Newsgroups: comp.parallel,comp.parallel.pvm
From: wrankin@ee.duke.edu (William T. Rankin)
Subject: Re: network parallel politics (net load)
Organization: Duke University EE Dept.; Durham, NC
Date: Tue, 21 Jun 1994 12:17:37 GMT
Message-ID: <Crqy5D.185@dcs.ed.ac.uk>

In article <CrG92z.FLA@dcs.ed.ac.uk>, edwards@cartel.colostate.edu (Jim Edwards) writes:

|> Can anyone point me to information regarding the load placed
|> on the network and on hardware on the same LAN but NOT running
|> a parallel application?  

The problem with trying to produce some number that satisfies this question
is that the amount of net bandwidth usage is very dependant upon the
granularity of the application being run.

For instance, I have a larger parallel program running currently on a
network of about 30 workstations.  It only produces a few small messages
per second, so it's impact on the network is not even measurable.
(It really kills the CPUs however :)

On the other hand, I am developing some Molecular Dynamics code
(FAWI: Parallel N-body using variations of Greengard's FMA) that 
has simultaneous large data transfers between all processors.
Theis code has a large impact on the network for the brief time
it is running.

Both programs run using PVM.  Both programs use the same clusters 
of workstations.  But the loads they produce are radically different.

|> The problem is that the site administrator here would like 
|> to limit access to the net for parallel applications.  
|> Her claim is that our application would negatively affect 
|> others on the same LAN.    

Yeah, the old "let's not waste those cycles getting work done,
they're expensive and the netrek-ers may get irritated" attitude.
Sheesh, damn cromagnon-man systems administration.  (Sorry, this
is a *very* sore point with me.  Never would have guessed, eh? ;)

You may wish to point out (in as non-confrontational a way as
possible :)  is that by the same logic, students should not be able to
run large Spice simulation jobs because it dramaticly impacts other
users CPU usage and response times.  In fact, access to the computers
themselves should be very very restricted unless your program is of
CRITICAL DEPARTMENTAL IMPORTANCE like, perhaps, the Deans LaTeX'd memo
regarding the current waste of paper on needless memos.

Anyway, one thing you can suggest, and what we've done here, is
have the workstation cluster you intend to use but on a subnet so that
the PVM traffic doesn't impact the rest of the departments work.
In addition, suggest that "they" may also want to run one of the
interactive network traffic monitors that will let them know 
*exactly* where their problems are wrt. bottlenecks.

The funny thing is that the reason we had to subnet around here 
was the DSP group here was running large simulations that produced
megs and megs of data that *all* went across NFS to a local server.
The network monitoring software revealed this problem.

|> Can anyone provide documentation to show that this is not the case?
|> Or, if it is, what can be done without completely eliminating
|> the possibility of parallel computing across the net? 

See above.  I guess the bottom line, IMNSHO, is that if you are worried 
about the net impact of the occasional PVM user, then most probably the
real problem is not PVM.

|> -- 
|> Jim Edwards 
|> edwards@cartel.atmos.colostate.edu

Good luck with it.

-- 
----                                /       __/    /    /
bill rankin                        /              /    /
wrankin@ee.duke.edu               ___  /    /    /    /
philosopher/coffee-drinker       /    /    /    /    /
                                /    /    /    /    /
                             _______/  __/  __/  __/





