From: "Alain Coetmeur" Newsgroups: comp.parallel.mpi Subject: Re: Lack of Performance under WINNT 4.0 Date: Tue, 10 Nov 1998 19:12:07 +0100 Organization: Informatique-CDC Message-Id: <729vr3$r8f2@puligny.idt.cdc.fr> References: <364879BE.5B2C8837@lhm.mw.tu-muenchen.de> Mime-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 8bit sorry I've forgotten some point Martin F. Schuster a écrit dans le message <364879BE.5B2C8837@lhm.mw.tu-muenchen.de>... >- we are using the TCP/IP protocoll under WINNT - is it possible that > data colissions occur? - is there a better way to go? MPI should avoid using UDP/TCP/IP NT layers for performance... maybe using lower layers (LLC?) or directly the network card... but then you would depend on the card you use... anyway that is not your business, but WMPI maker's. > >- dependent on the data rate, mpi errors occur: obviously there's a > limit for the maximum rate of data transmission - might this be >related > to the PC's network card ? Is there a buffer size to be extended ? look alos if ther are buffered/unbeffered version of MPI function I thin buffered if slower but safe... otherwise try to interlace communications so that ther are no deadlock, nor overflow... for example in a 2D neigbour border exchange the naive algorith is to send data to your left and get data from your right... but when sending if buffer is full, you must wait from the reader to read... but he cannot since he is sending to it's other neigbour... so deadlock or overflow ! the best technic is to send to right if your rank is even, and receive from left if your rank is odd then send to left if your rank is odd, and receive from right if your rank is even... smart is'nt it ?