Scholarlink

<p>Modern processors rely on multiple prefetching techniques to mitigate memory latency, but aggressive and uncoordinated prefetching can lead to contention for shared resources such as caches and memory bandwidth. To address this, a reinforcement learning-based approach is proposed that leverages Proximal Policy Optimization to manage prefetcher coordination efficiently. Additionally, a multi-objective reward function is employed that considers cache pollution, Miss Status Holding Register occupancy, Instructions Per Cycle, and bandwidth optimizing overall system performance. Unlike conventional classification-based methods, this reinforcement learning framework introduces significantly higher average coverage performance up to 80% in a single core and 81% in quadcores having homogeneous workloads while dynamically adapting to workload variations through online training. This approach effectively handles a wide range of workloads and configurations across multiple cache levels, including Level 1 Instruction Cache, Level 1 Data Cache, Level 2 Cache, and Last Level Cache, ensuring multilevel coordination, achieving 37. 66% increase in instruction per cycle over no prefetching system. The multiprefetcher system achieves an average coverage of 80.87% across single-core traces, improving to 81.57% in homogeneous quad-core workloads and 80. 86% in heterogeneous quad-core workloads, outperforming RL-CoPref by 4.39% in coverage. Scalability analysis shows that the multiprefetcher maintains coverage at 81.90% for 1 core, peaking at 82.69% for 2 cores, and declining to 55.35% for 16 cores, demonstrating resilience despite contention at higher core counts.</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

Multilevel Cache Prefetcher Management Using Reinforcement Learning

  • 期刊论文

  • 发表日期：

    2026年6月18日

查看全文
  • Vaishnavi Mulik,
  • Anvay Joshi,
  • Amaan Jamadar,
  • Amit D. Joshi,
  • N. Ramasubramanian

摘要

Modern processors rely on multiple prefetching techniques to mitigate memory latency, but aggressive and uncoordinated prefetching can lead to contention for shared resources such as caches and memory bandwidth. To address this, a reinforcement learning-based approach is proposed that leverages Proximal Policy Optimization to manage prefetcher coordination efficiently. Additionally, a multi-objective reward function is employed that considers cache pollution, Miss Status Holding Register occupancy, Instructions Per Cycle, and bandwidth optimizing overall system performance. Unlike conventional classification-based methods, this reinforcement learning framework introduces significantly higher average coverage performance up to 80% in a single core and 81% in quadcores having homogeneous workloads while dynamically adapting to workload variations through online training. This approach effectively handles a wide range of workloads and configurations across multiple cache levels, including Level 1 Instruction Cache, Level 1 Data Cache, Level 2 Cache, and Last Level Cache, ensuring multilevel coordination, achieving 37. 66% increase in instruction per cycle over no prefetching system. The multiprefetcher system achieves an average coverage of 80.87% across single-core traces, improving to 81.57% in homogeneous quad-core workloads and 80. 86% in heterogeneous quad-core workloads, outperforming RL-CoPref by 4.39% in coverage. Scalability analysis shows that the multiprefetcher maintains coverage at 81.90% for 1 core, peaking at 82.69% for 2 cores, and declining to 55.35% for 16 cores, demonstrating resilience despite contention at higher core counts.