[mpich-discuss] disable-auto-cleanup send/receive example
Stewart, Robert
R.Stewart at hw.ac.uk
Wed Nov 2 20:14:39 CDT 2011
On 11/02/2011 07:24 PM, Pavan Balaji wrote:
Hi Darius, Pavan,
> In your program, if I added a
> sleep(1);
> MPI_Iprobe(0, 0, MPI_COMM_WORLD, &flag, MPI_STATUS_IGNORE);
> in the error case of the receive, it did run as you expected.
> There's still a race, it's just less likely to hit in your example.
You're right, that did seem to work. May I ask what the need for the `sleep(1)' is? I'm trying now with another architecture - passing a message round a ring, that is slightly more unpredictable than my previous master/slace example.
Your suggestion does work for that too, but a lot more slowly once a remote process was killed. Is there a possibility to avoid the sleep(1) ? And can you explain for me why the MPI_Iprobe has any effect on fault tolerant behaviour. It does, as you have shown. But I don't know why.
Pavan,
> You cannot just "refresh" MPI_COMM_WORLD by resizing it, as the process
> ranks will be completely messed up if you do that. What you really want
> to do is create a new communicator with the remaining "alive" processes.
That's interesting, this is something I'd like to try. I can't find any compilable examples that make use of MPIX_Group_comm_create, would you be able to provide on? I tried, but wasn't sure how to create/manipulate the MPI_Group, for the 2nd parameter. Are there any code examples?
thanks,
--
Rob Stewart
--
Heriot-Watt University is a Scottish charity
registered under charity number SC000278.
Heriot-Watt University is the Sunday Times
Scottish University of the Year 2011-2012
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20111103/da506622/attachment.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: hw_uni_of_year.jpg
Type: image/jpeg
Size: 4803 bytes
Desc: not available
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20111103/da506622/attachment.jpg>
More information about the mpich-discuss
mailing list