[gambit-list] Termite problems
Nicholas Walton
nwalton at mac.com
Tue Feb 19 15:38:08 EST 2008
Hi,
I've just started working with Termite to distribute the processing
for a large data mining system over a multi core CPU and I seem to be
hitting a few problems for which any advice is welcome.
Small scale examples seem to work well but once I scale up I hit
problems. I'm running on a 1Gig Mac using OS X 10.4 with the latest
distribution of Gambit and the current tarball from the Termite site
together with the slib library.
At first I built a two process client server example but it hung. The
problem turned out to be the array.scm file in the slib distribution
redefining the equal? procedure. Once the redefinition was removed all
worked well. So long as I used tags and not pids to identify one
remote process to another. The pid returned by remote-spawn is not the
same as that returned by the process calling (self).
I then moved on to trying a parallel version of map, akin to that
given at http://theschemeway.blogspot.com/2007/06/parallel-map-with-termite.html
. That worked fine for small simple examples but when I threw in a
larger example, evaluating forty instances of a k-means cluster
procedure in parallel in one process, it consistently dumps core but
not in any predictable manner. All forty threads kick off fine but
while some succeed others fail catastrophically and at no particular
point resulting in Gambit dumping core. I've run the code using map
and all was well, and the code has nothing in it which seems likely to
cause the problem. I've run the code both compiled and fully
interpreted and the behavior is the same.
First is there any changes to the Gambit build which might help
identifying where the fault is occurring? Secondly are there known
problems with Termite especially when spawning moderately large
processes.?
Many Thanks
Nick Walton
More information about the Gambit-list
mailing list