Home Archive
Archive
2024 Volume 61 Issue 12  Published: 2024-12-01
    Computer Architecture
  • Zhang Zhang , Gang Shi , Qifan Wang , Yongbo Ma , Gang Liu , Libo Qian
    doi: 10.7544/issn1000-1239.202330364

    The fast and low-power in-memory computing architecture, which integrates memory and calculation, breaks through the traditional von-Neumann system that separates memory and calculation, and solves the problem of “memory wall” that limits the arithmetic power of the processor, which has become a research hotspot of new computing architecture. The basic devices for in-memory computing include fast and mature static random access memory (SRAM), low power, fast response and non-volatile memristor, and high density, low static power and non-volatile magnetic random access memory (MRAM). Up to the present, a great variety of in-memory computing studies have been proposed based on these devices, however, a systematic and comprehensive literature review on these in-memory computing architectures is still missing. In this paper, we firstly introduce the in-memory computing principles of different devices, the current development status of in-memory computing architectures, and the practical application scenarios from the three directions of SRAM, memristor, and MRAM. Next, the existing solutions and the future directions for the problems and challenges of current in-memory computing architectures are given. Finally, we summarize the research focus of in-memory computing based on the above devices, outline the shortcomings of the current research, and look forward to the future development direction.

  • Computer Architecture
  • Tao Li , Hui Yang , Junnan Li , Rulin Liu , Zhigang Sun
    doi: 10.7544/issn1000-1239.202220998

    In order to meet different requirements of performance, flexibility and quality of service in new network scenarios such as 5G and 8K video, network processor (NP) becomes increasingly complex and diverse. Traditional network processors try to integrate amount of heterogeneous processing resources such as processor cores, caches and accelerators on a single SoC (system on chip) to provide highly customizable capabilities. However, with the failure of Moore’s Law and Dennard’s scaling law, developing one-big network processor becomes unsustainable as it faces greater challenges in R & D cycle, cost and innovation iteration. We propose a novel agile customizable architecture for network processor, namely ChipletNP, which decouples heterogeneous resources and uses Chiplet technology to quickly customize new NPs by combining existing mature chip products. ChipetNP is highly flexible as it has an agile switching network which can connect diverse heterogeneous resources with high throughput and predictable delay. We develop a network processor chip, i.e., YHHX-NP, based on ChipletNP architecture, which integrates commercial CPU, FPGA (field programmable gate array) and agile switching chip. Our results show that ChipletNP supports various emerging network functions such as SRv6 (segment routing over IPv6) with ultra-low latency (<2.82 µs), and achieves more than two times energy efficiency improvement compared with commercial chips.

  • Computer Architecture
  • Zhiyuan Yan , Biwei Xie , Yungang Bao
    doi: 10.7544/issn1000-1239.202330204

    Sparse matrix-vector multiplication (SpMV) plays an important role in a wide variety of scientific and engineering applications.Executing SpMV efficiently on vector processors is challenging because of the irregular memory access of sparse matrices. We conduct a detailed analysis of the main factors that have impact on the efficiency of SpMV vectorization.In addition to the irregular distribution of non-zero elements within sparse matrices, different sparse matrices also exhibit huge variations in the distribution characteristics of non-zero elements. Therefore, it is difficult to apply a universal vector optimization method for matrices with diverse characteristics. Furthermore, there is a big difference in vector computing and vector instructions for various vector processors. The primary challenge of SpMV vectorization lies in mapping the irregular sparse matrices onto the regular vector processor. In this paper, we propose a hybrid vectorization-optimized mechanism of SpMV(HVMS). HVMS models the characteristics of vector processors and designs corresponding rules based on the abstracted basic operations to guide the regularization conversion of sparse matrices. HVMS divides the matrix into different parts according to the different characteristics. For each part, the non-zero distribution can be less irregular and then HVMS introduces corresponding optimization mechanisms to boost the vectorization efficiency of SpMV. We implement and evaluate HVMS on an Intel Xeon processor and compare it with three state-of-the-art approaches using 30 sparse matrices. Experimental results show that HVMS can achieve an average speedup of 1.60x, 1.72x, and 1.93x over CVR, SELL-C-σ, and Intel MKL, respectively.

