Top 500 trends

John Brookes johnb at quadrics.com
Wed Nov 27 05:10:26 PST 2002


Group/departmental clusters are without question a Good Thing (TM), but
don't be so down on building now or building the big 'uns:

1) If you knew exactly how to find the solution to problem X, it wouldn't
really be research. The corollary of that, of course, is that there exists
the possibility of waiting seven years to do something stupid and/or
pointless and have no time to rectify false assumptions etc. Result: only
half as many papers ;)

2) There seems to be an implicit assumption on this thread that clusters are
used by single groups. On the contrary, large machines may have thousands of
users - many of them neither knowledgable about, nor interested in, anything
more than how to turn the handle on their current project. Does it make
sense to give those groups time on some - inevitably large - central
resource that has staff to help with the problems that they will likely
have, or to give each of them $30k for something they don't (want to) know
how to build and that they may only need for a few weeks out of every year?

3) Cycles aren't everything. Some jobs aren't cycle-hungry (relatively at
least), but the only meaningful datasets are so vast that constructing a
beowulf to accomodate them would be impractical.

and obviously:

4) Big machines look cool - particularly ones with lots of blinkenlights.
"Oooh, shiny thing!"


Of course, I _like_ people buying BIIIG clusters :)

John Brookes
Quadrics
T: +44 (0)117 9155500
F: +44 (0)117 9075395
E: johnb at quadrics.com
W3: www.quadrics.com




More information about the Beowulf mailing list