Path: utzoo!mnetor!uunet!husc6!ukma!gatech!hubcap!Lori
From: grob@cmcl2.NYU.EDU (Lori S. Grob)
Newsgroups: comp.parallel
Subject: Re: Sandia scaling results
Message-ID: <1303@hubcap.UUCP>
Date: 5 Apr 88 12:06:33 GMT
Sender: fpst@hubcap.UUCP
Lines: 22
Keywords: formula (theoretically anyway) for determining the number of processors
Approved: parallel@hubcap.clemson.edu


There is a paper by Ron Cytron of IBM Research that was at the ICPP a few
years ago, that presents an analysis of performance with respect to the 
size of a task and the processor allocation for that task. It is theoretical
and does not claim to be for all topologies but in light of this discussion
I thought that the folks who did not  know about it might be interested.

Within the constraints of the problem as discussed in the paper it gives
the formula for figuring out the "breakeven point" for adding processors to
a problem. ie. The point beyond which it will cost more to throw more
processors at the problem then you will get help from those processors.

The paper applies the formula to a DOALL loop and to a vector sum reduction.

I think it was the 85 or 86 ICPP conference.



Lori S. Grob (NYU Ultracomputer Project)
grob@nyu.arpa
{mcvax!seismo,floyd,harpo,ihnp4,...}!cmcl2!grob   [That's c-m-c-ELL-2]
Courant Institute (NYU), 251 Mercer St., NYC 10012,  212-998-3350