RLOP: A Framework Design for Offset Prefetching Combined with Reinforcement Learning

Previous prefetching schemes have been found to be very effective at enhancing the performance of computers. However, speculative prefetching requests can have negative effects on computers, such as increased memory bandwidth consumption and cache pollution. To address the deficiencies of previous prefetching schemes, we propose the Reinforcement Learning Based Offset Prefetching Scheme (RLOP), an offset prefetching scheme based on reinforcement learning. As with previous offset prefetching schemes, RLOP evaluates multiple offsets and enables offsets that qualify to issue prefetching requests. RLOP, however, selects appropriate prefetch offsets through reinforcement learning, and the reinforcement learning reward scheme determines the goal of the prefetcher; we divide the rewards into four different rewards—accurate and timely prefetch, accurate but delayed prefetch, inaccurate prefetch, and no prefetch operation—and by increasing or decreasing the reward value, we facilitate or inhibit RLOP from future environments to collect such rewards, which enables or inhibits RLOP from collecting such rewards, which enables We evaluated and contrasted RLOP with various advanced data prefetchers and demonstrated that our scheme resulted in a 25.26% increase in system performance over systems without data prefetchers and a 3.8% increase over the previous best performing data prefetcher.