Closed mdavis36 closed 2 years ago
Hi mdavis36, we can provide a disable option for this (or perhaps and enable option if we change the defaults). This was necessary to support codes which relied on cuda aware mpi. But this should be selectable by the project depending on the needs. I'll work this into the next release.
Can you tell me which version of ATS you are using (ie the older python 2 or the current python 3 version)
We are unfortunately still on python2.7 (hoping to transition to Python3 in the near future) and are using ATS 7.0.5
thank you.
will work on an update to the python2 branch with this option perhaps by next week will also ensure it gets into python3 branch
Hi @dawson6 what is the status of this, it looks like #81 is the fix for this?
Hi mdavis. I have been testing David's updates to the main branch (python 3)(which would include this) and there are issues with that still, so can not make a release.
HOWEVER, as you are using the python2 branch I'll get that fix into the python 2 branch and get a release for that out. Will work on that this afternoon, update the public ATS versions and make a tarfile available in case you are installing ATS yourself. I can get that released on the CZ and the RZ, but will need someone on site this week to get the public ATS updated on the SCF. My usual person for doing that is gone this week I think, but I'll look around.
update for @mdavis36 am actively working on this.
The current python2 coding is as follows, that is it is alway on.
There are codes which do need this in order to run. But we should have way to disable it.
Hesitant to change this default, but will put in the option to disable this
# with MPI that it should be enabled.
str_smpi = "--smpiargs=\"-gpu\""
HI @mdavis36 Could you test the 7.0.9 install on rzansel, and give it one of these two options. Either will disable the smpiargs=-gpu option and replace it with --smpiargs=off or smpiargs=-show
--smpi_off Blueos option: Add --smpiargs=off to the lrun/jsrun line. Disables --smpiargs=-gpu --smpi_show Blueos option: Add --smpiargs=show to the lrun/jsrun
@dawson6 --smpi_off seems to do the trick for what we need. Thank you for getting this pushed through for us!
Our code is running into problems running gpu tests on blueos systems. This seems to be a result of ATS throwing
--smpiargs=-gpu
in thejsrun
command. Our tests pass when using--smpiargs=off
or--smpiargs=disable_gpu_hooks
however there doesn't seem to be a way to change this with ATS as far as I can see.