Home Latest Articles
Latest Articles
  • Shipeng Hu, Guangyan Zhang, Weimin Zheng
    Journal of Computer Research and Development. 2026, 63(4): 1021-1038.

    Large language model (LLM) inference requires buffering and reusing intermediate data, namely KV cache. However, the rapid growth in the size of KV cache imposes significant storage pressure on systems. On one hand, the large-scale KV cache of ongoing inference requests leads to GPU memory bottlenecks, thereby limiting inference parallelism. On the other hand, the long-term memory capability enabled by LLM service providers results in large-scale KV cache being stored on persistent storage devices, incurring high storage costs and loading overhead. To alleviate the storage pressure caused by KV cache, existing research proposes various KV cache compression strategies to reduce its size. Nonetheless, it remains challenging to achieve high compression ratios while minimizing the impact on model response quality and keeping the additional computational overhead at an acceptable level. We present a systematic survey of recent advances in KV cache compression strategies. First, we analyze the scale of KV cache and its impact on inference efficiency and storage cost. Based on this, we identify three major challenges in designing KV cache compression strategies: preserving model response quality, minimizing computational overhead, and ensuring generalizability. We then categorize existing KV cache compression approaches into five types: merge-based, quantization-based, token eviction-based, sharing-based, and attention head pruning-based strategies. Each category is elaborated upon with representative examples, discussing their design principles and implementation techniques. Finally, we outline potential future research directions in KV cache compression.

  • Rengang Li, Xia Li, Jun Liu, Pengyu Hou, Yanwei Wang, Guangwen Yang
    Journal of Computer Research and Development. 2026, 63(4): 1039-1062.

    Driven by emerging high-performance computing applications, the demand for high-speed interconnects across heterogeneous computing system components has surged dramatically, making interconnect technologies a key research focus in the computer architecture field in recent years. The open standard CXL (compute express link), with its advantages of high scalability, high bandwidth, and low latency, has become the mainstream of interconnect technology in heterogeneous computing systems, which has gained much attention from both academia and industry. This survey comprehensively analyzes and summarizes CXL technology from three aspects, including CXL protocol architecture, device types and its performance characteristics, as well as CXL memory management and application research. Firstly, the overall architecture of the CXL protocol is introduced. The cache coherence mechanism is analyzed based on typical protocol transaction flows. Then, we provide an overview of CXL device types and development progress in industrial products domestically and abroad, followed by systematical analysis of real CXL device’s performance characteristics. Finally, we summarize current academic application research, scenarios and challenges on the CXL technology from the perspective of CXL device types. On this basis, future research directions are discussed to promote its application research and industrial implementation in various fields, such as memory pooling, CXL-based processing-in-memory architecture and heterogeneous collaborative computing.

  • Jiaxuan Long, Chunhua Xiao, Taoning Sun, Haoyu Xiong, Xiaohan Zuo, Lin Wu, Yanke Chen
    Journal of Computer Research and Development. 2026, 63(4): 1063-1078.

    The development of the mobile application ecosystem has intensified memory pressure, while traditional physical expansion solutions are limited. As a result, compressed memory has become the mainstream technology for memory expansion. However, existing compressed memory systems still suffer from low file page performance, high compression overhead, and suboptimal storage resource utilization, negatively impacting overall system performance. To address these issues, we propose two optimization frameworks: EC-MemOpt and APC-MSO. EC-MemOpt improves file page I/O performance and storage resource utilization by migrating frequently accessed file pages to compressed memory and leveraging compression prediction techniques. APC-MSO optimizes the compression swap strategy for anonymous pages, reducing unnecessary compression computation overhead. Furthermore, a collaborative management architecture for hybrid memory pages (EC-MemOpt + APC-MSO) is designed to achieve differentiated partition management. Compared with the QCOM 8350R kernel maintained by OnePlus, the proposed architecture improves application switching speed and space utilization by 10.5% and 9.8%, respectively. During the experiments, large volumes of anonymous and file pages are tested under diverse workload conditions. The results verify that the hybrid framework maintains stable performance across varying memory access patterns. Experimental results demonstrate that the proposed optimization solutions significantly enhance the overall performance of compressed memory systems in mobile environments.

  • Li Chen, Jun Gao, Tianlei Zhao, Qiao Liu
    Journal of Computer Research and Development. 2026, 63(4): 1079-1093.

