From: rrzanny@cs.wmich.edu (Rodger R. Zanny) Newsgroups: comp.parallel.mpi Subject: Re: Questions using MPICH on SunOS 5.6 ? Date: 27 Jul 1999 14:50:24 -0400 Organization: Western Michigan University Message-Id: References: <379dc078.2968312@news.nic.utwente.nl> Xref: ukc comp.parallel.mpi:5383 >>>>> "A" == Andrei A Dergatchev writes: A> Hi, I've encountered a seemingly strange (for me) problem working A> with speeding up a simulation package. I've compiled the MPICH A> package on Sun E3500 under SunOS. The machine has 5 CPUs, and I A> expected to get a considerable increase in speed as calculations A> are highly parallized. However, currently I see a strong A> *increase* (~50%) in computation time going from 2 to 4 processors A> ("mpirun -np 2" -> "mpirun -np 4"). A> Why it would be ? Currently I'm thinking about trying PVM package IMO, I would doubt that if you get such a poor speedup with MPI, that PVM would significantly improve things. I would look at your algorithm. I work in numerical parallel integration (working with an asynchronous algorithm in a message passing environment); I've seen that the work required to get an answer increases drastically as #PE increases. I get speedup in time/work, but since the work required is increasing, the speedup is generally poor. (The reason is basically that the work being performed at any moment is inherently non-optimal due to the parallelization. I've seen similar problems in very different parallel algorithms for integration.) Might your problem be related? Can you measure the amount of *work* (calculations, etc.) being performed at 1, 2 and 4 PE's? --Rodger PS. I hope this gets posted ok. I'm gnew to gnus. -- *-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*- - Rodger R. Zanny Email: rrzanny@cs.wmich.edu.nospam * * Doctoral Student (remove .nospam to send mail) - - Computer Science Department Web page: I'm working on it! * * Western Michigan University - -*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*