  • Computer Architecture
  • Jia Wei , Xingjun Zhang , Longxiang Wang , Mingqiang Zhao , Xiaoshe Dong
    doi: 10.7544/issn1000-1239.202330164

    Deep neural network (DNN) have achieved state-of-the-art accuracy in many modern artificial intelligence (AI) tasks. In recent years, it has become increasingly popular to use high performance computing platforms for massively distributed parallel training of DNN. Energy consumption models have been crucial in designing and optimizing DNN for massively parallel training and restraining excessive energy consumption on HPC (high performance computing) platforms. Currently, most energy consumption models model the energy consumption of a single device or a cluster of multiple devices from a hardware perspective. From an energy consumption perspective, the need for disaggregated analysis of distributed parallel DNN applications has resulted in a dearth of energy consumption models that model the characteristics of distributed DNN applications. In this paper, we propose the “materials preprocessing-computing-communicating” three-stage MC2 model from the perspective of the essential features of DNN model training for the most commonly used DNN distributed data parallel training model. The model is validated by training the classical VGG16, ResNet50 networks and the latest Vision Transformer network using up to 128 MT nodes and 32 FT nodes on the domestic E-class prototype Tianhe-3. The experimental results show that the difference between MC2 and the actual energy measurements is only 2.84%. Compared with the four linear proportional energy models and the AR, SES, and ARIMA time prediction models, the accuracy of the model proposed is improved by 69.12%, 69.50%, 34.58%, 13.47%, 5.23%, 22.13%, and 10.53%, respectively. By using the models proposed in this paper, the energy consumption of DNN models at each stage and the overall energy consumption can be obtained on a supercomputer platform, which provides a basis for evaluating the efficiency of DNN energy-aware massively distributed parallel training and inference, as well as optimizing the strategies of task scheduling, job scheduling, model partitioning, and model pruning.

  • Newwork Information and Security
  • Jinping Jia , Shihan Xiao , Kun Qian , Yanqin Yang , Zhao Zhang
    doi: 10.7544/issn1000-1239.202330298

    The Web 3.0 digital economic system takes the blockchain platform as its infrastructure, and revolves around the digital assets such as cryptocurrencies, NFTs, digital collectibles, and decentralized applications (DApps) like decentralized finance (DeFi) and gaming finance (GameFi) to conduct various socio-economic activities. Smart contracts are the core of DApps on the public blockchains and the public permissioned blockchains, such as Ethereum, Solana, EOSIO, Findora, Antchain, ChainMaker, et al. Smart contracts can be deployed by any individual or organization, and are visible and accessible to all blockchain users. This openness brings new opportunities for economic development while also harboring numerous financial risks. We analyze potential risks for Web 3.0 digital economic by focusing on smart contract and summarize the current research on risk perception technology from three aspects: encoding, functionality, and application of smart contracts. We first introduce the research challenges, security vulnerability types, and four categories of vulnerability detection methods in smart contract vulnerability detection technology. Next, we analyze common types of smart contract scams and summarize existing scam recognition techniques based on different classifications of training data. Subsequently, we introduce the current state of technology for detecting four types of illicit transactions behaviors based on blockchain transaction records. Lastly, by analyzing the limitations of existing work, we envision the future research directions.

  • Newwork Information and Security
  • Junjie Zhuang , Shuang Hu , Baojian Hua , Yang Wang , Zhizhong Pan
    doi: 10.7544/issn1000-1239.202330049

    WebAssembly is an emerging binary instruction set architecture and code distribution format, providing a unified compiling target for high-level programming languages. Due to its design advantages such as efficiency, security and portability, WebAssembly has already been widely used in the Web and non-Web scenarios, and it is becoming one of the most promising platform-independent language runtimes. Although WebAssembly provides a variety of advanced features to guarantee its security, existing studies have demonstrated that WebAssembly still has unique attack surfaces leading to serious security issues. These security issues pose challenges to the security of whole ecosystems based on WebAssembly. Therefore, it is critical to study the security issues of WebAssembly and their mitigations. To the best of our knowledge, we present the first study of WebAssembly security, based on 42 published research papers in this area. First, we systematically analyze and summarize key security features of WebAssembly. Second, we propose the first four-layer threat model for WebAssembly: threats from high-level programming languages, compilation toolchains, binary files, and WebAssembly virtual machines. Third, we propose a taxonomy to classify current research efforts into four categories: empirical security study, vulnerability detection and exploitation, security enhancements, and formal semantics and verifications. Finally, we point out potential challenges to be addressed in this field, as well as five future research directions to be explored.

