, Moshref M, Li Y, et al. Poseidon: Efficient, robust, and practical datacenter CC via deployable INT[C]//Proc of the 20th USENIX Symp on Networked Systems Design and Implementation (NSDI’23). Berkeley, CA: USENIX Association, 2023: 255−274
42
Dragojevic A, Narayanan D, Hodson O, et al. FaRM: Fast remote memory[C]//Proc of the 11th USENIX Conf on Networked Systems Design and Implementation. Berkeley, CA: USENIX Association, 2014: 401−414
43
Hwang J, Cai Qizhe, Tang Ao, et al. TCP ≈ RDMA: CPU-efficient remote storage access with i10[C]//Proc of the 17th USENIX Symp on Networked Systems Design and Implementation (NSDI’20). Berkeley, CA: USENIX Association, 2020: 127−140
44
Cheng Huang, Simitci H, Xu Yikang, et al. Erasure coding in windows azure storage[C]//Proc of the 2012 USENIX Conf on Annual Technical. Berkeley, CA: USENIX Association, 2012: 1−12
45
Poke M, Hoefler T. DARE: High-performance state machine replication on RDMA networks[C]//Proc of the 24th Int Symp on High-Performance Parallel and Distributed Computing. New York: ACM, 2015: 107−118
46
Mitchell C, Geng Yifeng, Li Jinyang. Using one-sided RDMA reads to build a fast, CPU-efficient key-value store[C]//Proc of the 2013 USENIX Conf on Annual Technical. Berkeley, CA: USENIX Association, 2013: 103−114
47
Wu Haitao, Feng Zhenqian, Guo Chuanxiong, et al. ICTCP: Incast congestion control for TCP in data-center networks[J]. IEEE/ACM Transactions on Networking, 2013, 21(2): 345−358
48
Lao ChonLam, Le Yanfang, Mahajan K, et al. ATP: In-network aggregation for multi-tenant learning[C]//Proc of the 18th USENIX Symp on Networked Systems Design and Implementation (NSDI’21). Berkeley, CA: USENIX Association, 2021: 741−761
49
Zhang Jiao, Zhang Yali, Guan Zixuan, et al. HierCC: Hierarchical RDMA congestion control[C]//Proc of the 5th Asia-Pacific Workshop on Networking (APNet’2021). New York: ACM, 2021: 29−36
50
Kumar G, Dukkipati N, Jang K, et al. Swift: Delay is simple and effective for congestion control in the datacenter[C]//Proc of the Annual Conf of the ACM Special Interest Group on Data Communication on the Applications, Technologies, Architectures, and Protocols for Computer Communication. New York: ACM, 2020: 514−528
51
Addanki V, Michel O, Schmid S. POWERTCP: Pushing the performance limits of datacenter networks[C]//Proc of the 19th USENIX Symp on Networked Systems Design and Implementation (NSDI’22). Berkeley, CA: USENIX Association, 2022: 51−70
52
Zhong Xiaolong, Zhang Jiao, Zhang Yali, et al. PACC: Proactive and accurate congestion feedback for RDMA congestion control[C]//Proc of the 2022 IEEE Conf on Computer Communications (INFOCOM 2022). Piscataway, NJ: IEEE, 2022: 2228−2237
53
Infiniband Trade Association. InfiniBand architecture volume 1, general specifications, release 1.2.1[EB/OL]. [2024-04-23]. https://www.infinibandta.org/ibta-specification/
54
Recio R J, Culley P R, Garcia D, et al. A remote direct memory access protocol specification[EB/OL]. [2024-04-23]. https://doi.org/10.17487/RFC5040
55
Infiniband Trade Association. Supplement to infiniBand architecture specification volume 1 release 1.2.2 annex A16: RDMA over converged Ethernet (RoCE)[EB/OL]. [2024-04-23]. https://www.infinibandta.org/ibta-specification/
56
Infiniband Trade Association. Supplement to infiniBand architecture specification volume 1 release 1.2.2 annex A17: RoCEv2[EB/OL].[2024-04-23]. https://www.infinibandta.org/ibta-specification/
57
