From chemistry-request@server.ccl.net Fri Mar 8 06:01:40 2002 Received: from mailrelay1.lrz-muenchen.de ([129.187.254.101]) by server.ccl.net (8.11.6/8.11.0) with ESMTP id g28B1d507121 for ; Fri, 8 Mar 2002 06:01:39 -0500 Received: from sun4.lrz-muenchen.de by mailrelay1.lrz-muenchen.de with ESMTP for chemistry@ccl.net; Fri, 8 Mar 2002 12:00:47 +0100 Received: from localhost (ui22204@localhost) by sun4.lrz-muenchen.de (8.8.8+Sun/8.8.8) with ESMTP id MAA09684 for ; Fri, 8 Mar 2002 12:00:45 +0100 (MET) X-Authentication-Warning: sun4.lrz-muenchen.de: ui22204 owned process doing -bs Date: Fri, 8 Mar 2002 12:00:45 +0100 (MET) From: Eugene Leitl X-X-Sender: ui22204@sun4.lrz-muenchen.de To: chemistry@ccl.net Subject: FYI: superlinear speedups in GROMACS Message-Id: MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII ---------- Forwarded message ---------- Date: Thu, 7 Mar 2002 23:24:49 -0500 From: 'Velocet' To: Jim Fraser Cc: beowulf@beowulf.org Subject: Re: DUAL CPU board vs 2 Single CPU boards: bang for buck? On Thu, Mar 07, 2002 at 12:45:26PM -0500, Jim Fraser's all... > Amazing. 1 cpu 63 ps/day, 2 cpu = 179 ps/day, sounds like a case of 1 + 1 = > 2.84 (ie. 63 + 63 = 179 (not 126)?) > Sounds fishy to me but I am unfamiliar with your application. Ask David VanDerSpoel, he'll tell you that they see superlinear speedups in Gromacs jobs quite often. Has to do with the cache, I'd guess something like with one CPU the job cant fit in L2 cache by itself, so lots of main memory accesses are required - when the job is split for 2 cpus, more of the job is in L2 caches, and despite the shared memory that LAM uses for communications for dual nodes, the overhead for that doesnt outweigh the advantage of having the job divided into more cache ram... works out well! :) It happens sometimes - they see 20-25% for Pentium (3? 4?) in speedups, but Athlons really kick ass for CERTAIN JOBS in gromacs (small jobs dont do this because one CPU is enough for the whole job, and MONSTER jobs may well be too big for the cache to give a big advantage - and once you go onto a network it may too slow to take advantage of directly (dotn know if people have scali numbers for gromacs 3.0 on Dual athlons >=1.2GHz) - I see >100% scaling per CPU with 4 nodes with GBE with that job (d.dppc) in gromacs 3.0 with 1.333 Ghz CPUs. [ im not sure this is all how it works, perhaps someone on the list can explain it. ] Go read the factoring hack by DJB I think it was that makes RSA style public key easier by a magnitude or so (couple weeks ago) - there are big speedups for that stuff due to multiple CPU cache usage... it seems somewhat related. /kc > Jim > > -----Original Message----- > From: beowulf-admin@beowulf.org [mailto:beowulf-admin@beowulf.org]On > Behalf Of Velocet > Sent: Thursday, March 07, 2002 11:00 AM > To: beowulf@beowulf.org > Subject: Re: DUAL CPU board vs 2 Single CPU boards: bang for buck? > > > On Thu, Mar 07, 2002 at 07:24:08AM -0500, Jim Fraser's all... > > I don't desire to start a flame but it seems to me that for computational > > intensive and memory intensive work that 2 singles are better and in most > > cases cheaper then a dual setup. The dual SMP systems out there now have > to > > fight for bandwidth along the same bus. The AMD chips while fast appear > to > > be starved for data when *big* memory jobs are running. Further I don't > see > > the cost benefits, if you actually dig into the "bang-for-buck" duals > never > > seem to win. > > Counter example: > > d.dppc on gromacs 3.0 calculation on a dual tyan 2460 board: > > 1 cpu 63 ps/day > dual cpu 179 ps/day > > cache, baby. > > now, if only we could get 8 way athlons! > > /kc > > > The only significant benefit I see from dual is less hardware and higher > > computing density, that may be reason enough to go with duals esp. if your > > application does not require a lot of memory fetches. > > > > > > Jim Fraser > > > > _______________________________________________ > > Beowulf mailing list, Beowulf@beowulf.org > > To change your subscription (digest mode or unsubscribe) visit > http://www.beowulf.org/mailman/listinfo/beowulf > > -- > Ken Chase, math@velocet.ca * Velocet Communications Inc. * Toronto, > CANADA > _______________________________________________ > Beowulf mailing list, Beowulf@beowulf.org > To change your subscription (digest mode or unsubscribe) visit > http://www.beowulf.org/mailman/listinfo/beowulf > -- Ken Chase, math@velocet.ca * Velocet Communications Inc. * Toronto, CANADA _______________________________________________ Beowulf mailing list, Beowulf@beowulf.org To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf