<div>Kilian,</div>
<div>I'm a noobie at the hardware issues but your machine at Sanford sounds like great bang/buck. Plainly economies of scale work both ways (I can't imagine wiring up that many nodes with ethernet). The factor of five thing sounds like a careless gloss of some numbers from some edition of the top500 lists, from some date, by INQ. Adams' site seems more modest. What I love about it is that it completely specs out what he did, very clearly. He gets an A for documentation, I'm not such a n00b at documentation :-)
</div>
<div>Peter<br><br> </div>
<div><span class="gmail_quote">On 8/31/07, <b class="gmail_sendername">Kilian CAVALOTTI</b> <<a href="mailto:kilian@stanford.edu">kilian@stanford.edu</a>> wrote:</span>
<blockquote class="gmail_quote" style="PADDING-LEFT: 1ex; MARGIN: 0px 0px 0px 0.8ex; BORDER-LEFT: #ccc 1px solid">On Friday 31 August 2007 09:37:49 am Peter St. John wrote:<br>> I saw at wiki <a href="http://en.wikipedia.org/wiki/Beowulf_%28computing%29">
http://en.wikipedia.org/wiki/Beowulf_%28computing%29</a><br>> that someone posted a link to<br>> <a href="http://www.calvin.edu/~adams/research/microwulf/">http://www.calvin.edu/~adams/research/microwulf/</a>, a nice description
<br>> of a "microwulf" at Calvin College. It has brief but useful<br>> descriptions of it's design, cost broken down by parts in the<br>> Manifest, and price/performance specs.<br>> If LLNL is an Epic maybe this is only a limmerick, but it's a witty
<br>> limmerick.<br><br>It's also been posted on the INQ:<br><a href="http://www.theinquirer.net/default.aspx?article=42050">http://www.theinquirer.net/default.aspx?article=42050</a><br><br>It's very interesting from the design point of view, but I'm less
<br>convinced by the price tag argument.<br><br>Stanford University recently purchased a 276 dual-socket quad-core HPC<br>cluster, which is capable of 20.6TFlops (peak) and actually achieved<br>15.5TFlops (it has been ranked #54 in the latest Top500 list).
<br><br>The cost per GFlops has been $109, which is indeed more expensive than<br>the microwulf's $94/GFlops, but still far from being "five times as<br>much", as claimed by the INQ article. Besides, the price included
<br>software (a commercial scheduler) and support, management capabilities,<br>50TB high-performance storage, an administration GigE network and a DDR<br>Infiniband interconnect.<br><br>I guess it's easier to lower the cost per GFlop on large scale clusters,
<br>but I'd just wanted to put this into perspective, and show that you<br>don't necessarily need to build your cluster from scratch to get<br>comparable price/performance ratios.<br><br>Cheers,<br>--<br>Kilian<br>
_______________________________________________<br>Beowulf mailing list, <a href="mailto:Beowulf@beowulf.org">Beowulf@beowulf.org</a><br>To change your subscription (digest mode or unsubscribe) visit <a href="http://www.beowulf.org/mailman/listinfo/beowulf">
http://www.beowulf.org/mailman/listinfo/beowulf</a><br></blockquote></div><br>
!DSPAM:46d86c18106467095717635!