Zhu Yibo, Ghobadi M, Misra V, et al. ECN or delay: Lessons learnt from analysis of DCQCN and TIMELY[C]//Proc of the 12th Int Conf on Emerging Networking Experiments and Technologies. New York: ACM, 2016: 313−327
58
Intel. Intel E810 Ethernet network adapter[EB/OL]. [2024-04-27]. https://www.intel.cn/content/www/cn/zh/products/details/ethernet/800-network-adapters/e810-network-adapters/products.html
59
Broadcom. Broadcom BCM957508-P1200G NIC datasheet[EB/OL]. [2024-04-27]. https://docs.broadcom.com/doc/957508-P1200G-DS1XX
60
NVIDIA. NVIDIA Connectx-7 network adapter datasheet[EB/OL]. [2024-04-27]. https://nvdam.widen.net/s/csf8rmnqwl/infiniband-ethernet-datasheet-connectx-7-ds-nv-us-2544471
61
MARVELL. MARVELL converged Ethernet network adapter[EB/OL]. [2024-04-27]. https://cn.marvell.com/products/ethernet-adapters-andcontrollers/fastlinq-cna-adapters/documents.html
62
华为. 华为SP600网卡[EB/OL]. [2024-04-27]. https: //support.huawei.com/enterprise/zh/doc/EDOC1100309168?idPath=23710424%7C251364409%7C21782478%7C15791
Huawei. Huawei SP600 NIC[EB/OL]. [2024-04-27]. https://support.huawei.com/enterprise/zh/doc/EDOC1100309168?idPath=23710424%7C251364409%7C21782478%7C15791 (in Chinese)
63
Chelsio. Chelsio T7 DPU products[EB/OL]. [2024-04-27]. https://www.chelsio.com/wp-content/uploads/resources/t7-dpu-asic.pdf
64
Lee C, Park C, Jang K, et al. DX: Latency-based congestion control for datacenters[J]. IEEE/ACM Transactions Networking, 2017, 25(1): 335−348
65
Mittal R, Lam V T, Dukkipati N, et al. TIMELY: RTT-based congestion control for the datacenter[C]//Proc of the 2015 ACM Conf on Special Interest Group on Data Communication. New York: ACM, 2015: 537−550
66
Zhou Renjie, Dong Dezun, Huang Shan, et al. FastTune: Timely and precise congestion control in data center network[C]//Proc of the 2021 IEEE Int Conf on Parallel &amp; Distributed Processing with Applications, Big Data &amp; Cloud Computing, Sustainable Computing &amp; Communications, Social Computing &amp; Networking (ISPA/BDCloud/SocialCom/SustainCom). Piscataway, NJ: IEEE, 2021: 238−245
67
Arslan S, Li Yuliang, Kumar G, et al. Bolt: Sub-RTT congestion control for ultra-low latency[C]//Proc of the 20th USENIX Symp on Networked Systems Design and Implementation (NSDI’23). Berkeley, CA: USENIX Association, 2023: 219−236
68
Floyd S, Jacobson V. Random early detection gateways for congestion avoidance[J]. IEEE/ACM Transactions on Networking, 1993, 1(4): 397−413
69
Ramakrishnan K, Floyd S, Black D. The addition of explicit congestion notification (ECN) to IP[EB/OL]. [2024-05-02]. https://doi.org/10.17487/rfc3168
70
IEEE. IEEE 802.1 qau-congestion notification[EB/OL]. [2024-05-02]. http://www.ieee802.org/1/pages/802.1au.html
71
Wu Haitao, Ju Jiabo, Lu Guohan, et al. Tuning ECN for data center networks[C]//Proc of the 8th Int Conf on Emerging Networking Experiments and Technologies. New York: ACM, 2012: 25−36
72
Shan Danfeng, Ren Fengyuan. ECN marking with micro-burst traffic: Problem, analysis, and improvement[J] IEEE/ACM Transactions on Networking, 2018, 26(4): 1533−1546
73
Qian Kun, Cheng Wenxue, Zhang Tong, et al. Gentle flow control: Avoiding deadlock in lossless networks[C]//Proc of the ACM Special Interest Group on Data Communication. New York: ACM, 2019: 75−89
74
The P4. org Applications Working Group. In-band network telemetry (INT) dataplane specification[EB/OL]. [2024-05-12]. https://p4.org/p4-spec/docs/INT_v2_1.pdf
75
Ben Basat R, Ramanathan S, Li Yuliang, et al. PINT: Probabilistic in-band network telemetry[C]//Proc of the Annual Conf of the ACM Special Interest Group on Data Communication on the Applications, Technologies, Architectures, and Protocols for Computer Communication. New York: ACM, 2020: 662−680