  • Newwork Information and Security
  • Zili Meng , Mingwei Xu
    doi: 10.7544/issn1000-1239.202330240

    Real-time multimedia transmission is one of the most important applications of the Internet, with the applications such as videoconferencing, cloud gaming, virtual reality and so on. In the same time, the real-time multimedia transmission systems therefore demand high requirements for the end-to-end transmission latency. Among them, latency fluctuation is the most challenging problem in latency optimization. However, traditional ‘best-effort’ transmission services in the Internet cannot meet the requirements of latency fluctuations for real-time multimedia transmission in many cases. We firstly elaborates the main challenges faced by real-time multimedia transmission. Secondly, we analyze the key issues that need to be addressed to optimize the latency of real-time multimedia transmission. Based on these issues, two key paths (control path and data path) and five core components in the architecture of real-time multimedia transmission system are summarized. Around the technologies involved in each component, representative research results are summarized and discussed, especially for the research efforts in the recent years. Based on the analysis above, the research branches for real-time multimedia transmission and low-latency applications are summarized, and the optimization algorithms and applications for each research branch are reviewed. A key finding in this paper is that the fluctuation of latency is the key metric that research needs to work on. Finally, we also discuss and propose possible future research directions for the readers.

  • Newwork Information and Security
  • Lifang Bai , Yuefei Zhu , Yongjun Li , Shuai Wang , Xiaoqi Yang
    doi: 10.7544/issn1000-1239.202221052

    With the acceleration of the digitization process, the problem of data security and privacy protection has attracted much attention. Data encryption has always been an important means to solve this problem. However, it is common to store and transfer data in encrypted form. Once calculation is involved, it is often necessary to decrypt the ciphertext, perform the calculation in plaintext, and then encrypt the calculation result. Full homomorphic encryption (FHE) extends encryption to the computing, which can perform meaningful calculations in ciphertext without decryption, and the calculation process and result are encrypted, so it has the natural characteristics of protecting data security and user privacy. The first FHE scheme was proposed by Gentry in 2009, and then FHE has always attracted the attention of the industry and academia. After more than ten years of research, FHE has developed to the fourth stage, and substantial progress has been made. We review and analyze the research progress of FHE from the aspects of the construction idea, different research stages and problems faced, introduce the application progress of FHE from the aspects of algorithm library, standardization progress and typical application scenarios, and put forward suggestions for future research direction.

  • Newwork Information and Security
  • Zelin Su , Wenfang Zhang , Xiaomin Wang
    doi: 10.7544/issn1000-1239.202220133

    Attribute-based encryption is a new access control scheme based on cryptography, which is suitable for data sharing. However, the large computational and communication costs of encryption and access policy updating limit the practical application of attribute-based encryption. Moreover, most of proposed outsourcing ABE schemes do not provide a ciphertext correctness verification method for data owners. Thus, an outsourced ABE scheme with dynamic policy updating and real-time verification of ciphertext correctness is proposed to further protect data privacy in an untrusted cloud environment. In the scheme, the design of policy updating references outsourced encryption, which reduces the computational cost of generating update key. The design of ciphertext correctness verification algorithm refers to decryption operation and introduces verification transformation key to make ciphertext verification more efficient. According to different cloud environment models, efficient verification algorithm and strict verification algorithm are designed, which are suitable for honest but curious cloud environment and untrustworthy cloud environment respectively. The scheme is secure against chosen plaintext attack under the standard model. Performance analysis and efficiency comparison show that the computation of local encryption, policy updating and ciphertext verification are reduced, and the scheme is more lightweight, which is suitable for the application of computation-constrained devices in access control scenarios.

