osc/pt2pt fails with multiple threads #2614

jjhursey · 2016-12-20T20:53:38Z

Testing with the osc/pt2pt component revealed multiple hangs and wrong answers when running with two threads. Each thread is working with it's own communicator copy of MPI_COMM_WORLD and their own private windows.

The test is here:

% mpicc -o x mt_1sided.c mt_1sided_td1.c mt_1sided_td2.c
% mpirun -host hostA,hostB -mca osc pt2pt -mca pml ob1 -mca btl openib,self,vader ./x

PR #2630 will need to be reverted when a resolution to this issue is committed on v2.0.x branch.

The text was updated successfully, but these errors were encountered:

As a workaround for issue open-mpi#2614 for the v2.0.2 release, do not allow for selection of the OSC PT2PT when creating an MPI RMA window. Print a hopefully helpful message and return an not-supported error. This PR should be reverted once a fix for open-mpi#2614 is in place. Signed-off-by: Howard Pritchard <[email protected]>

hjelmn · 2017-01-07T00:10:00Z

I can't get this to hang with osc/rdma. @markalle Was the 1sided.c test hanging with osc/rdma or just osc/pt2pt?

This commit fixes a bug in the timer check. When -fPIC is used we need to save/restore ebx. The code copied from patcher was meant for 32-bit systems and did not work correctly on 64-bit systems. This commit updates the save/restore to use rbx instead of ebx. Fixes open-mpi#2614 Signed-off-by: Nathan Hjelm <[email protected]>

hjelmn · 2017-01-07T01:52:10Z

Opps. Wrong bug :). Deleting those.

hjelmn · 2017-01-09T18:40:17Z

Ok, I can reproduce the issue with osc/pt2pt. Looks like something is still not right with PSCW. Taking a look now.

hjelmn · 2017-01-09T22:17:30Z

Found the bug. Its an artifact of the original design. I haven't had the time to move the counters into the sync object. Think I have a workaround. Testing it now.

hjelmn · 2017-01-10T00:20:36Z

Ok, definitely fixed. Running it through the tests a couple more times to shake out any remaining bugs. Will have a PR open for master, v2.x, and v2.0.3.

hjelmn · 2017-01-11T15:54:16Z

Still have one stubborn bug holding up the fix. It now hangs about 10% of the time. Will hopefully have this resolved today.

hppritcha · 2017-05-22T11:12:03Z

@hjelmn can this issue be closed?

markalle · 2017-06-19T21:22:30Z

I had been meaning to re-evaluate this one for a while. I built a vanilla OMPI v2.0.x with --enable-mpi-thread-multiple, and my results were:

mpirun -mca osc rdma -host hostA:2 ... : passed
mpirun -mca osc rdma -host hostA:4 ... : passed
mpirun -mca osc rdma -host hostA:1,hostB:1 ... : segv quickly
mpirun -mca osc pt2pt ... : message that "OSC pt2pt component does not support MPI_THREAD_MULTIPLE in this release"

hppritcha · 2017-07-12T14:37:47Z

@gpaulsen will check to see if we have more data on this issue. @hjelmn says he'll work on it.

hjelmn · 2017-09-14T21:26:42Z

Finally found the time to track this down. I see what is going wrong in osc/pt2pt. I should have a fix ready for testing tomorrow.

hjelmn · 2017-09-20T14:51:32Z

Damn this is a nasty bug. Its getting a lot further but now I am running into another issue. Will keep cranking away at it until I know what the root cause is. Will start a parallel effort to ensure osc/rdma passes next month.

gpaulsen · 2017-09-26T16:00:25Z

@hjelmn, Is this issue specific to osc_pt2pt or osc progress / multithreaded in general and only tickled by osc_pt2pt?

jsquyres · 2018-03-26T19:50:11Z

Per discussion with @hjelmn and @hppritcha on 26 Mar 2018:

v2.0.x already disables osc/pt2pt in THREAD_MULTIPLE scenarios (because of this issue)
v2.x does not disable itself in THREAD_MULTIPLE scenarios
- It looks like we thought this would get fixed in time for v2.1.x. But it didn't.
At this point, it doesn't seem worth fixing this issue in the v2.x series.
@hjelmn said he will fix in the v3.0.x series.
- So I'll re-target this issue for v3.0.x.
I will PR cherry-pick d0ffd66 to the v2.x branch (i.e., disable osc/pt2pt in THREAD_MULTIPLE scenarios)

