Newsgroups: comp.parallel.pvm
From: bubeck@peanuts.informatik.uni-tuebingen.de (Till Bubeck)
Subject: Re: Timing PVM programs
Organization: Technische Informatik, Uni Tuebingen
Date: 14 Mar 1995 17:18:37 GMT

Niall Adams (cmsnadam@kryten.cms.livjm.ac.uk) wrote:
: Hi,
: 	I am interested in timing my PVM application (yes, I know this
: can be a mine field!). The application consists of a master program
: driving a number of slaves tasks. After some two way communication 
: between slaves and master, the slaves terminate, leaving the master to
: do a final computation and some output. I guess this is the classic 
: Master-slave mode. Anyway, if I use the UNIX time command, I get 
: something like;

:   % time pvm_prog.out
: 460.509u 1.792s 18:03.06 42.6% 130+344k 4+66io 4pf+0w

: So the wall clock time for execution is about 18 minutes. The sum of 
: cpu and system time comes to less than eight minutes. Is this eight
: minutes a valid estimate of the throughput time for this job? Is there
: a more meaningful method of measurement?

: Thanks,

: Niall Adams
: cmsnadam@kryten.cms.livjm.ac.uk

In my opinion, you can *only* measure in wall clock, because you do
not know, how the system(+user) time of the involved processes
(master+slave) have to be arranged.

Taking the user time of the master is wrong in the following case: if
the master spawns a slave, which does all the work, the master's user and
system time would be nearly zero, which is definetely wrong.

Summing up the times of all slaves and masters is also wrong, because
you do not know how they are interfered. If they run in parallel, you
have to take the largest of all measured times. But if they could only
run sequentially, you have to sum them up.

In contrast measuring in wall clock takes all this into account, but
has to be done, when the machines are as unloaded as possible. 

Read the following paper, for a good way to measure message passing
systems, without this problems:

\bibitem[Kar92]{Kar92}{Nicholas T.\ Karonis: {\em Timing
Parallel Programs That Use Message Passing.} In Journal of
Parallel and Distributed Computing 14(1), pp.\ 29-36,
January 1992.}


Till Bubeck
--
                ,,,
               (o o)
-----------oOO--(_)--OOo-------------------------------------------
|  Tilmann Bubeck            |   _||__                            |
|  Universitaet Tuebingen    |  /   . | <- Berlin                 |
|  Technische Informatik     |  |    _/                           |
|  Sand 13, 72076 Tuebingen  |  |_   \                            |
|  Germany                   |   /*__/ <- University of T"ubingen |
|------------------------------------------------------------------
|  Tel: +49 7071 295865     Fax: +49 7071 610399                  |
|  Email: bubeck@peanuts.informatik.uni-tuebingen.de              |
|  WWW/Mosaic: http://www-ti.informatik.uni-tuebingen.de/~bubeck  |
-------------------------------------------------------------------

