Improving Performance of the Irregular Data Intensive Application with Small Computation Workload for CMPs

Abstract

The data needs of scientific or commercial applications from a diverse range of fields have been increasing exponentially over the recent years. Although the traditional systems work well for computation that requires limited data handling, the CMPs in cloud computing may below performance for the computation that requires large amounts of intensive data. Conventional helper thread techniques try to improve the high performance overheads, but they can not improve performance of the irregular data intensive applications with small computation workload. Our goal is to provide a novel solution to improve the application performance in data intensive computing environments. By introducing the prepuce look ahead Size K, the prepush block size P and the synchronization block size B three operations to helper thread, we expect to reduce the overheads introduced by the traditional helper thread and leave the computing resources to perform useful prefetch work. As a starting point, we design the KPB interleaved data prepush algorithm, and use Q6600 and IBM 5110 multi-core computers as our test platforms to study behaviors of the benchmarks fromSPEC2006 suite and Olden suite. We construct the helper threads of mcf from SPEC2006, mst and em3d from Olden by using our method, the average result of speedup is 1.23, 1.32and 1.09 separately on the Q6600 machine, and 1.28, 1.35 and1.23 separately on another machine. Compared with the AP and PV methods, our method is less negative impact than both AP and PV, our KPB-method is also better than AP and PV in the prefetching timeliness and control ability.