    The large instruction fetch delay which is caused by the miss of L1 instruction cache is one of the most important bottlenecks that restrict the performance development of modern processors, especially in server applications which have a very large number of instruction traces. Instruction prefetching technology is a key technology to solve this problem, and it achieves the purpose of masking the high access latency by putting the instruction cache blocks to be used in L1 instruction cache in advance. In recent years, researchers have proposed a lot of instruction prefetching architectures to alleviate this problem, but due to poor temporal and spatial locality, long-distance function calls still bring a large number of missing instructions. In this paper, a new instruction prefetching mechanism is designed, which can achieve high coverage and high accuracy prefetching of the target instruction for function calls with low hardware overhead. Experiments show that after applying the optimization in this paper, the missing rate of the function call target instruction is about 45% lower than that of the current state-of-the-art instruction prefetcher, and the IPC (instruction per cycle) performance is about 11.9% higher than the baseline, and about 2.9% higher than the current state-of-the-art instruction prefetcher with similar overhead.

  • Guojie Li
    Journal of Computer Research and Development. 2026, 63(3): 539-547.

    This paper reconceptualizes AI safety from the lens of logical complexity, and establishes three levels of safety issues: R1 Level (decidable propositions, supporting formal pre-verification), R2 Level (semi-decidable propositions, supporting only post-hoc evidence discovery), and R3 Level (non-recursively enumerable, where even the identification of insecurity cannot be guaranteed). Distinguishing between R1 and R2 is critical: all engineering-tractable security issues reside within the R1 Level. Consequently, achieving security requires a dual-track effort focused on correctness verification and institutional safety nets. Regarding the high-profile field of AI security, while current risks have not yet escalated to the R3 Level, the governance trajectory must urgently shift from “pre-verification” to “runtime governance”, prioritizing external monitoring including gating, rollbacks, isolation, human-in-the-loop systems, and permission hierarchies. Furthermore, it necessitates a dual-sovereignty system combining embedded technology and external institutions, thereby maintaining human corrective sovereignty and civilizational security within a logically incomplete reality.

  • Jiwu Shu, Yijie Wang
    Journal of Computer Research and Development. 2026, 63(3): 548-549.
  • Xi Cao, Mingjie Li, Chaoshu Yang, Cheng Yang, Runyu Zhang
    Journal of Computer Research and Development. 2026, 63(3): 550-566.

    Emerging zoned namespace solid state drive (ZNS SSD) addresses several critical issues inherent in traditional SSDs, including high write amplification, low storage density, and complex I/O paths. These advancements create new opportunities for the progress of storage technologies. The B+tree, as an efficient tree index structure, is widely used in various databases and file systems to support the efficient loading of large models, the construction of external knowledge bases, and the management of structured metadata, thereby significantly improving training efficiency and knowledge invocation performance. However, due to the hardware characteristics of ZNS SSDs being different from traditional block devices, directly deploying B+tree to ZNS SSDs not only results in higher write amplification rates but also causes cascading updates, seriously affecting the performance of the storage system. For the above problems, we propose DPZB+tree, a B+tree indexing structure optimized for ZNS SSDs and incorporating persistent memory (PM). First, DPZB+tree adopts a hybrid DRAM-PM-ZNS SSD storage architecture, thereby effectively separating hot data and cold data. Second, DPZB+tree introduces a lightweight hot/cold node identification strategy, which in turn improves I/O efficiency. Third, to address the limited capacity of PM, we present a dynamic node placement strategy that adaptively migrates data between PM and ZNS SSD. Finally, we combine hardware characteristics and spatiotemporal locality principles to design leaf node splitting and merging operations. DPZB+tree is implemented using a ZNS SSD simulator and Intel optane PM. Experimental results show that under various workloads, DPZB+tree outperforms LSM tree, SSDB+tree, DZB+tree, and Baseline in terms of performance and recovery time.

  • Qiang Li, Tao Liu, Xueqing Zhang, Xianwei Meng, Minggang Sun, Bole Li, Xi Chen
    Journal of Computer Research and Development. 2026, 63(3): 567-584.

