On Jan 29, 2007, at 1:44 AM, Rajesh vanaparthi wrote:
> I am student doing project on MPI(implementing parallel matrix
> multiplication algorithms on several computers).i installed the MPI
> software in the server of my college. But there exists some
> problem in the server(during lamboot).
> Actually I have do master slave programming i.e one computer as
> master and remaining computer as slaves and execute my programs.As
> there is problem in the server I started implementing the same
> algorithm in a single machine with rank 0 as master and remaining
> processes(rank 1,2….) as slaves and calculated the computation time.
> My doubt is that am I doing the correct way (executing on
> single machine with several processes as I mentioned above) and
> another reason is that I am not having sufficient time to do
> another way.
> Is there any any difference in calculating the computation
> time in both the ways
>
> Briefly to say
> Is there any difference in executing the parallel algorithm (using
> MPI) on single machine with several processes(rank 0 as master and
> remaining as slaves) instead of executing on a parallel computer
> with several computers(nodes)(i.e node 0 as master and node 1,2
> ….as slaves).
> I am following the first approach (single machine) and doing the
> code for fox and cannons algorithm. Is that the correct method.
Generally, such things don't matter for the correctness of MPI
applications -- MPI hides all the communication details for you.
However, the change can have some huge effects on performance, either
good or bad. If the app needs more resources (processors or memory)
than one machine can make available, your app can see great speed
hits when running on a single machine. On the other hand, the
latency and bandwidth of shared memory are so much greater than
between machines (especially if you are using TCP between nodes) that
the increased communication cost can slow things down when running on
multiple machines.
Hope this helps,
Brian
--
Brian Barrett
LAM/MPI developer and all around nice guy
Have a LAM/MPI day: http://www.lam-mpi.org/
|
