From: Robin N . T. <rn...@br...> - 2002-08-30 16:14:38
|
In fact I used one of the programs in the LAM-MPI test suite on the wanglab cluster and it worked like a charm. Can't tell you about performance increases, but I can tell you that it seems that setting up MPICH is much simpler, as compared to LAM-MPI; used LAM-MPI anyway because that was requested. As Moshe already said, they work together very well. Shabbat shalom to everybody and a good weekend to all others, -Robin * on the Fri, Aug 02, 2002 at 01:53:21PM -0400, Abraham Schneider was saying: > Hello, > > My lab is currently experimenting with software for running on our > beowulf cluster. We first tried PBS, after being told it was the defacto > standard, but had to switch because a queue system does not really fit > our needs. openMosix seems like the perfect approach for us to take, it > works wonderfully (so far, anyways), and it makes life much much easier. > > Most of our computational work is done using a simple method of having > the same program run many times over on different sections of a data set > (usually octave or matlab [which we are quickly switching away from > since it does not seem to play nicely with openMosix]). > > Anyways, my question is this. Currently I'm working on a evolution > computer library to work on the cluster for fitting models to > experimental data. I would _very_ much like to use MPI, since in the > past I had to write my own TCP/IP code for message passing. After > talking with someone from a different lab I was told that MPI and > openMOSIX may not play nicely together (i.e. its okay, as long as you > lock spawned jobs to a specific node). I have looked everywhere for more > info, including this mailing list, but I haven't found enough > information on the subject. What are the complications to running MPI on > openMosix, do I have to write my programs differently to take anything > into account, where can I find more information on this subject. One > item I did find in the mailing list was someone claiming that I should > not use MPI if I use openMosix (if so, is there another good message > passing library out there to use?), but Moshe responded with: > > moshe> Agree with everything except that openMosix (can we please call > it that > moshe> way? OM is ugly ) does not come into the way of MPI *at all*. In > fact, > moshe> openMosix can only be beneficiary for MPI. MPI just executes > through rsh > moshe> jobs on a remote node. With openMosix if that remote node is too > busy, moshe> that job will migrate to a better node. > > I understand that this question might be better suited for the MPICH or > LAM mailing lists (btw, if MPI does work, does it matter how well it > plays with openMosix depend on the implementation?), but I found even > less information on their web pages. > > > Thanks, > > Abe |