Follow-up from "Default to using direct GPU communication with tMPI"
The following discussion from !2258 (merged) should be addressed:
-
@al42and started a discussion: (+4 comments) Now we explicitly disable direct comms for all single-GPU jobs with threadMPI. That is a pretty common config, so I think we should test the default code path, unaltered by any env. variables, as well.
KUBERNETES_EXTENDED_RESOURCE_LIMIT: 1 tags: - k8s-scilifelab needs: - job: gromacs:clang-12-cuda-11.4.1:build gromacs:clang-12-cuda-11.4.1:regressiontest: extends: - .gromacs:base:regressiontest - .rules:post-merge-acceptance image: ${CI_REGISTRY}/gromacs/gromacs/ci-ubuntu-20.04-llvm-12-cuda-11.4.1-hipsycl-7a309b808-rocm-4.3 variables: CMAKE: /usr/local/cmake-3.17.2/bin/cmake KUBERNETES_EXTENDED_RESOURCE_NAME: "nvidia.com/gpu" KUBERNETES_EXTENDED_RESOURCE_LIMIT: 1 REGRESSIONTEST_PME_RANK_NUMBER: 0 REGRESSIONTEST_TOTAL_RANK_NUMBER: 2 REGRESSIONTEST_OMP_RANK_NUMBER: 1
Edited by Szilárd Páll