[ucx-group] UCX support added to OpenMPI

Pritchard Jr., Howard howardp at lanl.gov
Wed Oct 21 12:32:10 EDT 2015


Hi George,

responses interweaved below

From: George Bosilca [mailto:bosilca at icl.utk.edu]
Sent: Wednesday, October 21, 2015 10:15 AM
To: Pritchard Jr., Howard
Cc: Yossi Itigin; Shamis, Pavel; ucx-group at elist.ornl.gov
Subject: Re: [ucx-group] UCX support added to OpenMPI

Howard,

I am a little bit concerned about your statement. From a community perspective (and I'm also sure that this is covered by our by laws) as the 2.0 release is still in infancy (basically not released yet), there is absolutely no way to prevent one of the project contributors from brining their contribution in. Obviously, this only holds as long as no technical issue is on the way.
[howardp] its not a question of not bringing it in to v2.x, but just in terms of timing which minor version of v2.x.  One of the main reasons for moving to the new release system – with minor numbers – was to help facilitate bringing in new features as they become ready.  It would be great to get it out in v2.0, but on the other hand, if under usage a bunch of bugs show up either in the ucx pml or ucx itself, I don’t think we want to hold up v2.0 just for this work.  It could get in to v2.1.

This brings me to my question: Why a low priority at 5 ? I know we have an ongoing discussion on the OMPI mailing list regarding the selection logic, but UCX is not the root of the problem. Thus no conditions should be attached to the acceptance of the UCX support in OMPI 2.x. Moreover, if Mellanox decide to promote UCX PML instead of the MXM, they should be allowed to handle the priority of their components the way they want (of course the community will appreciate any heads-up).

[howardp] well, the problem here is UCX has multiple tl’s that run over a variety of interconnects in addition to mlx5, etc.  It presents some of the same problems in this respect as we would have seen with ofi mtl if it had a higher priority.

Thanks,
  George.

PS: The above does not apply to any already released version of the code.




On Wed, Oct 21, 2015 at 11:46 AM, Pritchard Jr., Howard <howardp at lanl.gov<mailto:howardp at lanl.gov>> wrote:
Hi Yossi,

I think as long as the ucx pml keeps its low priority of 5,
I would be okay with it going in to v2.0, although the time
is pretty tight. I can’t speak for Jeff.

There may be some dependence on the outcome of a discussion
on the ompi devel mail list at the moment.

Howard


From: Yossi Itigin [mailto:yosefe at mellanox.com<mailto:yosefe at mellanox.com>]
Sent: Wednesday, October 21, 2015 8:30 AM
To: Shamis, Pavel

Cc: ucx-group at elist.ornl.gov<mailto:ucx-group at elist.ornl.gov>
Subject: Re: [ucx-group] UCX support added to OpenMPI

Yes, that’s the plan.

From: Shamis, Pavel [mailto:shamisp at ornl.gov]
Sent: Wednesday, October 21, 2015 5:24 PM
To: Yossi Itigin <yosefe at mellanox.com<mailto:yosefe at mellanox.com>>
Cc: ucx-group at elist.ornl.gov<mailto:ucx-group at elist.ornl.gov>
Subject: Re: [ucx-group] UCX support added to OpenMPI

Fantastic ! I guess from OMPI master it will make its way to v2.0 and v1.10 versions ?

Pavel (Pasha) Shamis
---
Computer Science Research Group
Computer Science and Math Division
Oak Ridge National Laboratory



On Oct 21, 2015, at 10:08 AM, Yossi Itigin <yosefe at mellanox.com<mailto:yosefe at mellanox.com>>
 wrote:

Hi,

I’m happy to inform that initial UCX support has been added to OpenMPI, with PRhttps://github.com/open-mpi/ompi/pull/1008
This adds support for both MPI (as pml) and SHMEM (as spml).
Thanks to everybody who took part and helped review the code.

--Yossi
_______________________________________________
ucx-group mailing list
ucx-group at elist.ornl.gov<mailto:ucx-group at elist.ornl.gov>
https://elist.ornl.gov/mailman/listinfo/ucx-group
To unsubscribe, send a blank email to ucx-group-unsubscribe at elist.ornl.gov<mailto:ucx-group-unsubscribe at elist.ornl.gov>


_______________________________________________
ucx-group mailing list
ucx-group at elist.ornl.gov<mailto:ucx-group at elist.ornl.gov>
https://elist.ornl.gov/mailman/listinfo/ucx-group
To unsubscribe, send a blank email to ucx-group-unsubscribe at elist.ornl.gov<mailto:ucx-group-unsubscribe at elist.ornl.gov>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://elist.ornl.gov/mailman/private/ucx-group/attachments/20151021/7a5ff67b/attachment-0001.html>


More information about the ucx-group mailing list