Newsgroups: comp.parallel.pvm
From: phv@ipp-garching.mpg.de (Philippe    Verplancke     TE  )
Subject: Very slow execution on SUNos4-Workstations with AFS
Organization: Rechenzentrum der Max-Planck-Gesellschaft in Garching
Date: 20 Dec 1994 21:43:44 +0100
Message-ID: <3d7fm0INN416@spcfnh.aug.ipp-garching.mpg.de>


Hello, 

I parallellized a fortran code which computes the trajectories of
about 80000 charged particles in a self-consistent electric
field.  When I run with 2 SUNos4-Workstations (Sparc classic),
each one "does" 40000 particles. For each time step of the calculation,
only a few quantities, like the cell densities, are "sum-reduced" and
exchanged between the processors (4 vectors of 1600 REAL8). I wrote this 
up to give you an idea of the degree of "parallellizability" of this code.

My problem is, that this code runs about 100% slower on two processors,
as the unparallellized version does on one processor only! 
What I observe during run-time, is a continuous use of the disk,
although paging and swapping are negligible.  The two machines are working
with the Andrew File System, and I use PVM version 3.3.0.
Since the code normally doesn't use the disk during the run, I wonder why
this happens during parallellized execution with PVM. Could this be the
consequence of my bad programming?  Or does PVM perhaps use the disk, 
because of some feature of the AFS system? In any case, I guess that it
explains the slow execution.

Thanks for any hints!

Philippe Verplancke