76
Chelsio. Chelsio T5 packet rate performance report[EB/OL]. [2024-05-13]. https://www.chelsio.com/wp-content/uploads/2013/08/Chelsio-T5-Packet-Rate-Performance-Report.pdf
77
Silicom. Dual port 10 gigabit server adapter with precision time stamping[EB/OL]. [2024-05-13]. https://www.silicom-usa.com/wp-content/uploads/2016/08/PE210G2TSI9-10G-Precision-Time-Stamping-Server-Adapter.pdf
78
ntop. Using hardware timestamps with PF RING[EB/OL]. [2024-05-13]. https://www.ntop.org/pf_ring/hardware-timestamps-with-pf_ring/#:~:text=PF_RING%20has%20been%20enhanced%20to%20support%20hw%20timestamps%2C,the%20NIC%20that%20received%20packets%20must%20be%20timestamped
79
ntop. Who (really) needs sub-microsecond packet simestamps[EB/OL]. [2024-05-13]. https://www.ntop.org/pf_ring/who-really-needs-sub-microsecond-packet-timestamps/
80
NVIDIA Mellanox. Highly accurate time synchronization with connectX-3 and timeKeeper[EB/OL]. [2024-05-13]. https://enterprise-support.nvidia.com/s/article/highly-accurate-time-synchronization-with-connectx--174--3-and-timekeeper--174-x
81
Brakmo L S, O’Malley S W, Peterson L L. TCP Vegas: New techniques for congestion detection and avoidance[J]. SIGCOMM Computer Communication Review, 1994, 24(4): 24−35
82
Wei D X, Cheng Jin, Low S H, et al. FAST TCP: Motivation, architecture, algorithms, performance[J]. IEEE/ACM Transactions on Networking, 2006, 14(6): 1246−1259
83
Tan Kun, Song Jingmin, Zhang Qian, et al. A compound TCP approach for high-speed and long distance networks[C]//Proc of the 25th IEEE Int Conf on Computer Communications (INFOCOM’2006). Piscataway, NJ: IEEE, 2006: 1−12
84
Jiang Huiling, Li Qing, Jiang Yong, et al. When machine learning meets congestion control: A survey and comparison[J]. Computer Networks, 2021, 192: 108033
85
Fuhrer B, Shpigelman Y, Tessler C, et al. Implementing reinforcement learning datacenter congestion control in NVIDIA NICs[C]//Proc of the 23rd IEEE/ACM Int Symp on Cluster, Cloud and Internet Computing (CCGrid). New York: ACM, 2022: 331−343
86
Dong Mo, Meng Tong, Zarchy D, et al. PCC vivace: Online-learning congestion control[C]//Proc of the 15th USENIX Conf on Networked Systems Design and Implementation. Berkeley, CA: USENIX Association, 2018: 343−356
87
Sutton R S, Barto A G. Reinforcement Learning: An Introduction[M]. Cambridge, MA: A Bradford Book, 2018
88
Tessler C, Shpigelman Y, Dalal G, et al. Reinforcement learning for datacenter congestion control[J]. SIGMETRICS Performance Evaluation Review, 2022, 49(2): 43−46
89
Callon R. RFC1925: The twelve networking[EB/OL]. [2024-05-02]. https://doi.org/10.17487/RFC1925
2026年第63卷第2期
196
74
0
引用本文
BibTeX
文章信息
doi: 10.7544/issn1000-1239.202440623
  • 接收时间:2024-07-20
  • 首发时间:2026-11-06
  • 出版时间:2026-02-01
补充材料
相关文章
文章信息
作者
出版历史
  • 收稿日期:2024-07-20
  • 修回日期:2025-04-21
  • 录用日期:2025-06-11
基金
作者信息
    国防科技大学计算机学院 长沙 410073

通讯作者:

李韬()
参考文献
分享链接
https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/10.7544/issn1000-1239.202440623
分享至
全文二维码

扫描看全文

引用本文
BibTeX
本文的引用情况
2种不同金属材料的力学参数
科
Family
属数
Number of
genus
种数
Number of
species
占总种数比例
Percentage of
total species (%)
属
Genus
种数
Number of
species
占总种数比例
Percentage of total
species (%)
鹅膏菌科Amanitaceae 2 11 5.26 鹅膏菌属 Amanita 10 4.78
小菇科 Mycenaceae 2 12 5.74 丝盖伞属 Inocybe 5 2.39
多孔菌科 Polyporaceae 8 14 6.70 蜡蘑属 Laccaria 5 2.39
红菇科 Russulaceae 3 23 11.00 小皮伞属 Marasmius 6 2.87
小菇属 Mycena 11 5.26
光柄菇属 Pluteus 5 2.39
红菇属 Russula 17 8.13
栓菌属 Trametes 5 2.39
关闭全屏