Newsgroups: comp.parallel.pvm
Path: ukc!uknet!pipex!howland.reston.ans.net!EU.net!Germany.EU.net!netmbx.de!zib-berlin.de!math.fu-berlin.de!inf2hro!infhro!tbahls
From: tbahls@informatik.uni-rostock.de (Thomas Bahls)
Subject: Re: algorithm for default spawning?
Message-ID: <Cn6HKC.K47@informatik.uni-rostock.de>
Sender: news@informatik.uni-rostock.de (News Account)
Organization: University of Rostock, CS Dept. (Germany)
X-Newsreader: TIN [version 1.1 PL9]
References: <2mn0tq$f5k@quabbin.crl.dec.com>
Date: Thu, 24 Mar 1994 16:52:58 GMT
Lines: 28

William Hsu (hsu@crl.dec.com) wrote:
: When spawning a number of tasks (pvm_spawn) with the flag set to
: PvmTaskDefault pvm is allowed to choose the processors as it seems
: fit. I'm a little disconcerted that it sometimes chooses to run two pvm
: processes on a single machine before using all available machines.
: That is, when spawning a set of tasks from machine A, with machines
: A, B, C, D configured in pvmd, sometimes the children will be 
: spawned on B, C, A, instead of B, C, D.  Can someone explain to me
: the reasoning behind this?

That's extremely interesting. The behaviour of pvm_spawn you got is right
that what I want to get. But I cannot. I am confused about it now.

I wanted to check out the "intelligence" of PVM some days ago cause it is
said so in the manual (...transparent mode: process is executed on the
least loaded computer...). Hence I configured a machine of two RS/6000 and
executed on the one a complicated numerical integration method to increase
the load of this machine. What happened ? Right the same like without the
integration, though the one machine *was* very slow indeed.
To sum it up, PVM always spawnes the tasks "uniformly distributed" over the 
entire virtual machine. It doesn't depend on load etc. as mentioned in the
manual. How did you get those behaviour ???
I use PVM 3.2. Is it an old version now ? Isn't sth. implemented yet ???

I'm interested in any ideas. E-mail me, please.

TIA, Tom.  (tbahls@informatik.uni-rostock.de)


