newbie: 16-node 500Mbps design

Laurent Itti itti at cco.caltech.edu
Mon Aug 21 16:26:46 PDT 2000


Hi all:

thanks again for all the very good comments. I'll summarize a few
highlights:

ethernet:
---------
- "check out www.netperf.org for benchmarks"
- "DEC Tulip chips seem to have a very good reputation re. throughput"
- "the 3com (or any tulip) does full busmastering; the 8139
   sort of does, but has limitations that require the OS to copy
   the packet much of the time.  which is not necessarily a problem,
   since people have been doing copy-and-checksum in one step for years.
   the 3c9xx's, though, offload both the copy and the checksum."

I see; that indeed can justify the price premium.


chipsets:
---------
- "in general, chipsets don't require "support" since they're usually
  quite back-compatible."

But some chipsets do have explicit support in the kernel... and I guess it
makes a big difference whether you access your disk via DMA
("supported chipset") or PIO ("fallback if chipset is unsupported").

- "ATA-100: why want it, current disks max out at 40MB/s?"

then, when I add more drives in a year or two, I won't be limited by the
bus speed?  (2*40 > 66; but I agree that it's a bit overkill; just since
the price premium is low, I'll give it a try).

- "forget gentus; they're just hangers-on.  you can get the official stuff
from linux-ide.org or kernel.org"

thanks!

- "Abit will have a new dual-coppermine M.B. soon (whatever that means)."

sounds good, I'll try it for the head of the system.

buying linux:
-------------
- "[buying one copy/node for retail Mandrake costs] less than 5% of total
  machine cost.  The idea was to say "thank you" for the great OS.
  But saying "thank you" to mandrake is a bit silly, since they're 
  just packagers."

well, yes and no.  Most of the time in trying to design the system I have
spent in trying to match chipsets to drivers.  Ultimately, I don't see why
we should waste so much volunteer effort on hacking out sub-optimal linux
drivers from partial docs. There are many more interesting programming
challenges around. In the future, I want fully-optimized linux support for
all hardware I buy -> only manufacturers can provide that. Packagers are
the ones who increase general linux awareness so that manufacturers
finally start realizing that they must provide Linux drivers...  with so
many boxes for RedHat and Mandrake in stores everywhere, it's getting more
and more difficult to ignore them.

just my 2 cents but I don't want to get too much off-topic.

Thanks a lot to all, I'll post some benchmarks once we have the machine
running ;-)

  -- laurent

with contributions from

Scott L. Delinger <Scott.Delinger at ualberta.ca>
Gerry Creager N5JXS <gerry at cs.tamu.edu>
Mark Hahn <hahn at coffee.psychology.mcmaster.ca>
Ioannis F Sotiriadis <admin at elcomsoft.gr>






More information about the Beowulf mailing list