From: Vasiliu Bogdan Newsgroups: comp.parallel.mpi Subject: Re: Speed using MPI Date: Thu, 24 Sep 1998 15:46:09 -0500 Organization: Iowa State University, Ames, Iowa, USA Message-Id: <360AAF91.62424C39@scl.ameslab.gov> References: <6u64r3$9a5$1@news1.tc.umn.edu> <6u6stn$5mq@news1.newsguy.com> <6u8n37$pek$1@news1.tc.umn.edu> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit AJ Lewis wrote: > Thanks to everyone who responded to my question. The answers I got back are > very helpful in evaluating the use of MPI for the project I am working on. > > I have another question for you now: Is there any way to specify the packet > size in MPI? I know with sockets you can specify a packet size up to a > certain value around 15-16kb. Is there any way to do this with MPI, or does > MPI optimize the packet size for you? > > Thanks, > AJ Lewis I do not know if you can specify the packet size in MPICH unless you modify the p4 library. But you can modify the TCP buffer size without recompiling the MPICH code each time. You have to put in your .cshrc file the command: setenv P4_SOCKBUFSIZE (where buffer_size is a number that will give the size of the TCP buffer; for example: setenv P4_SOCKBUFSIZE 16384) (make sure you type "source .cshrc" each time you change the value) You can test different values for the P4_SOCKBUFIZE and finally choose the one that gives you the best results. I tested sizes ranging from 8kbytes to 8Mbytes and I have found the best size to be 64kbytes (on my PC cluster). Another issue that you should take into account is the stability of the SMP kernel. This kernel does not seem to be stable at all. Basically we could not run application that require large amounts of memory or extensive computations (the computers crashed), so we switched back to an old kernel that enables only one processor. Bogdan -- Bogdan Vasiliu Research Assistant - Ames Lab - ISU 304 Wilhelm Hall bogdan@scl.ameslab.gov Iowa State University Phone:(515)-294-7336 Ames, IA 50011