Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: [OMPI devel] sm BTL flow management
From: Eugene Loh (Eugene.Loh_at_[hidden])
Date: 2009-06-23 21:40:11

I have a fix for ticket 1944 working, but the broader problem is
unpleasant. E.g., let's say we have zillions of uncountered Bcasts or
something. Say, the root is repeatedly emitting sends, but never
polling its in-coming FIFO. Return fragments will be accumulating, the
FIFO will be congested, pending-send queues on peer processes will be
growing, etc. The code now handles this (by growing the pending-send
queues and eventually draining them, pre-1.3.2 we would also have
handled this by growing the FIFO and using up the shared memory), but
the whole thing is disturbing. E.g., queues might drain only when the
root reaches MPI_Finalize. (Okay, unclear to me what sort of real
application would have communications only from one process going out.)

So, is this (one-way communications, e.g., repeated Bcasts) pathological
and not worth worrying about. Or, are other solutions worth
considering? E.g., I'd like to have a sending process run
mca_btl_sm_component_progress occasionally, even if it is successfully
completing its sends.