As a workaround for issue open-mpi#2614 for the v2.0.2 release, do not allow for selection of the OSC PT2PT when creating an MPI RMA window. Print a hopefully helpful message and return an not-supported error. This PR should be reverted once a fix for open-mpi#2614 is in place. Signed-off-by: Howard Pritchard <[email protected]> (cherry picked from commit d0ffd66) The original commit message is shown above. Followup: as of this writing (26 Mar 2018), we do not plan to fix this issue for the v2.0.x or v2.x. Hence, the osc/pt2pt component will continue to disable itself in THREAD_MULTIPLE scenarios for the life of all v2.x series. It is possible (likely?) that this will be fixed in a v3.0.x release (where x>1).

jsquyres · 2018-05-29T15:17:01Z

Per discussion 2018-05-29: @hjelmn says that a better solution would be to get everything to support osc/rdma. E.g., get vader, TCP, and the upcoming OFI BTLs to support RDMA, which then works with osc/rdma.

So here's the way forward:

Disable osc/pt2pt in 3.0.x and 3.1.x (just like we did in all of 2.x)
Add put/get support in vader, and (upcoming) OFI BTL
Eventually delete osc/pt2pt (perhaps after v4.0.x series? Would be good to keep it around just until all the new osc/rdma support in the BTLs solidifies/matures)

As a workaround for issue open-mpi#2614 for the v2.0.2 release, do not allow for selection of the OSC PT2PT when creating an MPI RMA window. Print a hopefully helpful message and return an not-supported error. This PR should be reverted once a fix for open-mpi#2614 is in place. Signed-off-by: Howard Pritchard <[email protected]> (cherry picked from commit d0ffd66)

Per discussion at open-mpi#2614 (comment), do not allow for selection of the OSC PT2PT when creating an MPI RMA window when THREAD_MULTIPLE is active. Print a helpful message and return a not-supported error. Signed-off-by: Howard Pritchard <[email protected]> Signed-off-by: Jeff Squyres <[email protected]> (cherry picked from commit d0ffd66)

Per discussion at open-mpi#2614 (comment), do not allow for selection of the OSC PT2PT when creating an MPI RMA window when THREAD_MULTIPLE is active. Print a helpful message and return a not-supported error. Signed-off-by: Howard Pritchard <[email protected]> Signed-off-by: Jeff Squyres <[email protected]> (cherry picked from commit d0ffd66) Signed-off-by: Jeff Squyres <[email protected]>

matcabral · 2018-05-29T17:50:37Z

Eventually delete osc/pt2pt (perhaps after v4.0.x series? Would be good to keep it around just until all the new osc/rdma support in the BTLs solidifies/matures)

Currently psm and psm2 MTLs use the osc/pt2pt understanding the limitations. There is a workaround using the OFI path, but still a workaround.

jsquyres · 2018-05-29T17:57:33Z

@matcabral I think @hjelmn's plan is to make all the BTL's support put/get, and then all transports can use osc/rdma -- therefore the need for osc/pt2pt can go away. That would be the purpose of the OFI BTL -- just for one-sided. Make sense?

matcabral · 2018-05-29T18:17:38Z

@jsquyres, yes, thanks for the clarification.

hjelmn · 2018-05-29T19:01:28Z

I expect all transports that do not currently work with osc/rdma will see a performance improvement. How large will depend on the AMO and RDMA implementations.

Per discussion at open-mpi#2614 (comment), do not allow for selection of the OSC PT2PT when creating an MPI RMA window when THREAD_MULTIPLE is active. Print a helpful message and return a not-supported error. Signed-off-by: Howard Pritchard <[email protected]> Signed-off-by: Jeff Squyres <[email protected]> (cherry picked from commit d0ffd66) Signed-off-by: Jeff Squyres <[email protected]> (cherry picked from commit 5b7c866)

