Newsgroups: comp.parallel.pvm
From: eckhard@ts.go.dlr.de (Eckhard Rueggeberg)
Subject: Re: Timing Distributed Processes
Organization: DLR, Abt SM-SK
Date: Wed, 17 Aug 1994 06:54:38 GMT
Message-ID: <Cuo373.Fn2@dnsserv.go.dlr.de>

In article Evq@IRO.UMontreal.CA, le_saux@IRO.UMontreal.CA (Eric Le Saux) writes:
>   We want to time master-slave PVM programs.  All the slaves run on
>different but identically configured machines (Sparc 5).
>
>   Lets suppose we are alone on the network then a real-time clock
>gives quite stable measurements.  It can be used to compare runs
>on different problem sizes.
>
>   The problem is that most of the time, those machines are used for
>other purposes (as you would have guessed).  
> [...]
>    So if you have any comments on all this, particularly if you have a
> better function based on cpu time, I would greatly appreciate it.

I don't think that either formula measures what you want to know
(well at least what I want to know). Imagine an idle Workstation
farm : If you measure CPU time (given by getrusage, for example),
you won't get the effects of CPU's waiting for communication. We 
had such an example here : A programmer measured only CPU time, 
and was happy that the sum of two CPU times was only epsilon larger
than sequential execution. I had to show him that on two otherwise
idling workstations on a quiet ethernet, the total execution time
was LONGER than sequential execution on a single machine, because 
the efficiency was extremely poor. The processes got only 30% CPU
time on their machines, which were idling 70%. This obviously is
not what you want, but you wouldn notice if you rely on any formula
on slave CPU times.


---
Eckhard R"uggeberg
eckhard@ts.go.dlr.de



