From: Lars Rzymianowicz Newsgroups: comp.parallel.mpi Subject: Re: multi-cpu machines and mpi Date: Wed, 19 Aug 1998 09:22:59 +0200 Organization: Dept. of Computer Engineering, University of Mannheim, Germany Message-Id: <35DA7D53.8B7F6B24@mufasa.informatik.uni-mannheim.de> References: <6rcen0$qju$1@node17.cwnet.frontiernet.net> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit don schad wrote: > [...] The question came up > as to whether we should purchase (N/2) two-processor > intel machines or N single-processor machines. [...] A general rule seems to be: If you build a distributed cluster, make the nodes as strong as possible. Of course, that's also a price/performance issue. Dual-processor nodes, i think, have a reasonable price. I don't know exactly, if Quad-processor boards still deliver enough performance and bandwidth to justify their price. Of course, the message passing libs must make use of the locality. E.g., the MPICH ch_p4 communication device is in fact a multi-protocol device. It can be configured (-comm=shared option) to use TCP/IP between nodes and shared memory within nodes. We have 2 dual-cpu PC's running this configuration, it works fine. This is especially useful, if your nodes are interconnected with a low-performance network like (Fast-)Ethernet. The step from one- to dual-cpu nodes unloads 50% of the communication load from the interconnect, in average. A lot of clusters (e.g. ASCI RED, 4500+ dual-PPro nodes) are installed this way. Question to all others: are there any good quad-cpu boards around? I think, i read in the PII manual, that its bus interface can not handle more than 2 cpu's in an SMP configuration. Same for AMD, Cyrix, etc. The PPro can, but is dead. So that might be the reason, why quad-boards are not a matter today. And Merced, McKinley, ..., we have to wait. How about Alpha? A cluster of 4-cpu Alpha nodes, wow, that sounds good ;-) Lars -- email: lr@mufasa.informatik.uni-mannheim.de Homepage: http://mufasa.informatik.uni-mannheim.de/lsra/persons/lars/