Message-Id: <36B584B2.1F3C3D8@dolphinics.com> Date: Mon, 01 Feb 1999 10:40:50 +0000 From: James Cownie Organization: Dolphin Interconnect Solutions Mime-Version: 1.0 Newsgroups: comp.parallel.pvm,comp.parallel.mpi Subject: Re: Switching HUB for beowulf cluster? References: <78osd7$9ji$1@news.kren.nm.kr> <78piue$lrv$1@pegasus.csx.cam.ac.uk> Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Xref: ukc comp.parallel.pvm:7956 comp.parallel.mpi:4572 Nick Maclaren wrote: > For a system of that size, using MPI, the hardware letency is > negligible compared with the software latency. Concentrate on > getting a tuned MPI implementation. > > |> 2. How much is the need for 100Mbps bandwidth for a beowulf cluster? > |> Will 10M bps only switching hub be ok for clustering or > |> will 100M bps switching hub lead to better overall performance? > > In terms of latency, there is no difference. In terms of bandwidth, > a lot. As the extra expense of 100M bps is now small, I advise > going for that. The reason to get a switching hub is to increase the bi-sectional bandwidth. As Nick says, it is unlikely to affect the point-to-point latency in a measurable way, but the extra bi-sectional bandwidth can have a significant effect on the performance of any application which uses broadcast, reduce, edge exchange etc... To see this you need a benchmark which measures the performance when all the nodes are communicating at the same time. -- Jim James Cownie Etnus, Inc. +44 117 9071438 http://www.etnus.com