mindist mapper broken on master

[jsquyres]

A bunch of Mellanox Jenkins runs have been failing on master with this kind of error:

18:32:14 + /var/lib/jenkins/jobs/gh-ompi-master-pr/workspace-2/ompi_install1/bin/mpirun -np 8 --map-by dist -mca rmaps_dist_device mlx4_0 -x TEST_CLOSEST_NUMA -x TEST_PHYS_ID_COUNT -x TEST_CORE_ID_COUNT /scrap/jenkins/jenkins/jobs/gh-ompi-master-pr/workspace-2/jenkins_scripts/jenkins/ompi/mindist_test
18:32:16 
18:32:16 Success rank - 4: only one NUMA is scheduled.
18:32:16 
18:32:16 Success rank - 6: only one NUMA is scheduled.
18:32:16 
18:32:16 Success rank - 0: only one NUMA is scheduled.
18:32:16 
18:32:16 Success rank - 2: only one NUMA is scheduled.
18:32:16 
18:32:16 Error rank - 5: scheduled on wrong NUMA node - 1, should be 0
18:32:16 
18:32:16 Error rank - 3: scheduled on wrong NUMA node - 1, should be 0
18:32:16 
18:32:16 Error rank - 1: scheduled on wrong NUMA node - 1, should be 0
18:32:16 
18:32:16 Error rank - 7: scheduled on wrong NUMA node - 1, should be 0
18:32:17 -------------------------------------------------------
18:32:17 Primary job  terminated normally, but 1 process returned
18:32:17 a non-zero exit code. Per user-direction, the job has been aborted.
18:32:17 -------------------------------------------------------
18:32:17 --------------------------------------------------------------------------
18:32:17 mpirun detected that one or more processes exited with non-zero status, thus causing
18:32:17 the job to be terminated. The first process to do so was:
18:32:17 
18:32:17   Process name: [[34935,1],5]
18:32:17   Exit code:    1
18:32:17 --------------------------------------------------------------------------

In #1612 (comment), @jladd-mlnx reported that the PMIx external component changes (i.e., the changes from that PR) seem to be what broke this test.

@rhc54 Can you investigate?

[jladd-mlnx]

I see it in Brice's unmerged PR too. HWLOC change makes sense, but how has it crept into master? Don't see it on 2.x. Strange...

#1584

[jsquyres]

@jladd-mlnx Did you git bisect and find that the place where it came into master was the commit that came from #1612?

[jladd-mlnx]

No, just combing through closed PRs to see where it first showed up. It may have nothing to do with this commit. It shows up in Brice's unmerged HWLOC update - which would make more sense, but then I don't understand how it would have leaked in master. Not sure I will have time to chase much down today. If I get a few spare cycles, I can bisect.

[jsquyres]

Discussed on teleconf today: @rhc54 is going to have a look, but will need help in testing it because he doesn't have any HCAs.

[jsquyres]

@jladd-mlnx @Di0gen Can you temporarily disable the mindist test while we know that it is broken? It's making all other PR's fail the Mellanox Jenkins test. We can re-enable it when mindist gets fixed.

[jladd-mlnx]

Done. I plan to re-enable it in one week.

On Wed, May 4, 2016 at 6:58 AM, Jeff Squyres notifications@github.com
wrote:

@jladd-mlnx https://github.com/jladd-mlnx @Di0gen
https://github.com/Di0gen Can you temporarily disable the mindist test
while we know that it is broken? It's making all other PR's fail the
Mellanox Jenkins test. We can re-enable it when mindist gets fixed.

—
You are receiving this because you were mentioned.
Reply to this email directly or view it on GitHub
#1623 (comment)

[ggouaillardet]

@jladd-mlnx does mellanox Jenkins explicitly configure ompi with external pmix ?
if yes, then both external hwloc and libevent are required.
if not, then external pmix should not be built, and this is a bug.
@rhc54 do you agree with the last statement ?

[rhc54]

Yes, that would certainly be true. However, someone on the call looked back at the logs and claimed it started with a different PR, so there is some uncertainty here.