Newsgroups: comp.parallel.pvm
From: vanvoors@boreas.src.honeywell.com (Brian VanVoorst)
Subject: Q's about Paragon and PVM
Organization: Honeywell Technology Center
Date: 8 Sep 1995 08:55:00 CDT
Message-ID: <42phvk$grl@moon.src.honeywell.com>

I am interested in understanding more about other people's experience using
PVM on the Paragon and how PVM is implemented on the Paragon.
 
Let me say that I already know that PVM is going to perform poorly on the
Paragon, but I am required to get some timings of an application implemented
with PVM on the Paragon so that is just the way things are.
 
It's a simple application and when it runs using NX (the Paragon's native
communication lib) on four nodes it take about 9 seconds, the send
side communication cost is about 2% of that.
 
When the code is ported to PVM (the communication is simple, it's a direct
port), it takes about 81 seconds.  Now, I expected the communication to be
much much slower, but it isn't THAT slow.
 
I have a bunch of questions floating around my head, if anyone has any
answers they would like to offer please respond.
 
   What is happening that is making it so slow?
      ... Is there a PVM daemon on each node that is swapped in from disk
            for every communication?
      ... How does PVM on the Paragon work?
 
   What are other peoples experience with PVM on the Paragon?
      ... Do other people get these performance problems that are above and
           beyond what you would expect for communication overhead?
      ... Is there a real code that got good performance that I can look at?
 
   Are there any papers out there on PVM on the Paragon that I should find?
 
Thanks for your help.
Brian
 
vanvoors@src.honeywell.com

