Fun_People Archive
22 Aug
Prodding for Parallel Program Performance

Date: Tue, 22 Aug 95 12:25:47 -0700
From: Peter Langston <psl>
To: Fun_People
Subject: Prodding for Parallel Program Performance

Forwarded-by: bostic@bsdi.com (Keith Bostic)
Forwarded-by: "Harry I. Rubin" <harry@farside.sj.unisys.com>

	    NAS Special Presentation 


Speakers:   Sekhar Sarukkai, Recom Technologies
	    Rob Van der Wijngaart, MCAT Incorporated
	    NASA Ames Research Center

Topic:	    Prodding for Parallel Program Performance
	    (Techniques and Tools for Performance Tuning of
	    Parallel and Distributed Scientific Applications)

Abstract:   Many papers on high-performance computing talk about
	    harnessing the power of parallel machines, as if there were this
	    uncontrollable source--like nuclear fusion--under the slick
	    surface of the computer, ready to be unleashed. But there are
	    no rampant compute demons running around inside such machines.
	    Instead, machines are akin to a bunch of happy-go-lucky elves
	    who take their time chatting and twiddling their thumbs. These
	    creatures should be prodded, not harnessed.

	    This presentation discusses where and how to apply the parallel
	    performance prods. The first part of the presentation
	    categorizes scientific applications in terms of data accesses
	    and data dependencies; important factors influencing performance
	    of some representative examples are considered. In the second
	    part, techniques for node code and parallel (message-passing)
	    code optimizations are discussed systematically. The third and
	    final part presents application of available tools in measuring
	    and tuning program performance.

	    Note: This presentation is a compressed version of a tutorial
	    that the speakers will present at the International Parallel
	    Processing Symposium `95.  Also note the extended time.

© 1995 Peter Langston