From: "Alain Coetmeur" Newsgroups: comp.parallel.pvm,comp.parallel.mpi Subject: Re: HW cluster design (newbie) Date: Tue, 18 May 1999 12:07:45 +0200 Organization: Informatique-CDC Message-Id: <7hre89$d9u2@puligny.idt.cdc.fr> References: <37405145.314147FD@si.unirioja.es> <7hqft3$1q8e@news2.newsguy.com> Mime-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 8bit Xref: ukc comp.parallel.pvm:8397 comp.parallel.mpi:5102 Greg Lindahl a écrit dans le message <7hqft3$1q8e@news2.newsguy.com>... >Eduardo Bergasa writes: >> 1) switched Fast ethernet >> 2) Myrinet LAN for our cluster it would habe double the price. valuable if I/O are the botteleneck since it is 5-10 times faster for bandwidth and even more (with good MPI implementation) for latency >[...] >> Can you give me any recomendation or advice? Is the performance >> improvement (using myrinet) worth the cost difference? >Benchmark your codes. If you don't like your performance and >scaling with fast ethernet, then maybe Myrinet is right for your >applications. There is no general answer. that is it. my only 0.05 Euro advice : the most important in your benchmark is to know the number of synchronisation and megabytes by megaflop for example one of my problem is 60Mflop/s with 2kb and 2 sync by node transmited every 500s . the problem is solved after 150Tflop, which should take less than 48h. with this you can check that it easily supports 200us latency and 80Mb/s bandwidth on our 8 bi-P2 beowulf. as an exercise try to imagine how many pentium 2 (at 60Mflop) I need which BW and latency I can tolerate... I'm sure it would pass through ISDN, e-mail, or even optic telegraph ! do the same for your own problems. maybe are there no real machine that can do the job, maybe only SMP or supercomputers, maybe myrinet only... probably is it possible to reduce sync or data xfer by a few power of ten ? look after the "BSP: Bulk synchrnous model" modeling technic. it will show you how to build a model of your algorithm (my opinion is not to use it as a programming technic) so that you can check what latency and what bandwidth you need to avoid I/O bottleneck. anyway it can give you idea or justification to change the algorithm, avoid sync or data transfer. note that it is a very FAQ, but if this question have survived so long it is because there is no definite answer. look in the dejanews or alike for similar discussion