Evaluating the performance of a DSM cluster with improved communication subsystem

Abstract

Parallel applications on network-based computing are most sensitive to communication overhead. The performance of a cluster computing system largely depends on the bandwidth, latency, and communication software processing overhead of the communication subsystem. The currently available fiber optics and opto-electronic device technologies offers solution to the bandwidth problem. However, achieving low latency remains a challenge and is considered by the research community as one critical issue to attain high-performance cluster computing. Different approaches to solve the problem have been proposed which deals with latency reduction and latency hiding. In this paper, we report on the new implementation of a DSM cluster computing system using the low-level application programming interface (LAPI) provided as part of IBM SP2 software. We discuss and evaluate the new (DSE-LAPI) and previous (DSE-TCP) implementations using application benchmarks with SPLASH-2 programs, i.e., FFT, radix, and LU. Likewise, we evaluate the scalability performance of the new implementation. Experimental results show promising performance of the new implementation and it further demonstrate the relative merit of adapting LAPI on a DSM cluster.