deadlock possible in mpi-conduit w/ blocking sends

• Previous message: Wes Bethel: "possible upc_memget() and upc_memput() bugs"

Hi all,
When using the GASNet mpi-conduit built upon the AMMPI library, I think 
that there is a deadlock possible when using blocking sends 
(AMMPI_NONBLOCKING_SENDS=0).

Consider the following GASNet client  pseudocode code (based on 
testcore1.c):

SENDER
iter=0;
while (iter++<n)
     gasnet_AMRequestShort(RECEIVER);

reply_handler()  {}
------------------------------------------
RECEIVER
while (iter<n)
     gasnet_AMPoll();

request_handler(sender)
     iter++;
     gasnet_AMReplyShort(sender);

If blocking MPI sends are used, then we can have the following scenario:

SENDER                           RECEIVER
t=0: send M1                      t=0: AMPoll() - no message
t=1:     AMPoll() - no message    t=1: AMPoll() - no message
t=2:     BSend(M1)                t=2: AMPoll() - no message
t=3: send M2                      t=3: AMPoll() - message M1
t=4:     AMPoll() - no message    t=4:    process M1
t=5:     BSend(M2)                t=5:    BSend(R1)

In this case, we get a deadlock since both processes are waiting in a 
blocking send.

Is this a problem with AMMPI/mpi-conduit or am I missing something? If 
it is indeed a problem, is the only solution using blocking sends?

Thank you,
George

• Previous message: Wes Bethel: "possible upc_memget() and upc_memput() bugs"