Temporal, Processor, and Spatial Locality in Multiprocessor Memory References.

摘要

The performance of cache-coherent multiprocessors is strongly influenced by locality in the memory reference behavior of parallel applications. While the notions of temporal and spatial locality in uniprocessor memory references are well understood, the corresponding notions of locality in multiprocessors and their impact on multiprocessor cache behavior are not clear. A locality model suitable for multiprocessor cache evaluation is derived by viewing memory references as streams of processor identifiers directed at specific cache/memory blocks. This viewpoint differs from the traditional uniprocessor approach that uses streams of addresses to different blocks emanating from specific processors. Our view is based on the intuition that cache coherence traffic in multiprocessors is largely determined by the number of processors accessing a location, the frequency with which they access the location, and the sequence in which their accesses occur. The specific locations accessed by each processor, the time order of access to different locations, and the size of the working set play a smaller role in determining the cache coherence traffic, although they still influence intrinsic cache performance. Looking at traces from the viewpoint of a memory block leads to a new notion of reference locality for multiprocessors, called processor locality. In this paper, we study the temporal, spatial, and processor locality in the memory reference patterns of three parallel applications. Based on the observed locality, we then reflect on the expected cache behavior of the three applications. (kr)