Bottleneck-free interconnect and IO subsystem in SPAX

• Published in: Proceedings 1997 International Conference on Parallel and Distributed Systems pp. 524 - 533
• Main Authors: Kangwoo Lee, Woo-Jong Han, Dubois, M.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 1997
• Subjects: Benchmark testing; Computer aided instruction; Computer buffers; Concurrent computing; Disk drives; Electronic equipment testing; Hardware; Neck; Parallel architectures; Scientific computing
• ISBN: 9780818682278; 0818682272
• DOI: 10.1109/ICPADS.1997.652595

The performance of parallel systems under commercial applications strongly depend on the speeds of IO subsystems and interconnects. Any single bottleneck in there hinder one from fully taking advantages of today's faster CPUs and better instruction architectures. Therefore, configuring bottleneck-free IO subsystems and interconnects is essential. In this paper we considered IO subsystem and interconnect in SPAX system. When the SPAX's IO subsystem is lightly equipped, the first bottleneck appears at disk drives. Although the disk bottleneck can be removed by putting more disks, another bottleneck was detected at the data buffer which is the first component in communication path. That is to say, all components in IO subsystems and interconnects must be sufficiently equipped, simultaneously, to thoroughly eliminate the bottleneck. We could reach a cost-effective bottleneck-free configuration of SPAX by making use of its flexible hardware design. The simulation results of SPAX then show significant performance gain from faster CPUs with better instruction architectures.