On Tue, 25 Aug 2009, Eric Niebler wrote:
Matthias Troyer wrote:
On Aug 25, 2009, at 3:15 PM, Eric Niebler wrote:
We have some MPI test results now. Matthias should have a look and report on the failures there. Then we can decide whether there are any showstoppers.
Marshal, can you make sure Matthias has access to your test results?
I saw some test results, and they are the failures I also sometimes see: the skeleton&contents tests sometimes fail, and sometimes work. This is the same as in the past and Doug Gregor had explained it as a problem of the underlying MPI implementation. Looking deeper into it would require more time than is available. These are not regressions but issues that have been around for a while.
Thanks, Matthias. This is good enough for me to feel comfortable moving ahead with the 1.40 release. Beman, Jeremiah, any objections?
I am able to reproduce the skeleton&content errors fairly consistently now, but I don't understand the mechanism enough to debug it. If you use "-mca btl tcp,self" on the mpirun command line (running the test by hand), the memory errors turn into test failures, at least for me. I think I need Matthias to help with debugging the problems. -- Jeremiah Willcock