Scholarlink

With the increasing core count in modern multi-core processors, directory-based cache coherence protocols face significant challenges related to the “three-hop problem,” which introduces substantial latency in shared data access. This paper proposes a Fast Access Method for Shared Data (Facess) to address this issue. The Facess approach introduces a directory cache between each processor’s private cache controller and the network interface, modifies cache block status tag bits, and redesigns bus, directory, and processor state machines. Experimental results using the GEM5 simulation platform with PARSEC benchmarks show that Facess achieves remarkable latency improvements: approximately 84% for instruction-related data, 80% for data load operations, and 78% overall latency reduction.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Facess: A Fast Access Method for Shared Data Among Multi-cores in Parallel Programs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Junhui Wang,
  • Jierong Tang,
  • Shiyuan Wang,
  • Yijing Peng,
  • Hongwei Zhou,
  • Quanyou Feng,
  • Libo Huang,
  • Yongwen Wang

摘要

With the increasing core count in modern multi-core processors, directory-based cache coherence protocols face significant challenges related to the “three-hop problem,” which introduces substantial latency in shared data access. This paper proposes a Fast Access Method for Shared Data (Facess) to address this issue. The Facess approach introduces a directory cache between each processor’s private cache controller and the network interface, modifies cache block status tag bits, and redesigns bus, directory, and processor state machines. Experimental results using the GEM5 simulation platform with PARSEC benchmarks show that Facess achieves remarkable latency improvements: approximately 84% for instruction-related data, 80% for data load operations, and 78% overall latency reduction.