On Apr 27, 2007, at 6:29 AM, Götz Waschk wrote:
I'm testing my new cluster installation with the hpcc benchmark and
openmpi 1.2.1 on RHEL5 32 bit. I have some trouble with using a
threaded BLAS implementation. I have tried ATLAS 3.7.30 compiled with
pthread support. It crashes as reported here:
http://sourceforge.net/tracker/index.php?
func=detail&aid=1708575&group_id=23725&atid=379483
If I link to the ATLAS version without pthread support hpcc runs fine.
I have a problem with Goto BLAS 1.14 too, the output of hpcc stops
before the HPL run, then the hpcc processes seem to do nothing,
consuming 100% CPU. If I set the maximum number of threads for Goto
BLAS to 1, hpcc is working fine again.
openmpi was compiled without thread support.
Did you configure with --without-threads or did you just not specify
any options? Specifying --without-threads is a promise that you will
never ever have any threads anywhere in your code, especially on
Linux. This is because we have to use our own memory manager code to
get around the memory pinning problem with Myrinet/GM and
InfiniBand. You might want to configure with --without-memory-
manager and see if that helps with your crashes.
Brian
