Or Gerlitz wrote:
Steve Wise wrote:
Or maybe map PS_TCP port space to the be the unified space, and have
a new PS_RDMA or PS_IB port space that doesn't? At and second
Please note that the rdma-cm derives the IB Service-ID from the port
space and the port provided to rdma_resolve_addr - and this is done
along the lines of the IBTA "RDMA IP CM Service" annex A11.
So we would need to have PS_IB_UD and PS_IB_RC that SID(ps=TCP,
port=X) = SID(ps=IB_RC, port=X)
at first thought this is possible and would add too much compatibility
issues - except for the need to provide an app with the means to sense
if the underlying librdmacm support the _IB port spaces.
Still, I would like to avoid the case where someone runs MPI job over
RDMA-CM over IB and librdmacm opens/bounds to all these sockets - so
the preferred solution - for my taste would be some setting done by
the user to have librdmacm allocate socket / carve a port - and it
wouldn't be the default. For example, using an environment variable,
e.g similar to the libibverbs fork support.
I could live with that. It would allow taking advantage of this
functionality without changing all the apps.
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
