On Tue, 2009-03-24 at 07:03 -0800, Eugene Loh wrote:
I'm not sure I understand this suggestion, so I'll say it the way I
understand it. Would it be possible for each process to send an "all
done" message to each of its neighbors? Conversely, each process would
poll its neighbors for messages, either processing graph operations or
collecting "all done" messages depending on whether the message
indicates a graph operation or signals "all done".
Ashley Pittman wrote:
Exactly, that way you have a defined number of messages which can be
calculated locally for each process and hence there is no need to use
Probe and you can get rid of the MPI_Barrier call.
Hi Eugene,
By `poll its neighbours', do you mean posting an MPI_Irecv for each
neighbour, and working until an `all done' message (sent using
MPI_Send) has been received from each neighbour?
As long as each process posts its MPI_Irecv before starting the
MPI_Send, are we guaranteed that two processes won't deadlock by
MPI_Send-ing to each other? I avoided this method at first because I
didn't understand that the MPI_Irecv would stick around regardless of
a message being ready to receive; I figured that it had no effect if
no message was ready to receive.
In this implementation, the graph is partitioned arbitrarily; the
vertices are distributed based on a hash function of each vertex's
unique ID, not based on the topology of the graph (which would be
nice, but difficult). So, every process is a neighbour of every other
process.
Thanks for your help. Cheers,
Shaun
