From: Mark Gebhardt Newsgroups: comp.parallel.mpi Subject: mpich send delays Date: Tue, 15 Jun 1999 20:49:43 +0200 Organization: RRSG Message-Id: <3766A047.97583711@rrsg.ee.uct.ac.za> Reply-To: mark@rrsg.ee.uct.ac.za Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Xref: ukc comp.parallel.mpi:5205 Hi I have run into what seems to be a strange problem using mpich1.1.2 on linux 2.0.36 My application is master:slve, with 2 levels of slaves, ie master:slave1:slave2 The data (4MBytes) is sent from the master to a lev1 slave and then to lev2 slaves. I have noticed that if the lev1 and lev2 slaves exist together on a single-processor node, the comms time between them shoots up as they fight for processor time. The increase is so great that it is doubling the execution time of my entire application. eg: time for send on slv1 (2 processor node): 0.21 secs time for send on slv1 (1 processor node): 68.95 secs !!! I am simply using MPI_Send and MPI_Recv. I also notice that this does not happen in my PVM implementation of the same code! Any clues as to what exactly mpich is doing? and, more importantly, how do I get around this? Thanks for your assistance Mark -- Mark Gebhardt Radar Remote Sensing Group University of Cape Town South Africa email: mark@rrsg.ee.uct.ac.za www: http://rrsg.ee.uct.ac.za -- Mark Gebhardt Radar Remote Sensing Group University of Cape Town South Africa tel: +27 21 6503756 email: mark@rrsg.ee.uct.ac.za www: http://rrsg.ee.uct.ac.za