Kurt, > I have put together a �Cluster that addresses some of these concerns.
Very interesting, yours is a more solid-engineering type of approach, not a surprise considering you address!... > I eschewed latency concerns in favor of bandwidth ones and implemented a > channel-bonding scheme much like the one used at http://ilab.usc.edu/beo. How does bonding of 100 Mbps cards compare with Gigabit Ehternet these days? The cards are not all that expensive anymore, but I guess the network equipment still is. We got a 3Com server card for about US$ 500 some time ago, but I hear there are some going for less than US$ 200 now. > I agree with many of Jorge's suggestions (I ran his page through the URL > translator at worldlingo.com) If this worked well, Andrew's translation job may already be half done! > and for raw heat-spewing, number-crunching, beowulf-thrashing > performance, AMD CPUs are the way to go... Yes, and for having a bit of heat trouble too. One of the nodes has been bothering us sporadically. So we have 50 large fans in the buying queue. > ...now if only we could get the data to and from the AMD MoBos faster! Well, speaking of latency, I understand there are two major causes: the kernel's TCP stack and the switches. With wire-speed switches I presume the former is the main one these days. I hear there is a public drive somewhere for 3Com Gigabit cards for using them without going through the kernel networking layer. Does anybody know anything about this? If true it could be a big plus for reducing latency. Cheers, ---------------------------------------------------------------- Jorge L. deLyra, Associate Professor of Physics The University of Sao Paulo, IFUSP-DFMA For more information: finger [EMAIL PROTECTED] ----------------------------------------------------------------
