[Beowulf] Re: next gen / parallel processing

Vincent Diepeveen diep at xs4all.nl
Mon Apr 3 15:06:47 PDT 2006


Well, doubling the bandwidth is great of course, but will it sell more machines?

Important is the price a socket. I tend to remember that the initial bare price of a node (12 sockets) is around 50000 dollar
and within 1 node you've got a one way pingpong latency of 1.5 us from cpu to cpu.

That's pretty expensive, considering that just the bare machine itself without too much RAM and without too much i/o
already gets very expensive.

A 2400 socket machine is soon then like 200 * 50k = 10 MLN dollar without additional network, without additional RAM
and with little i/o. 

So basically money talks too.

50k for 12 socket node at a great supercomputer isn't much, but you really want something then that is great everywhere.

Further the latency within 1 node is not exactly magnificent. The entire machine has a great latency from node to node,
but majority of scientists in first place run something at a cpu or 4-12 and want that application to perform fast. In general
that application simply can't handle 1.5 us latencies in such low number of cpu's.

IMHO doubling the bandwidth is nice on paper, but 90% of the jobs are jobs of just a few processors, if *those* run slow
then the machine simply doesn't get bought.

So theoretically spoken the improvement is great, practical i doubt it will change decision taking.

Vincent
  ----- Original Message ----- 
  From: Ed Karns 
  To: beowulf at beowulf.org 
  Sent: Monday, April 03, 2006 8:57 PM
  Subject: [Beowulf] Re: next gen / parallel processing


  " ... the HyperTransport Consortium is weeks away from launching HT 3.0, which is expected to at least double the bandwidth of that interconnect while lowering latency and leaving the underlying protocol largely unaltered. ..." From http://www.eetimes.com/news/latest/showArticle.jhtml?articleID=181502574 ... article on Cray's latest efforts.


  This next generation of parallel processing scenarios appears dovetail nicely with the Beowulf concept. Anyone looking into this? 



  Ed Karns

  FireWireStuff.com 





------------------------------------------------------------------------------


  _______________________________________________
  Beowulf mailing list, Beowulf at beowulf.org
  To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20060403/2a17d71d/attachment.html>


More information about the Beowulf mailing list