Newsgroups: comp.parallel
From: dbakken@bbn.com (David Bakken)
Subject: Re: Help - explain superlinear speedup?
Organization: BBN Inc.
Date: Tue, 17 Jan 1995 12:57:25 GMT
Message-ID: <3feb9q$hhj@info-server.bbn.com>

In article <3f61t5$etd@mozart.convex.com>,
Joel Williamson  <joelw@convex.convex.com> wrote:
>Superlinear speedups always result from additional resources (almost
>always memory) relieving bottlenecks in the computation.  So either the
>additional physical memory eliminated some paging and/or the problem fit
>in cache once you got to four processors.

I would say "usually" here, not "always".  Superlinearity can sometimes be
the result of architectural specialization.  Consider a program that does
some searching and some matrix multiplication.  Assume that, for this 
program, the searching runs fast on a hypercube but much slower on
a MIMD multiprocessor, and the matrix multiplication runs slow on the 
hypercube but much faster on the MIMD machine.  If the program can be
structured so that the searching can be done on the hypercube while the
matrix multiplication is done on the MIMD multiprocessor, then the program
may finish in much less time (at least less than half) than either machine
could run the program by itself.
-- 
David E. Bakken 	dbakken@bbn.com		+1 617 873 6072
Distributed Systems Department; BBN Systems and Technologies;
10 Moulton St. MS 6/3D; Cambridge, MA 02138 USA


