gpu-margo-p2p-bw assertion on thetagpu

[carns]

I can compile and run the thetagpu regression test with the latest version of spack after the updates in #58 and mochi-hpc-experiments/platform-configurations#18, but the test fails at runtime with the following:

gpu-margo-p2p-bw: ../perf-regression/gpu-margo-p2p-bw.cu:275: int main(int, char**): Assertion `ret == 0' failed.

(that indicates a failure of margo_bulk_create_attr()). I assume this is the libfabric problem where it looks like the libfabric library itself needs to be explicitly configured with gpu memory support?

@vchoi-hdfgroup @jhendersonHDF

[vchoi-hdfgroup]

Hi Phil, Yes it is. It can be fixed by configuring libfabric with cuda support. See the info in perf-regression/theta-gpu/README.md in this PR on how to do this: HDFGroup#4 Vailin From: Phil Carns ***@***.***> Date: Sunday, November 27, 2022 at 10:48 AM To: mochi-hpc-experiments/mochi-tests ***@***.***> Cc: Vailin Choi ***@***.***>, Mention ***@***.***> Subject: [mochi-hpc-experiments/mochi-tests] gpu-margo-p2p-bw assertion on thetagpu (Issue #59) I can compile and run the thetagpu regression test with the latest version of spack after the updates in #58<#58> and mochi-hpc-experiments/platform-configurations#18<mochi-hpc-experiments/platform-configurations#18>, but the test fails at runtime with the following: gpu-margo-p2p-bw: ../perf-regression/gpu-margo-p2p-bw.cu:275: int main(int, char**): Assertion `ret == 0' failed. (that indicates a failure of margo_bulk_create_attr()). I assume this is the libfabric problem where it looks like the libfabric library itself needs to be explicitly configured with gpu memory support? @vchoi-hdfgroup<https://github.com/vchoi-hdfgroup> @jhendersonHDF<https://github.com/jhendersonHDF> — Reply to this email directly, view it on GitHub<#59>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ANF3JBFABIAJNKACYSZODTLWKOGE3ANCNFSM6AAAAAASMTGNRQ>. You are receiving this because you were mentioned.Message ID: ***@***.***>

[carns]

Great, thanks! I'll leave this issue open for now just to track this topic; we can close it once the fix is in place here. I put a comment on that other PR; it would be great if you could contribute the new variant directly to the mochi-spack-packages repo.