Per discussion at #2614 (comment), do not allow for selection of the OSC PT2PT when creating an MPI RMA window when THREAD_MULTIPLE is active. Print a helpful message and return a not-supported error. Signed-off-by: Howard Pritchard <[email protected]> Signed-off-by: Jeff Squyres <[email protected]> (cherry picked from commit d0ffd66) Signed-off-by: Jeff Squyres <[email protected]> (cherry picked from commit 5b7c866)

jjhursey added bug Severity: critical Target: v2.0.x Target: v2.x labels Dec 20, 2016

jjhursey added this to the v2.0.3 milestone Dec 20, 2016

jjhursey assigned hjelmn, jjhursey and markalle Dec 20, 2016

jjhursey mentioned this issue Dec 20, 2016

osc/pt2pt multithread fixes #2615

Closed

gpaulsen mentioned this issue Dec 21, 2016

Topic/osc controls #2626

Closed

hppritcha mentioned this issue Dec 22, 2016

osc/pt2pt: workaround for issue 2614 for v2.0.2 #2630

Merged

hppritcha modified the milestones: v2.0.3, v2.0.4 Jun 1, 2017

hppritcha modified the milestones: v2.0.4, v2.1.3 Sep 19, 2017

jsquyres removed the Target: v2.0.x label Sep 19, 2017

jsquyres added the Target: v3.0.x label Sep 19, 2017

gpaulsen assigned nysal Sep 20, 2017

hppritcha modified the milestones: v2.1.3, v2.1.4 Mar 15, 2018

jsquyres modified the milestones: v2.1.4, v3.0.2 Mar 26, 2018

jsquyres mentioned this issue Mar 26, 2018

v2.x: osc/pt2pt: disable osc/pt2pt for THREAD_MULTIPLE scenarios #4972

Merged

jsquyres removed the Target: v2.x label May 29, 2018

jsquyres mentioned this issue May 29, 2018

osc/pt2pt: disable when THREAD_MULITPLE #5200

Merged

jsquyres closed this as completed in #5200 May 29, 2018

This was referenced May 29, 2018

v3.0.x: osc/pt2pt: disable when THREAD_MULTIPLE. #5201

Merged

v3.1.x: osc/pt2pt: disable when THREAD_MULTIPLE. #5202

Merged

dalcinl mentioned this issue Mar 16, 2021

no mpiexec: mpi4py.MPI.Exception: MPI_ERR_WIN: invalid window mpi4py/mpi4py#34

Closed

jkbk2004 mentioned this issue Oct 19, 2023

Update to spack-stack-1.5.0 (update to gocart hash f7b34e) ufs-community/ufs-weather-model#1920

Merged

43 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

osc/pt2pt fails with multiple threads #2614

osc/pt2pt fails with multiple threads #2614

jjhursey commented Dec 20, 2016 •

edited by hppritcha

Loading

hjelmn commented Jan 7, 2017

hjelmn commented Jan 7, 2017

hjelmn commented Jan 9, 2017

hjelmn commented Jan 9, 2017

hjelmn commented Jan 10, 2017

hjelmn commented Jan 11, 2017

hppritcha commented May 22, 2017

markalle commented Jun 19, 2017

hppritcha commented Jul 12, 2017

hjelmn commented Sep 14, 2017

hjelmn commented Sep 20, 2017

gpaulsen commented Sep 26, 2017

jsquyres commented Mar 26, 2018

jsquyres commented May 29, 2018 •

edited

Loading

matcabral commented May 29, 2018

jsquyres commented May 29, 2018

matcabral commented May 29, 2018

hjelmn commented May 29, 2018

osc/pt2pt fails with multiple threads #2614

osc/pt2pt fails with multiple threads #2614

Comments

jjhursey commented Dec 20, 2016 • edited by hppritcha Loading

hjelmn commented Jan 7, 2017

hjelmn commented Jan 7, 2017

hjelmn commented Jan 9, 2017

hjelmn commented Jan 9, 2017

hjelmn commented Jan 10, 2017

hjelmn commented Jan 11, 2017

hppritcha commented May 22, 2017

markalle commented Jun 19, 2017

hppritcha commented Jul 12, 2017

hjelmn commented Sep 14, 2017

hjelmn commented Sep 20, 2017

gpaulsen commented Sep 26, 2017

jsquyres commented Mar 26, 2018

jsquyres commented May 29, 2018 • edited Loading

matcabral commented May 29, 2018

jsquyres commented May 29, 2018

matcabral commented May 29, 2018

hjelmn commented May 29, 2018

jjhursey commented Dec 20, 2016 •

edited by hppritcha

Loading

jsquyres commented May 29, 2018 •

edited

Loading