A trace-driven analysis of parallel prefetching algorithms for parallel and distributed systems
CaiBin, Diqing Hu, Changsheng M. Xie · 2005
I/O for parallel and distributed systems has drawn increasing attention over the past decade as it has become apparent that I/O performance, rather than CPU performance, may be the key limiting factor in the performance of future systems. Prefetch is the fundamental approach for improving the overall read performance. At present, there are some well-known prefetching algorithms in parallel and distributed systems, such as LRU-Lookahead, Fixed Horizon, and greedy algorithmin. In this paper, we study these parallel prefetching algorithms and explore the performance characteristics of each of the algorithms using the trace-driven simulation. We find that when performance is limited by I/O stall, aggressive prefetch helps to alleviate the problem. However, conservative prefetch performs well in computing-bound situations. Moreover, we find that carefully choosing replacement decision is not necessary to balance the load across the disks when the data is well laid out.