message size not multiple of 4 when sending extremely large messages

[mattijsjanssens]

We're occasionally seeing assert message of the form

ips_proto.c:1646: (scb->payload_size & 0x3) == 0

which seem to originate from somewhere in the network stack (e.g. https://github.com/intel/psm/blob/master/ptl_ips/ips_proto.c) when the size is not a multiple of 4.

• this only happens occasionally
• and only for extremely large messages (not sure but could be 100Mb or even Gb)
• and only on omnipath

Is this a known problem? We don't pad our mpi messages to be multiple of 4 bytes.

[hestela]

Hi, the psm2 library for omnipath (hfi cards) is in this repo: https://github.com/01org/opa-psm2
this repo is for truescale psm (without the 2), for the QLEXXXX cards