Path: utzoo!attcan!uunet!husc6!purdue!gatech!hubcap!brooks
From: brooks@lll-crg.llnl.gov (Eugene D. Brooks III)
Newsgroups: comp.parallel
Subject: Re: parallel numerical algorithms
Message-ID: <1688@hubcap.UUCP>
Date: 24 May 88 12:44:41 GMT
Sender: fpst@hubcap.UUCP
Lines: 22
Approved: parallel@hubcap.clemson.edu

In article <1684@hubcap.UUCP> eugene@pioneer.arc.nasa.gov (Eugene N. Miya) writes:
>a computer scientist [he was humbled by our problems].  Then he placed a view
>graph of a 3-D plot of Amdahl's law, one axis was "percent vectorization" and
>the other domain axis was "percent parallelism" done using the well known
>NCAR graphics package displaying the spike at 100% vectorization and
>parallelization (blah, what a word).  This graph disturbed me, there was
>...

His point was that on a parallel/vector architecture such as the Cray X/MP
or Cray 2 series you have to do well on both the MIMD parallel and Vector
fronts in order to approach the peak performance for a machine.  Although
the space of applications may not be evenly distributed on the x-y surface
of his plot, the notion that it presented that a rather small number of
applications would do well on both the MIMD and Vector parallel fronts
seems to be quite valid and is not disturbing at all.  A good example of
an application that fits the distinction of (MIMD,Vector) parallel is
Gauss elimination.  One vectorizes the SAXPY operations to mask memory and
functional unit latency, one parallizes across the rows of the matrix to
use more than one vector CPU in a MIMD mode.  Of course, there is more than
one way to skin this cat, but the notion of simultaneously exploiting MIMD
parallelism and Vector (SIMD) parallelism in a given architecture or algorithm
is quite useful.