  • Newwork Information and Security
  • Xin He , Fan Wu , Yujun Zhu , Yong Xu , Panlong Yang
    doi: 10.7544/issn1000-1239.202221010

    The increasing number of the Internet of things (IoT) devices in our environment presents significant opportunities as well as challenges, including limited battery life, low computing power, and scalability for multiple access. Recent advances in backscatter communication have enabled ubiquitous IoT devices to communicate in an energy-effective manner. However, a fundamental challenge in backscatter communication is to design a multiple access protocol to support concurrent communications from numerous tags. Aiming at tackling the high bit error rate and low transmission rate of large-scale tag concurrent communication, we design a coding-based code division multiple access technique to facilitate multi-tag backscattering with reliable communication. Firstly, the chip discrimination tags with strong orthogonality such as Walsh code are combined with error correction coding to enhance the anti-interference ability. Secondly, the joint decoding of correlation detection and decoding is used to reduce the bit error rate and realize the reliable communication from the backscatter link. Finally, a grouping algorithm is designed to cluster tags based on the K-means clustering algorithm, which fully utilizes the time-code resources and enhances the capacity. Experimental results demonstrate that our scheme can achieve reliable communication with 100+ tags, with a bit error rate of approximately 2%, and the achieved throughput of 39 Mbps.

  • Artificial Intelligence
  • Junwei Li , Quan Liu , Zhigang Huang , Yapeng Xu
    doi: 10.7544/issn1000-1239.202220970

    As a common temporal abstraction method for hierarchical reinforcement learning, Option framework allows agents to learn strategies at different time scales, which can effectively solve sparse reward problems. In order to ensure that options can guide agents to access more state space, some methods improve the diversity of options by introducing mutual information in internal reward and termination functions. However, it will lead to slow algorithm learning speed and low knowledge transfer ability of internal strategy, which seriously affect algorithm performance. To address the above problems, diversity-enriched option-critic algorithm with interest functions(DEOC-IF) is proposed. Based on the diversity-enriched option-critic (DEOC) algorithm, the algorithm constrains the selection of the upper-level strategy on the internal strategy of Option by introducing the interest function, which not only ensures the diversity of the Option set, but also makes the learned internal strategies focus on different regions of the state space, which is conducive to improving the knowledge transfer ability of the algorithm and accelerating the learning speed. In addition, DEOC-IF introduces a new interest function update gradient, which is beneficial to improve the exploration ability of the algorithm. In order to verify the effectiveness and option reusability of the algorithm, the algorithm comparison experiments are carried out in four-room navigation task, Mujoco, and MiniWorld. Experimental results show that DEOC-IF algorithm has better performance and option reusability compared with other algorithms.

  • Artificial Intelligence
  • Yiwang Huang , Yuxin Huang , Sheng Liu
    doi: 10.7544/issn1000-1239.202330382

    Training deep learning models with noisy data containing lossy labels is a hot research topic in machine learning. Studies have shown that deep learning model training is susceptible to overfitting due to noisy data. Recently, a method combining meta-learning and label correction can make the model better adapt to the noisy data to mitigate the overfitting phenomenon. However, this meta-label correction method relies on the model’s performance, and the lightweight model does not have good generalization performance under noisy data. To address this problem, we propose a knowledge distillation-based meta-label correction learning method (KDMLC), which treats the meta label correction model (MLC) composed of a deep neural network and a multilayer perceptron as a teacher model to correct the noise labels and guide the training of the lightweight model, and at the same time, KDMLC adopts a two-layer optimization strategy to train and enhance the generalization ability of the teacher model, so as to generate a higher-quality pseudo-labels for training the lightweight model. The experiments show that KDMLC improves the test accuracy by 5.50% compared with MLC method at high noise level; meanwhile, using Cutout data enhancement on the CIFAR10 dataset, KDMLC improves the test accuracy by 9.11% compared with MLC at high noise level, and the experiments on the real noisy dataset, Clothing1M, also show that KDMLC outperforms the other methods, verifying that KDMLC is better than the other methods, which verifies the feasibility and validity of KDMLC method.

