ohpc icon indicating copy to clipboard operation
ohpc copied to clipboard

Support for OpenMPI v4.0 + UCX

Open mtds opened this issue 5 years ago • 7 comments

Is there a plan to add OpenMPI version 4.0.x with UCX support? If not in the next minor upgrades (e.g. from 1.3.x to 1.4) maybe directly in the 2.0 tree?

I saw there is already an opened issue about Open UCX (#921) but I did not see any other opened issues about newer version of OpenMPI with UCX support.

According to the OpenMPI README released with version 4.0, support for Infiniband and RoCE through the openib plugin is deprecated and superseded by the ucx PML component (see line 740 of the README).

mtds avatar Oct 24 '19 12:10 mtds

Right now openmpi4 is targeted for the 2.0 release. There are a few more 2.0 highlights covered in the meeting notes from the last TSC meeting. I'm not sure specifically about UCX, but given that #921 has a target milestone of 2.0 I'm thinking that is a likely candidate.

amblakem avatar Oct 30 '19 23:10 amblakem

Hello, there's a way to keep OpenMPI 3 too in OpenHPC 2.0? I'm requesting this because of this issue: https://www-lb.open-mpi.org/software/ompi/major-changes.php

Legacy MPI-1 symbols that were deprecated in MPI-2.0 in 1996 and then deleted from the MPI-3.0 specification in 2012 are no longer prototyped in mpi.h by default. This will almost certainly cause some legacy MPI applications to fail to compile.

So the version can be selected through the modules framework and we can keep the compatibility with legacy code needing the removed symbols?

viniciusferrao avatar Jan 09 '20 17:01 viniciusferrao

Another option is to configure OpenMPI 4.x with --enable-mpi1-compatibility....

boegel avatar Jan 09 '20 17:01 boegel

I would prefer @boegel's suggestion over maintaining a separate build for symbols that were deprecated over 20 years ago.

koomie avatar Jan 09 '20 19:01 koomie

@boegel @koomie There's a warning about that in the link that @viniciusferrao posted.

... WARNING: These legacy symbols may disappear
 in a future major release series of Open MPI!.

IMHO, the less risky option would be to have both openmpi3 and openmpi4 in 2.0 release with a warning about it, and then drop the 3.x OpenMPI afterwards (e.g: OHPC 3.0)

fcannini avatar Jan 09 '20 19:01 fcannini

Thanks for pointing that out. I will bring this up for discussion at our next TSC meeting but I'm definitely in the camp that strongly thinks a 20 year warning is plenty...

Barring another motivating reason I really don't want to carry both the 3/4 variants in 2.0 out of the gate. I'm not against the --enable-mpi1-compatibility option for 4.x, but that may also just be kicking the proverbial can down the road ...

koomie avatar Jan 09 '20 22:01 koomie

@koomie I'm of the same opinion as yours, I just rather not be caught with my pants down when suddenly prof John Doe's software doesn't compile anymore. ;)

fcannini avatar Jan 10 '20 00:01 fcannini

Closing as 3.x switched to Open MPI 5

adrianreber avatar Mar 01 '24 11:03 adrianreber