Newsgroups: comp.parallel.pvm
From: shaikh@brando.ece.utexas.edu (Saghir Ahmed Shaikh)
Subject: Performace issue with PVM based applications
Organization: The University of Texas at Austin, Austin, Texas
Date: 23 Jun 1996 07:43:38 GMT
Message-ID: <4qisja$moi@geraldo.cc.utexas.edu>


Dear PVM User

I need some suggestions regarding my program using the PVM 3.3.10 and its 
performance measurement. 

I have developed a simulation program which spawns around 150 process. The
processes communicate very heavily during the execution. I wanted to do
the performance comparison on the workstation cluster of 5 and 10 to see
the scalability. 

Well, the results are against the expectations. The time on the 5
workstations is better (or almost the same)  than the 10 workstations! In
my opinion this is happening because in the case of 5 workstations each
workstation gets around 30 processes thus all the communication between
the "blocks of 30 process assigned to that workstation" communicate faster
(I think PVM daemon recognizes the fact that the destined processes are on
the same host or not and communicates faster if they are on the same
processor). Well in the case of 10 workstations, each workstation get 15
processes, thus now the block assigned to each host is reduced and thus
the "inter host communication" increases (inter-process is the same).


Well could you suggest a strategy where I can show the both sides of the
story. i.e. the actual happening (which I described earlier) and
theoretical case when the 10 hosts show better performance than 5 hosts.


Please let me know if you have any suggestions. 

My email is saghir@mail.utexas.edu

Thanks
 
-- 
--
peace, saghir
-------------------------------------------------------------------+
My Home Page: http://www.ece.utexas.edu/~shaikh/		   |
-------------------------------------------------------------------+
Saghir A. Shaikh	       | Email: saghir@mail.utexas.edu	   |
Dept. of Elec. & Comp. Engr.   | The University of Texas at Austin | 
-------------------------------------------------------------------+