    Global cache is a core technology for enterprise-grade storage arrays, whose primary function is to improve data write performance. I/O operations return immediately once data is written into the memory cache of storage controllers, without waiting for the data to be persisted to slow HDDs or SSDs. However, the design of global cache in enterprise-grade storage products must comprehensively consider factors such as the number of replicas, system high availability, and failure scenarios. Constrained by the mutual interactions of these factors, the number of controllers in storage clusters typically cannot exceed 4, and system high availability fails to support the scenario where three out of four controllers fail concurrently. Furthermore, the transmission of cache replica data between controllers via RDMA(remote direct memory access) technology also impairs storage performance. Thus, a memory-disaggregated global cache system prototype is designed based on the latest CXL(compute express link) technology. With memory cache disaggregated from within controllers to external, the system offers three key advantages: first, controllers and memory cache can scale out independently, supporting multiple cache replicas; second, controller failures do not cause cache data loss, supporting concurrent failures of n−1 out of n controllers; third, CXL technology significantly optimizes storage performance while enhancing system scalability and high availability. Experimental results demonstrate that the CXL-based global cache achieves a 38% performance improvement compared with the RDMA-based global cache.

  • Jian Gao, Jiwu Shu
    Journal of Computer Research and Development. 2026, 63(3): 585-596.

    Distributed lock is a crucial component in distributed storage systems. The performance of the lock protocol significantly influences the overall performance of the entire system. Remote direct memory access (RDMA) is an emerging data center networking technology that supports one-sided communication verbs and offers low CPU overhead, low latency, and high throughput. It presents new opportunities for designing high-performance distributed lock protocols. However, designing such RDMA based network protocols faces significant challenges, for example scalability and fairness. We address these challenges by proposing FeLock, an RDMA-based distributed lock protocol. FeLock achieves high performance while tackling the above challenges by leveraging different RDMA communication verbs, which enables the clients to communicate directly with both the server for lock acquisition and release, and other clients to hand over lock ownership. Specifically, first, to improve performance, FeLock introduces a per-node lock management mechanism to reduce network roundtrips on the critical paths of the lock protocol. Second, to achieve scalability, FeLock incorporates a round-robin handover mechanism, in which nodes are logically organized into a ring, and clients hand over lock ownership sequentially according to their positions within the ring. Third, to ensure fairness and prevent clients from starvation, FeLock employs a node credit mechanism that limits the number of consecutive lock acquisitions by any single node, thereby preventing clients on any node from being indefinitely blocked by others. Experimental results demonstrate that FeLock achieves performance comparable to or exceeding that of existing one-sided RDMA lock protocols, such as DSLR, while exhibiting better fairness and scalability. With 3 to 120 clients, FeLock achieves throughput 1.01 to 7.51 times of DSLR, with its fairness improved by up to 2.24 times.

  • Fangyu Wei, Tianyun Chen, Yang Zhou, Yulai Xie, Fang Wang
    Journal of Computer Research and Development. 2026, 63(3): 597-614.

    In the context of the information age, the explosive growth of data volume and the diversification of compression application scenarios raise higher demands for the flexibility and efficiency of compression strategies. The LZ77 algorithm, a classical lossless compression method, is widely used in mainstream compression tools such as Zstandard (ZSTD). However, achieving higher compression ratios often requires larger history windows and more complex compression strategies, which in turn lead to frequent cache misses and reduced efficiency of lazy matching in ZSTD’s LZ77 implementation. To address these issues, two optimization strategies are proposed. First, the multi-level region search strategy (MLRS) introduces a hierarchical matching region mechanism along with access threshold control, enabling adaptive adjustment of search depth and restricting memory access range during the matching process to alleviate cache pressure. Second, the extended search-based lazy matching strategy (ESLM) reuses search paths and incorporates approximate substitution techniques to reduce redundant computation while improving matching efficiency. These strategies are implemented and evaluated under ZSTD level 12 configuration on Kunpeng 920 server platform. Experimental results demonstrate that MLRS significantly reduces last-level cache miss rates across various datasets and improves compression throughput to 118.34%−149.50% of the baseline while maintaining a compression ratio within 94.65%−99.58%. ESLM improves compression throughput to 113.49%−117.46% of the original while further increasing the compression ratio on most datasets. When both strategies are applied jointly, the compression speed increases to 134.53%−171.17% of the original scheme, while sustaining a compression ratio in the range of 94.18%−99.80%.