Hello, I have a big problem that I thing needs about 6Gb of memeory. Th total number of elements is about 8.25 Million with the ratio b/w tet to hybrid 7.0/1.3 I am setting up the parallel computer, 2 linux and one windows xp. I have the 2 linux machines running in parallel, but I am not sure how to add the windows xp. questions 1) How much memory will it take to solve this problem ? 2) Does anyone has an idea how to get a windows machine to join the setup ? Please help !

Hi Rao, As I think you need to edit hosts.ccl on your master machine (Is your master machine linux?) to include your XP machine with appropriate architecture string. This could be done by using cfx5parhosts or manually. Our current experience with windows parallel computing is bad - very slow (we build cluster consisted of computers running under windows 2000 and perfomance degrade with each additional machine so three computers performed worse than 1). On linux everything is OK. For problem size you could run job in serial and sum memory requirements in begining of Solver section. (Of course you job will be completed with error - because of memmory)

Hi Rao/deLuther,

The parallel performance in the CFX5.7 release (coming soon, hopefully only a few weeks) is much improved over 5.6. In my benchmarking of the beta release it is as fast in Windows using the new MPICH parallel option as it is in Linux.

Regards, Glenn

You only need to include the host name and installation directory (if not installed in the default location). The solver will determine the appropriate architecture.

-RK

Interesting Glenn,


So do you think that pvm will be replaced in the future by this MPICH technology?


Cheers, G. G.

Hi Gloria,

Hoepfully it will not replace it will be a continuing option. For pure windows clusters MPICH seems to be by far the best choice, but for mixed windows/Linux/Unix clusters the PVM option may end up being faster. Don't know, I haven't tried it.

Regards, Glenn

MPI works well for local parallel, but is problematic for distributed parallel. You should continue to use PVM for distributed parallel.

Regards, Robin

No, MPICH will never replace PVM. MPICH is not robust to processes dying or a machine dying.

So, slow and robust (PVM) or faster and not so robust (MPICH). That's the options. Bottom line is that there's no free lunch with this stuff.

Neale.