Closed davidozog closed 11 months ago
Thanks @davidozog. SHMEMX_TEAM_NODE
is better.
LGTM! Should things like SHMEM_TEAM_HOST_INDEX and shmem_internal_team_host be SHMEMX_TEAM_HOST_INDEX/shmemx_internal_team_host though? From a readability perspective I definitely prefer things as they are, just not sure about the technicalities of using shmem vs shmemx. Are we good to use shmem internally for variable name prefixes to shmemx related objects so long as what is exposed to the user abides by the shmem/shmemx rules?
Thanks @philipmarshall21 - only the user-facing API needs the "X" to indicate it's an extension. The runtime can handle internal symbols however, and as you say, it kinda makes sense to withhold the "X/x" from symbols that aren't visible to users anyway.
Thanks @davidozog.
SHMEMX_TEAM_NODE
is better.
I think so too - it's renamed now.
This should be handy for supporting multi-NIC and/or topologically-aware collectives.
@wrrobin - perhaps this could be named
SHMEMX_TEAM_NODE
because it's relying on PMI's notion of a "node" (viashmem_runtime_get_node_rank()
). Here is a short summary of where this info comes from across the SOS "runtime-pmi
" options:gethostname
gethostname
PMIX_LOCAL_PEERS
queryMPI_COMM_TYPE_SHARED
But
SHMEMX_TEAM_HOST
is fine with me if that's preferred/simpler.