System-Aware Hybrid Prefetcher : A Reinforcement Learning Based Approach

The growing need for high-performance computing demands continuous innovation as the modern processors face a significant memory wall due to the growing latency gap between processor and main memory. Effective memory management is crucial for reducing latency, boosting data throughput, and enhancing overall system performance giving birth to cache hierarchy. There have been significant advancements being done in micro-architectural research over cache hierarchy like prefetching, for attaining optimal performance. Hardware-based Data prefetching is one of the most sought-after topics. Various prefetchers are being developed that targets different cache levels. Most prefetchers target the top level cache as it has direct access to the data being requested by the processor, hence direct access to the different memory access patterns. Some of these prefetchers employ hybrid approach to cover wide-range of memory access patterns to reduce demand misses. Instruction Pointer Classifier based spatial prefetcher is one such hybrid prefetcher. It works at top level cache with low storage overhead showing promising results in prefetching on regular workloads. However it leads to degradation in performance when it encounters irregular memory accesses. Pythia, an reinforcement learning based framework allows to do hybrid prefetching by selecting different prefetchers according to the system state but results in significant storage overhead managing the prefetchers. Inspired by Pythia, this work introduces a reinforcement learning controller enhancement to Instruction Pointer Classifier based spatial prefetcher to adapt it's prefetching for the irregular memory access patterns, particularly in graph processing workloads. The enhancement achieves overall speedup of 7.2% and increment in accuracy and coverage of prefetching by 126% and 15.88% respectively over original IPCP with 2.3KB of approximate storage overhead.