[Beowulf] I'm so glad I didn't buy one of these
stewart at serissa.com
Thu Jul 3 08:24:19 PDT 2014
Regarding the side question of "what experience do folks have with HPCC"
I agree with the sentiment that HPL is pretty useless, but
I do think some of the others are quite useful.
The communications tests are fairly useful if read carefully.
PTRANS might tell you something about communications bandwidth,
but it might also just tell you about the cleverness of the
job layout scheduler.
RandomRing on the other hand, is fairly hard to fool, and will
be informative about bandwidth to expect with "everyone talking
RandomAccess is also quite instructive, as it is very sensitive to
end to end latency on a reasonably loaded network. The "SANDIA"
modification of a few years ago to RandomAccess obscures this a little
bit, but it is still very hard to have a system score that is anywhere
close to the scaled single processor score.
And as noted, FFT is just hard! It has a modest amount of computation, but
a great deal of communications, and the communication scales superlinear
with the job size.
Its kind of fun to take the Top500 list and resort it on each benchmark
to see how the order changes, and how different interconnects affect
the order. You have do some sort of mental corrections for system scale
HPC isn't about computation anymore, it is about communications and memory.
More information about the Beowulf