IMP: Indirect Memory Prefetcher

Abstract

Machine learning, graph analytics and sparse linear algebra-based applications are dominated by irregular memory accesses resulting from following edges in a graph or non-zero elements in a sparse matrix. These accesses have little temporal or spatial locality, and thus incur long memory stalls and large bandwidth requirements. A traditional streaming or striding prefetcher cannot capture these irregular access patterns.

A majority of these irregular accesses come from indirect patterns of the form A[B[i]]. We propose an efficient hardware indirect memory prefetcher (IMP) to capture this access pattern and hide latency. We also propose a partial cacheline accessing mechanism for these prefetches to reduce the network and DRAM bandwidth pressure from the lack of spatial locality.

Evaluated on 7 applications, IMP shows 56% speedup on average (up to 2.3×) compared to a baseline 64 core system with streaming prefetchers. This is within 23% of an idealized system. With partial cacheline accessing, we see another 9.4% speedup on average (up to 46.6%).