Synchronization-aware Prefetching

Data prefetching is an important part in increasing the performance of the majority of programs written. Most of the designs for data prefetchers today try to find patterns in the data access. Other prefetchers attempt to correlate addresses, so that the access to one address is indicative of another access to a different address in the near future. Multi-threaded programs use locks or barriers to synchronize accesses to shared data. The data that is shared between threads of a program tend not to move much in memory, due to the desire to keep as much of the memory processor specific to avoid coherence misses and costs. The prefetcher that we propose in this paper takes advantage of how those programs use locks and the nature of shared data in general. In this project, we examine the effectiveness of using synchronization primitives as prefetch hints to improve performance in scientific workloads.