Bundling Spatially Correlated Prefetches for Improved Main Memory Row Buffer Locality
Patrick Judd · TSpace (University of Toronto) · 2014
Row buffer locality is a consequence of programs' inherent spatial locality that the memory system can easily exploit for significant performance gains and power savings. However, as the number of cores on a chip increases, request streams become interleaved more frequently and row buffer locality is lost. Prefetching can help mitigate this effect, but more spatial locality remains to be recovered. In this thesis we propose Prefetch Bundling, a scheme which tags spatially correlated prefetches with information to allow the memory controller to prevent prefetches from becoming interleaved. We evaluate this scheme with a simple scheduling policy and show that it improves the row hit rate by 11%. Unfortunately, the simplicity of the scheduling policy does not translate these row hits to improved performance or power savings. However, future work may build on this framework and develop more balanced policies that leverage the row locality of Prefetch Bundling.