Newsgroups: comp.parallel.mpi From: Reto Koradi Subject: Re: MPI doesn't perform well on the Cray J90 Organization: Swiss Federal Institute of Technology (ETHZ) Date: Fri, 26 Sep 1997 13:25:26 +0200 Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Message-ID: <342B9BA6.5115@mol.biol.ethz.ch> Roberto Lionello wrote: > I'm doing some tests on a J90 using a program developed on the > T3E. With a given size of the problem, it scales nicely on the T3E > using up to 128 processors. I compiled it on the J90 using > TCP sockets and I was disappointed by the performance: > it saturates with 8 processors! :-( The performance of MPI over TCP is indeed extremely slow. But as long as you only calculate on one machine, you can use the shared memory implementation, this is orders of magnitude faster. All you need to do is compile your code with the "-a taskcommon" switch (that's for f90, check the manual for other compilers), and use "mpirun -nt n" instead of "mpirunt -np n" to start the program. -- Reto Koradi (kor@mol.biol.ethz.ch, http://www.mol.biol.ethz.ch/~kor)