Scholarlink

<p>The imbalance between computing power and memory development in computing systems has given rise to the memory wall problem, which has become a significant bottleneck in the rapid advancement of AI and HPC. As a memory expansion solution, CXL-SSD significantly alleviates this issue. Prefetching technology predicts and preloads the data to be used, further mitigating the impact of long access latencies. However, the prefetching algorithm in the existing CXL-SSD is unable to flexibly adjust the prefetching strategy according to the workload characteristics, which affects the efficiency of the data access and the system performance. To this end, we designed a prefetching optimization scheme called LDBO based on local feedback and dynamic adjustment of the prefetch amount. LDBO evaluates local metrics to assess prefetching effectiveness and adjusts the Best-Offset prefetch amount accordingly to enhance performance. We compare it with different prefetching algorithms and demonstrate that this method significantly reduces access latency.</p>

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

Local feedback and dynamic adjustment best offset prefetcher in CXL-SSD

  • 期刊论文

  • 发表日期：

    2025年11月11日

查看全文
  • Shibao Li,
  • Zhou Yang,
  • Chengzhi Wang,
  • Yunwu Zhang,
  • Wei Dou,
  • Xuerong Cui,
  • Jianhang Liu

摘要

The imbalance between computing power and memory development in computing systems has given rise to the memory wall problem, which has become a significant bottleneck in the rapid advancement of AI and HPC. As a memory expansion solution, CXL-SSD significantly alleviates this issue. Prefetching technology predicts and preloads the data to be used, further mitigating the impact of long access latencies. However, the prefetching algorithm in the existing CXL-SSD is unable to flexibly adjust the prefetching strategy according to the workload characteristics, which affects the efficiency of the data access and the system performance. To this end, we designed a prefetching optimization scheme called LDBO based on local feedback and dynamic adjustment of the prefetch amount. LDBO evaluates local metrics to assess prefetching effectiveness and adjusts the Best-Offset prefetch amount accordingly to enhance performance. We compare it with different prefetching algorithms and demonstrate that this method significantly reduces access latency.