Path: utzoo!mnetor!uunet!husc6!rutgers!iuvax!pur-ee!j.cc.purdue.edu!k.cc.purdue.edu!l.cc.purdue.edu!cik
From: cik@l.cc.purdue.edu (Herman Rubin)
Newsgroups: comp.arch
Subject: Re: Single tasking the wave of the future?
Message-ID: <635@l.cc.purdue.edu>
Date: 25 Dec 87 21:35:32 GMT
References: <18@amelia.nas.nasa.gov> <2341@encore.UUCP> <25@amelia.nas.nasa.gov> <1030@alliant.Alliant.COM>
Organization: Purdue University Statistics Department
Lines: 26
Keywords: parallel processing today
Summary: It ain't necessarily so

In article <1030@alliant.Alliant.COM>, muller@alliant.Alliant.COM (Jim Muller) writes:
								As for
>    parallelism being harder to exploit than vectorization, the reverse is more
>    likely to hold.  Parallel execution can be applied to practically any
>    vectorizable structure, while the reverse is decidedly not true. 

If you are doing even moderately efficient methods of generating non-uniform
random numbers, this is likely to be false.  A flexible vector structure like
the CYBER 205 makes the acceptance-rejection step trivial.  A rigid vector
structure like the CRAY makes it somewhat harder.  On a MIMD machine, we
have to wait for the unluckiest processor to finish.  On a SIMD machine, we
have this cost, and also must have both an acceptance and a rejection step
for each cycle which has a rejection, as well as testing at each cycle
whether a rejection has occurred _on any processor_.

							More to the point,
>    however, for those problems which do not exhibit inherent parallelism,
>    nothing requires the machine (or compiler) to use it where inappropriate.
>    All that is needed is that the unused processors be doing something else.

This eliminates SIMD machines, which are currently the ones with the greatest
(by far) parallelism.
-- 
Herman Rubin, Dept. of Statistics, Purdue Univ., West Lafayette IN47907
Phone: (317)494-6054
hrubin@l.cc.purdue.edu (ARPA or UUCP) or hrubin@purccvm.bitnet