On Apr 7, 2015, at 7:45 PM, Jeff Hammond <[email protected]> wrote:

...

...

...

I don't expect MPI to make async progress on Ethernet unless you have a comm thread in MPI poking the network.

Friendly reminder, Jeff: Ethernet != sockets. The statement you are making is generally true for sockets but not necessarily for Ethernet.

I bit my tongue for the same reasons… ;-)

Correction: I do not expect MPICH to make async progress on Ethernet (because it uses sockets).

Does MPICH make async progress on InfiniBand using Verbs without calling into MPI_Test*/MPI_Wait*? Does MPICH not fragment large RDMAs? Yes, the HCA (why can’t they call it a NIC?) will offload much of the work, but not all of it. Matching is still done in the host and only when invoked by the app calling into Test/Wait or a progress thread doing it on the app’s behalf.

I'm not sure how this in any way affects the conversation being had, other than to demonstrate that chemists are unreliable sources of information about networking hardware ;-)

It is not just chemists. I hear HPC computer scientists say the same thing… ;-) Scott _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss