On Sat, 2009-08-08 at 21:42 -0700, Martin Schulz wrote:

Hi all,

Our next MPI3 tools WG meeting will be on Monday 8/10 at 8am PDT.

Apologies as I won't be a able to make the call this week due to other commitments.

So far the only agenda point that I have

Jeff and I spoke over my collectives proposal discussed before, the general consensus at the time was that I should work with Jeff to integrate this into being a part of his general handles debug interface. Having discussed this we now realise that's a sensible or realistic approach. As such my proposal stands as it originally did, I'd like to propose a small collective extension built on the existing message queue interface. To some extent this is therefore blocked on that work being written up and adopted before I can proceed. I'm easy on the specifics of the function parameters but the basics are the library maintains a pair of counters for each possible collective (Bcast/Barrier...) for each communicator. An extra callback function is added which returns this information for the "currently active communicator" in much the same way as message queue information is returned. The query operation is entirely local and is only aggregated into something more useful inside the tool which has access to the results from all processes. The two counters are initiated at zero and incremented when A) A call to the collective is made and B) that call returns. The interface returns both the first counter and either the second counter or potentially a single bit of information specifying whether the counters are equal or not (i.e. if the collective is "active" on that rank). As before the proposal is on-line at http://padb.pittman.org.uk/extensions/

To help this discussion, Jeff has offered to host this meeting via Webex. ... that this will make it free even for non US callers :).

That's appreciated from this end. Ashley Pittman. -- Ashley Pittman, Bath, UK. Padb - A parallel job inspection tool for cluster computing http://padb.pittman.org.uk