  • Artificial Intelligence
  • Xu Yu , Qinglong Peng , Dingjia Zhan , Junwei Du , Jinhuan Liu , Junyu Lin , Dunwei Gong , Ziying Zhang , Jie Yu
    doi: 10.7544/issn1000-1239.202330017

    Traditional cross-rating collaborative filtering paradigm ignores the influence of rating density in the target domain on the accuracy of user and item latent vectors, resulting in less accurate rating prediction in regions with sparse ratings. To overcome the influence of regional rating density on rating prediction, based on the thought of transfer learning, a cross-region and cross-rating collaborative filtering recommendation algorithm (CRCRCF) is proposed. Compared with the traditional cross-rating collaborative filtering paradigm, CRCRCF algorithm can effectively exploit not only the important knowledge from the auxiliary domain, but also the important knowledge from the rating-dense regions in the target domain, which can further improve the rating prediction accuracy of the whole target domain, especially the rating-sparse regions. Firstly, for users and items, active users and inactive users, popular items and unpopular items are divided respectively. Graph convolution matrix complementation algorithm is used to extract the latent vectors of active users and popular items in the target domain and all users and items in the auxiliary domain. Secondly, for users and items in rating-dense regions, deep regression models based on self-taught learning are constructed to learn the mapping relationships between latent vectors in the target domain and in the auxiliary domain, respectively. Then the mapping relationships are generalized to the whole target domain, and the relatively accurate latent vectors of inactive users and unpopular items in the auxiliary domain are used to derive their latent vectors in the target domain, which achieves the cross-region mapping relationships transfer and cross-rating latent vector information transfer successively. Finally, the restricted graph convolutional matrix completion model is proposed with the obtained latent vectors of inactive users and non-popular items in the target domain as constraints, and the corresponding recommendation results are given. The simulation experiments on MovieLens and Netflix datasets show that the CRCRCF algorithm has obvious advantages over other state-of-the-art algorithms.

  • Artificial Intelligence
  • Huiju Wang , Weixuan Huang , Xiao Yue
    doi: 10.7544/issn1000-1239.202330791

    With the promotion of data as a production factor, traditional query pricing methods face tremendous challenges in practical applications due to their overly strict premise assumptions, limited support for flexibility and dynamics, and inadequate consideration of key factors. To address these issues, we innovatively design a query pricing algorithm based on the labor game model. This algorithm models the participants in data transactions as labor unions and employers, and treats the data trading platform and data buyers as the labor union and employers, respectively. The data trading platform (labor union) is responsible for the fair and transparent calculation of the value of each traded dataset (wages), aiming to facilitate transactions as much as possible. Data buyers determine the purchase quantities of datasets based on their estimated value, personal needs, and budgets, thereby achieving a pricing strategy that balances the interests of all three parties. Experimental results demonstrate that compared with the popular Stackelberg game model, our algorithm better accommodates the interests of all parties and ensures greater fairness. Compared with traditional query-based data pricing methods, our pricing algorithm is more practical, offers greater flexibility and dynamics, and can dynamically adjust prices in response to changes in query results. The time complexity of our pricing algorithm is O(N), where N is the number of datasets related to the query, and it also guarantees no arbitrage.

  • Artificial Intelligence
  • Xuan Tian , Zezhou Xu , Zihan Wang
    doi: 10.7544/issn1000-1239.202220837

    Query suggestion (QS) is an indispensable part of search engines. It can provide query candidates before users entering a complete query to help express their information needs more accurately and more quickly. Deep learning helps to improve the accuracy of QS and it has become the mainstream technology to promote the development of QS in recent years. We mainly summarize, analyze and compare the research status of deep learning based QS (DQS). According to the different application stages of deep learning, DQS methods are divided into two categories: generative QS methods and ranking-based QS suggestion methods, and the modeling ideas of each model are analyzed. In addition, the data sets, baselines and evaluation indexes commonly used in the field of QS are introduced, and the technical characteristics and experimental results of different models are compared. Finally, the current challenges and future development trends of QS research based on deep learning are summarized.