Scholarlink

<p>As an emerging and powerful approach to break through the memory wall, CXL-SSD bridges the performance gap between CPU and memory. To further reduce memory access latency, prefetching techniques have been introduced in the CXL-SSD system. However, modern prefetching techniques are mostly designed for DRAM and conventional storage devices, which are not suitable for predicting CXL-SSD access patterns. We propose a dynamic feedback prefetcher that combines local spatial and global performance information (LG-Prefetcher) for CXL-SSD. It learns some prefetch offsets by analyzing the differences between recent access history records and dynamically adjusts prefetch offset filtering criteria based on global performance of the current prefetcher, providing feedback for the next prefetching to adjust the prefetching range accordingly. Simulation results demonstrate that LG-Prefetcher exhibits superior performance compared to other prefetchers, with average memory access latency reduction of up to 62.86%.</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

LG-Prefetcher: a prefetcher combining local spatial and global performance information for CXL-SSD

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Shibao Li,
  • Chengzhi Wang,
  • Zhou Yang,
  • Yunwu Zhang,
  • Longfei Li,
  • Chenxu Ma,
  • Xuerong Cui,
  • Lianghai Li,
  • Jianhang Liu

摘要

As an emerging and powerful approach to break through the memory wall, CXL-SSD bridges the performance gap between CPU and memory. To further reduce memory access latency, prefetching techniques have been introduced in the CXL-SSD system. However, modern prefetching techniques are mostly designed for DRAM and conventional storage devices, which are not suitable for predicting CXL-SSD access patterns. We propose a dynamic feedback prefetcher that combines local spatial and global performance information (LG-Prefetcher) for CXL-SSD. It learns some prefetch offsets by analyzing the differences between recent access history records and dynamically adjusts prefetch offset filtering criteria based on global performance of the current prefetcher, providing feedback for the next prefetching to adjust the prefetching range accordingly. Simulation results demonstrate that LG-Prefetcher exhibits superior performance compared to other prefetchers, with average memory access latency reduction of up to 62.86%.