Newsgroups: comp.parallel.pvm
From: "Bo-sung, Lee" <bslee@pvmcube4.snu.ac.kr>
Subject: Re: Run time analysis on PVM?
Organization: Aerodynamics Lab. Dept. of Aerospace Eng. Seoul National Univ.
Date: Sun, 15 Dec 1996 13:32:37 +0900
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <32B37F65.2CC@pvmcube4.snu.ac.kr>

nis2@Lehigh.EDU wrote:
> 
> Hi All..
> 
> I have started to use PVM to implement my parallel numerical codes. What I
> need to do is to make predictions about the run time of the algorithms and
> compare them with the experimental values. However, they differ quite a lot
> than what I predicted. I couldn't figure out why? What latency value should I
> assume for pvm programs. What is the cost of putting a message into the
> network for example? I am working on IBM workstations here..
> 
> Thanks in advance for any suggestions?
> 
> NihalIn my experience, use 'PvmRouteDirect' and 'PvmDataRaw' if you connected
same workstations. And, message send/receive overhead is induced not from
message size you send/receive but from how frequently you packing data, this
is alos my experience.

Using my pvm code-Navier Stokes solver for 2D problem, the speed up is about
3.6 on my 4 Linux Pentiums. And, 16 node SP2 (using high-performance switch)
the speed-up is only about 11. In 16 node SP2, comm./computaion is very large
so, I can't get some good speed-up. But, In 8 node SP2, speedup is about 6.7 

If you want to evaluate your parallel code, compare speed up and scale up.
The scale up is that you increase the problem size, e.g. matrix size as incresing
node number. That is, if you use 4 noded and problem size was 100x100, and if
you want to use 16 node and evaluate parallelism, use 200x200( 4 times larger
than 100x100) and the time is about 1.0 

Bo-sung, Lee Ph.D. Candidate
Aerodynamics Lab. Dept of Aerospace Eng.
Seoul National Univ. Korea
e-mail : bslee@pvmcube4.snu.ac.kr

