Running MPICH on constrained CI environments under oversubscription or close to is usually slow by default.
Things can be improved by tweaking some environment variables, eg:
https://github.com/pmodels/mpich/blob/main/.github/workflows/ci-workflow.yml#L300
However, that only applies to CH4, and these variables may eventually change over time or be added/removed.
Would it be possible to add a special mode, turned on with let say a MPICH_CVAR_OVERSUBSCRIBE boolean var that is basically a request from users for MPICH to use different defaults for any other CVAR with the objective starving CPUs as little as possible? In short, something that would go as easy as possible with the system, like the old and gentle ch3:sock?
IIUC, Open MPI has an oversubscription mode were some internal settings/default change to better accommodate for oversubscription.
PS: Maybe something like that already exists and I'm not aware of it. In that case please let me know.
Running MPICH on constrained CI environments under oversubscription or close to is usually slow by default.
Things can be improved by tweaking some environment variables, eg:
https://github.com/pmodels/mpich/blob/main/.github/workflows/ci-workflow.yml#L300
However, that only applies to CH4, and these variables may eventually change over time or be added/removed.
Would it be possible to add a special mode, turned on with let say a
MPICH_CVAR_OVERSUBSCRIBEboolean var that is basically a request from users for MPICH to use different defaults for any other CVAR with the objective starving CPUs as little as possible? In short, something that would go as easy as possible with the system, like the old and gentlech3:sock?IIUC, Open MPI has an oversubscription mode were some internal settings/default change to better accommodate for oversubscription.
PS: Maybe something like that already exists and I'm not aware of it. In that case please let me know.