[mvapich-discuss] xcbrd tests

[Shaun Rowland]

Bas van der Vlies wrote:

First, thank you for mentioning that you were using BLAS instead of
ATLAS. That is much easier to build and deal with :-) I had not tried
this previously.

> I have done some more tests only with mvapich2 0.9.8 and build 
> BLACS/SCALAPACK without any optimization (-O0 -g) for gcc version 3.4.6 
> and 4.1.1. For fortran gfortran 4.1.1 is used
> 
> I can run all tests with gcc version 3.4.6 ;-)
> With gcc version 4.1.1: 2 tests do not finish:
>  1) xzgsep
>  2) xzllt
> 
> We want to use MVAPICH2 so i did not test MVAPICH1 today,

I am not sure what is going on with your gcc 4.1.1 attempts. Since you
sent this, I built gcc 4.1.1 and 4.1.2 on the systems I am using. I just
built MVAPICH2 0.9.8, BLACS, BLAS, and ScaLAPACK using the gcc 4.1.1
build, since that matches the default version you were trying to use. I
ran all 102 ScaLAPACK tests from my build, and there were no hangs or
errors.

Could you give me some more details about how you are running these:

- which architecture (IA32, EM64T, X86_64 [Opteron])?
- how many processes per node?
- are you setting any specific MVAPICH2 runtime environment variables?

I've been testing with 4 processes on 2 nodes.  Each node has dual
dual-core Opteron 270 processors. I have not tried MVAPICH with BLAS in
place of ATLAS yet. I did not change any of the default optimization
levels, so this should be a mix of -O3 and -O4.

Also, the version of MVAPICH2 I am using has now been made into a new
"p1" release tarball, which you can get from our download page for
MVAPICH2. There were some updates in the 0.9.8 release branch, and the
new tarball has those - and this corresponds to the SVN repository
version I mentioned previously. This would be the latest MVAPICH2
release available.
-- 
Shaun Rowland	rowland at cse.ohio-state.edu
http://www.cse.ohio-state.edu/~rowland/