Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!husc6!bloom-beacon!gatech!hubcap!delp
From: delp@udel.EDU (Gary Delp)
Newsgroups: comp.hypercube,comp.arch
Subject: Amdahl's Law
Message-ID: <449@hubcap.UUCP>
Date: Mon, 14-Sep-87 13:32:40 EDT
Article-I.D.: hubcap.449
Posted: Mon Sep 14 13:32:40 1987
Date-Received: Tue, 15-Sep-87 06:22:54 EDT
Sender: fpst@hubcap.UUCP
Lines: 27
Keywords: amdahl, speedup, parallel
Approved: hypercube@hubcap.clemson.edu
Summary: Do you have references?
Xref: mnetor comp.hypercube:85 comp.arch:2172

I have heard the following observation attibuted to Gene Amdahl, but
have not been able to find it in the literature.  If you have seen it,
and have some hard information on it, I would appreciate it if you
would let me know directly.  I will report back on the results.

The "Law" as I understand it states that there is a fundamental limit
to the speedup that can be acheived by taking a serial task and
running it on a parallel processor.  This limit is set by the portions
of the task that still be serialized.  the formula is:

           S + P            S = time required by serial portion of task
speedup = --------- , where P = total time required by parallel portion
           S + P/N          N = Number of Processors

This obviously ignores such factors as communication overhead and
latency, but is meant to provide a fundamental limit, not a firm
prediction of performance.

Harold Stone's new book "High-Performance Computer Architecture",
which is, by the way, a fine addition to the choices of textbooks in
the field, covers the issues in this area without mentioning the
fundamental limit.

Gary Delp               ARPA:   delp@udel.edu
123 Evans Hall          BITNET: delp@udel.edu
EE U of D               CSNET:  delp%udel.edu@relay.cs.net
Newark, DE 19716