Newsgroups: comp.parallel.pvm
From: jtdougl@hubcap.clemson.edu (John T. Douglass)
Subject: UDP and TCP
Organization: Clemson University
Date: 22 Mar 1995 14:04:05 GMT


I have a simple distributed time driven simulation implemented using pvm.  
The traffic pattern is simple and the processors are laid out in a    
mesh connected grid with no wrap around.  Messages passing between 
processor are basically synchronization messages, with some carrying more
information than others.  The simulation works fine using default routing,
however if I switch to PvmRouteDirect the simulation will now hang 
about 20% of the time inside of a call to pvm_send().

I was wondering if anyone out there has any ideas to why this would be
occurring.  I have used direct routing in the past and had nice performance
gains. Direct routing seems like a logical choice here due to the limited
number of connections needed (at most 4 for any 1 processor).  And over
the life of a simulation run the savings should be great.

Any pointers on where to look further or suggestions would be greatly 
appreciated.


-- jd

John T. Douglass  (douglass@cs.clemson.edu)
Clemson University


 

