On Feb 28, 2008, at 11:29 PM, Supalov, Alexander wrote:
> - Communicator & group management: better memory footprint.
Take this point to an extreme - it may be possible to say "this app
only uses MPI_COMM_WORLD". In this case, you can remove the
communicator field from network packets for a small gain in latency,
or perhaps reduce it from 4 to 2 bytes or 1 byte (e.g., if an app says
"I'll only use 4 communicators").
> - Message tagging: better support for stable dataflow exchanges,
> smaller
> packets.
Two points here:
- allow app to eliminate MPI_ANY_TAG
- just like with communicators, allow the app to say "I'll only use N
tags", where N can save you space in network packets (e.g., if N==1,
no need for tag on the wire; if N == 2, then you only need 1 byte for
the tag, etc.).
> - Non-blocking communication: easier ordering, simplified request
> handling.
If there is no non-blocking communication, enormous chunks of the
progression engine can be optimized in terms of memory (i.e., remove
lots of now-unnecessary code) and probably a little speed.
On the teleconf (sorry I missed it), was there discussion of how to
specify these hints? Perhaps a new function: MPI_INIT_INFO (pass an
MPI_Info handle to MPI_INIT)? Or is it something that needs to be
specified at compile/link time?
--
Jeff Squyres
Cisco Systems
|
![[Flag]](../../../../../images/flag-us.gif){kind=small}