[mvapich-discuss] message coalescing

[Dhabaleswar Panda]

> Hello!
>
> I've read your presentation: Reducing Connection Memory Requirements of
> MPI for InfiniBand Clusters: A Message Coalescing Approach.
> Could you please explain one moment.
> I see that in the examples provided in your presentation different
> benchmarks were started with 256 processes, 4 per node.
> So messages going from different processes say of node A to node B are
> coalesced, aren't they?

No, Coalescing is done per QP-basis. We are talking about MPI library
here. So, messages across a pair of processes (say process 1 on node A
with process 1 on node B) can be coalesed. There is no aggregation of
messages from different processes on node A to other node.

> Does it work when I have several applications communicating from one
> node with other nodes,
> could messages from different applications be coalesced?

No.

Thanks,

DK

> Best regards,
> Arthur
>
> _______________________________________________
> mvapich-discuss mailing list
> mvapich-discuss at cse.ohio-state.edu
> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>