From: Yon Han CHONG <Y.H.Chong@cranfield.ac.uk>
Newsgroups: comp.parallel.mpi
Subject: Re: Dynamic load determination
Date: Wed, 18 Nov 1998 16:55:47 +0100
Organization: Cranfield University
Message-Id: <3652EE03.D56DE0B9@cranfield.ac.uk>
References: <3651E148.12E39523@usm.edu>
Mime-Version: 1.0
Content-Type: text/plain; charset=EUC-KR
Content-Transfer-Encoding: 7bit


Matt Bettencourt wrote:
> 
> Is there any command built into MPI that will let me figure out what the
> current load on all my processors are??  Example, I have a 64 proc
> machine.  I start up on all the procs.  I determin which are idle (or
> close to that) then start the job just on those procs.  After about 5
> minutes of run time, I check again and run on the now idle procs.  Is
> there any routine(s) that will help in doing this??
[...]

I was looking for something simular but the easiest way is using UNIX's
"ruptime" command. If your computer system doesn't have "ruptime" you
can "rsh <machine_name> uptime" to each machines to find out its current
load. By the way information from "uptime" is more recent than
"ruptime". Probably the best way is to write a C program to produce a
file containing name of machines you can run (or application schema,
have a look at your MPI implimentations manual) in every interval. 

To check whether one of the working machine is relatively too slow for
other machines use "MPI_WTIME" to measure the running time. You can also
check anybody is loged on with "ruptime". I found that the best way to
migrate a run is:

1. Save all the information you need to restart.

2. "MPI_Finish" on all the running hosts.

3. Exit except one host.

4. Choice a new set of hosts to run from the file generated with another
program.

5. Execute another "mpirun" on new hosts.

6. The last remaining host exits.

7. New hosts read in the informatin required to restart.

As you can see it is not easy but it is possible without buying extra
management software which could be quite expensive.

-- 
----------------------------------------------------------------------
Yon Han Chong                     |
Flow Control and Prediction Group | Tel: + 44 1234 750 111 ext. 5420
Cranfield College of Aeronautics  | E-mail: Y.H.Chong@cranfield.ac.uk
Cranfield University              | Fax: + 44 1234 752 149
United Kingdom                    | Mobile: + 44 958 381 395
----------------------------------------------------------------------

