Actually, there are some action items left. For example, some of the paths are hard-coded in there. -- Pavan On 10/05/2007 11:22 AM, Rajeev Thakur wrote:

I think you can just cut a new version and announce it on mpich-discuss.

Rajeev

...

-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of William Gropp Sent: Friday, October 05, 2007 11:21 AM To: [email protected] Subject: Re: [mpich2-core] [Fwd: Re: [MPICH] Bug in rings test of the MPI Intel Test suite ?]

Eek! Yes, we should. Are there any pending items, or should I just cut a new version?

Bill

On Oct 5, 2007, at 10:58 AM, Pavan Balaji wrote:

...

I think it's time we work on the release of the 1.1 version of the Intel test suite. The "latest" version out there is several years old and there have been many bug fixes since then. Thoughts?

-- Pavan

-------- Original Message -------- Subject: Re: [MPICH] Bug in rings test of the MPI Intel Test suite ? Date: Fri, 5 Oct 2007 08:41:24 -0500 From: William Gropp <[email protected]> To: Sylvain Jeaugey <[email protected]> CC: [email protected] References: <[email protected]>

Thanks for the note. This is fixed in our most current version; do you have the latest version of the Intel suite (there have been many minor but important fixes like this, particular for bugs that show up on 64 bit platforms)?

Bill

On Oct 4, 2007, at 10:46 AM, Sylvain Jeaugey wrote:

...

Hi all,

I think I just found a bug in the "rings" test of the Intel Test Suite. I didn't find a dedicated mailing list, so I'm posting here.

At line 231 of node.c, we call : ierr = MPI_Bcast(&seed, 1, MPI_LONG, 0, MPI_COMM_WORLD);

However, seed is an "int". On itanium, sizeof(long) is 8 whereas sizeof(int) is 4. This causes the Bcast of seed as well as .... the lower 4 bytes of the rings pointer. In my case, process 0 and process 1 don't have the same pointer for rings, hence causing quickly a segmentation fault.

I changed the type to MPI_INTEGER : ierr = MPI_Bcast(&seed, 1, MPI_INTEGER, 0, MPI_COMM_WORLD); and everything works fine.

Sylvain

-- Pavan Balaji http://www.mcs.anl.gov/~balaji

-- Pavan Balaji http://www.mcs.anl.gov/~balaji

Thanks! I was trying to look for this email, but couldn't find it. We'll need some discussion about the non-implementation specific job compilation/execution options, though. -- Pavan On 10/05/2007 01:38 PM, William Gropp wrote:

On Oct 5, 2007, at 11:44 AM, Pavan Balaji wrote:

...

Actually, there are some action items left. For example, some of the paths are hard-coded in there.

-- Pavan

Here's a list I found in my mail from Pavan. Are there any others? Note that (5) needs to be modified for systems that don't have compilation scripts; however, I agree that eliminating implementation-specific options is a good goal (but note also that mpiexec isn't universally available :( ).

Bill

...

Just making a note of the fixes that I think are necessary for the Intel test suite for your comments. I can make these changes myself once my current urgent task list is done:

1. Add a few testlists with descriptive names (e.g., complete.list, io.list, c.list, fortran.list) and get rid of all the current testlists that are available, but with very non-descriptive names.

2. Decouple compilation and testing. Currently "make" runs the tests too.

3. Add the default installation path to /usr/local or something. It currently points to one of the MCS machine directories. The user can, of course, specify any other directory with prefix.

4. Make it easier to run individual tests. I could not get this working as yet, without creating a separate testlist and modifying the Makefile.

5. Get rid of the --with-mpich2 and equivalent options. The test suite should not depend on any supported MPI implementation. It should just use "mpicc" for compilation. We can have some option for the job launcher, e.g., mpiexec, mpirun, etc.

6. Allow user to specify a timeout value for each test. If nothing is specified, the default timeout is used.

-- Pavan Balaji http://www.mcs.anl.gov/~balaji