The open archive for STFC research publications

Full Record Details

Persistent URL http://purl.org/net/epubs/work/39945
Record Status Checked
Record Id 39945
Title Workstation clusters for scientific applications: an assessment of various networks
Abstract We report the results of running a variety of scientific application codes on a cluster of HP and a cluster of IBM workstations. The HP cluster comprises 4 HP720 machines linked via Ethernet and Fibre Distributed Data Interface (FDDI), while the IBM cluster comprises 3 IBM 530H workstations linked via Ethernet and IBM's proprietary Serial Optical Channel Converter (SOCC). The latency and bandwidth of each network has been measured. The former is found to be roughly equivalent (2-3msec) for the three different interconnects, although it decreases to about 1msec if the SOCC is used with IBM's optimized version of the Parallel Virtual Machine (PVM) message passing harness. The measured bandwidth of FDDI and SOCC are below 30% of their quoted peaks, which is attributed to limitation in the workstation/network interface. For application codes with a high compute to communicate ratio, FDDI or SOCC has little to offer over a dedicated Ethernet link. For applications requiring high bandwidth communications (e.g. two-electron integral transformation quantum chemistry) FDDI and SOCC are of considerable benefit. Applications requiring many short messages show some benefit from the increased bandwidth, but this is tempered by latency considerations.
Organisation CSE , CSE-CEG , CSE-CMSG , SERC
Funding Information
Related Research Object(s):
Licence Information:
Language English (EN)
Type Details URI(s) Local file(s) Year
Report DL Technical Reports DL-SCI-TM99T. 1994. DL-SCI-TM99T.pdf 1994