Newsgroups: comp.parallel.mpi
From: kielmann@informatik.uni-siegen.de (Thilo Kielmann)
Subject: Re: timing with MPICH on WS
Organization: University of Siegen
Date: 9 Jun 1996 12:45:50 GMT
Message-ID: <4peh1u$sob@si-nic.hrz.uni-siegen.de>

In article <4p9a6j$dne@neptune.ethz.ch>, cesari@inf.ethz.ch (Giovanni Cesari) writes:
|> I'm using MPICH on a network of SPARC workstations and I have the following 
|> problem with timing:
|> my algorithms  are implemented using a master slave approach.
|>  
|> 
|> What I would like is somethink like:
|> 
|> if (myid == 0) {
|>   t0 = time();
|>   send task to slaves;
|>   receive results;
|>   t1 = time();
|>   t2 = t1 - t0;
|> }
|> 
|> If now I use the wall clock MPI_Wtime I measure also 
|> the load on the workstations due to other users.
|> On the other side, if I use clock() I don't measure the
|> time needed by the slaves.
|> The manual doesn't help me.
|> 

Well, why not use clock() on all nodes and finally collect the values?


But I doubt that this will provide useful information.
Can you be sure that processor time of your application processes caused
by other user processes (e.g. in form of memory paging...) will not be
included in your timings?

On workstation clusters, you have (IMHO) two general choices:
1. Measure wall clock time on otherwise unloaded machines (and networks!)
2. Define a given measure of workload caused by other users and measure
   wall clock time in the presence of this workload.


Thilo
-- 
Thilo Kielmann
Univ. of Siegen, Dept. of Electrical Engineering and Computer Science (FB 12)
Hoelderlinstr. 3, D-57068 Siegen, Germany
Phone:  +49 271 740 2642                                               /\,/\
Fax:    +49 271 740 2532                                              / u u \
E-Mail: kielmann@informatik.uni-siegen.de                             \<_Y_>/
WWW:    http://www.informatik.uni-siegen.de/~kielmann/                /  "  \

