From: "Alain Coetmeur" <alain.coetmeur@icdc.caissedesdepots.fr>
Newsgroups: comp.parallel.mpi
Subject: Re: LAM question: what does mpirun -c2c ?
Date: Mon, 17 May 1999 16:33:12 +0200
Organization: Informatique-CDC
Message-Id: <7hp9g5$d9j3@puligny.idt.cdc.fr>
References: <7hp20m$gat$1@sunu789.rz.ruhr-uni-bochum.de>
Mime-Version: 1.0
Content-Type: text/plain; charset="iso-8859-1"
Content-Transfer-Encoding: 8bit
Xref: ukc comp.parallel.mpi:5101


Tony Stoecker a écrit dans le message <7hp20m$gat$1@sunu789.rz.ruhr-uni-bochum.de>...
>
>Hi ,

>>> mpirun -O -c2c appschema
>
>to reduce the transfer-times. (but I really did not understand
>why this option should lead to much improved performance, as
>it is stated in the mpirun-manpage.)

-c2c, as I suppose, make the programs communicate directly
between themselves (client to client I think) instead of using
LAM MPI daemons as thirdparty.
anyway the daemons are usefull for tracing purpose and some other
things I ignore yet (who can complete)

>The following happened:
>The transfer-time are little smaller for x <= 2^15 byte , but
>when each slave send back >= 2^16 byte the transfer-time
>increased strongly.  (factor 13 larger than without the -c2c option)
strange, on our 8 bi-P2 clusters on fast ethernet, -c2c make all much better.

One problem we had was about memory, when one one was
short on RAM the performance were trashing above a few megs.

note that 32 or 64k is about a break in the slope of the delay curve,
probably because it change protocole (UDP-> TCP...)

I don't know the internals of LAM but any comment would be appreciated.

>Can anybody please help me what is going on?
>Maybe I have to adjust mpi-packet-sizes or something like that?