In this paper we propose and evaluate the Adaptive prefetching technique for Software DSM (Software-based Distributed Shared-Memory) systems over three different communication layers on top of Myrinet an Interconnection Network. Use of a User-Level Communication Layer provides us a low-overhead and high-bandwidth that make it possible to analyze the prefetching effect on Software DSM systems using high performance network. GM-VIA and S-VIA are two User-Level Communication Layers proposed for this paper to analyze prefetching effect on Software DSM systems. And for comparison raisons, and since most of all previous existing works on prefetching for DSM were tested on a traditional communication protocol TCP/IP, we also propose an implementation of TCP/IP over GM that provides the same quality of message delivery as a TCP/IP but on the top of Myrinet. Adaptive Prefetching Technique improves the performance for regular parallel applications running on Software DSM system by using past history of memory access faults to adapt between the phase mode and stride mode. This technique doesn’t issue prefetches during periods when the application is behaving irregularly. The experimental results obtained on two Splash-2 applications show that prefetching can provide more performance improvement with S-VIA than other communication layers even if GM-VIA provides more speedup than S-VIA when no prefetching is issued. Our main conclusion is that prefetching techniques have to be taken in consideration by software designers as an effective way of tolerating the overhead of remote data accesses even when high interconnection network is used. In term of overall performance the Adaptive Prefetching Technique can provides speedup improvements as significant as 10% on 8 processors with S-VIA.
본 논문에서는 다양한 통신계층상에서 소프트웨어 분산공유메모리 시스템을 위한 적응적 선인출 기법을 제안한다. 본 연구에서 이용한 한 통신계층은 기존의 통신계층인 TCP/IP와 고성능을 제공하는 사용자 수준 통신계층인 GM_VIA, S-VIA를 사용하였으며, 이들은 모두 미리넷(myrinet)상에서 구현되었다.
적응적 선인출 기법은 메모리 접근패턴이 정규적인 응용프로그램의 성능을 향상시켰다. 하지만 응용프로그램이 비정규적으로 메모리를 접근하는 부분에서는 성능향상에 도움이 되지 않았다.
실험결과 통신성능에서는 GM-VIA 보다 S-VIA 가 앞서지만, 응용프로그램의 성능은 S-VIA 가 더 좋게 나타났다. 결과적으로 선인출기법을 통하여 메모리 접근시간을 효과적으로 줄일 수 있었다.