Newsgroups: comp.parallel.pvm
From: beaudoin@robot.ireq.ca (Martin Beaudoin)
Subject: Question regarding Hence and PVM
Organization: La division Robotique de l'Institut de recherche d'Hydro-Quebec
Date: Wed, 25 Sep 1996 17:50:43 GMT
Message-ID: <BEAUDOIN.96Sep25135043@ozias.robot.ireq.ca>

I've got a question regarding the package hence.

From my understanding of the FANIN/FANOUT node, every subgraph
included between my FANIN and FANOUT nodes will be executed in
parallel, BUT all the subgraphs need to be terminated before the
FANOUT node can be reach.

So if I include a FANIN/FANOUT block inside a LOOP block, the LOOP
block will begin the next iteration only when all the subgraph inside the
FANIN/FANOUT have terminated.

So, a FANIN/FANOUT group execution time is determined by the slowest
machine that will be use for computing the FANIN/FANOUT group...

Any comments on this?

What I'm trying to do is this:

Using a network of workstations with big differences in CPU power,
compute the same algorithm in parallel on all the CPU avalailable, but
on different data set. 

So if machine A is twice as fast as machine B, I would expect machine
A to crunch twice as much data as machine B....

I also want to limit the number of identical tasks that may run on a
given machine (1 or two at the same time, at most). So if machine A is
to run the task 50 time on a given data set, I don't want machine A to
run these 50 tasks simultaneously but only 1 or 2 at a time.

I cannot seems to find a proper conctruct using Hence to do just that;
I want asynchronous tasks dispatching, and Hence seems to like the
synchronous way of life....

Any comments are more than welcomed.


--
Martin Beaudoin                            beaudoin@robot.ireq.ca
Institut de recherche d'Hydro-Quebec(IREQ)
Division Robotique
1740 montee Ste-Julie			      tel: (514) 652-8235
Varennes (Quebec) J3X 1S1 Canada	      fax: (514) 652-1316

