home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.arch
- Path: sparky!uunet!ukma!netsys!decwrl!elroy.jpl.nasa.gov!news.claremont.edu!jarthur.claremont.edu!tlilley
- From: tlilley@jarthur.claremont.edu (Thomas (Ted) Lilley)
- Subject: Re: Hardware Support for Numeric Algorithm
- Message-ID: <1992Nov13.195926.29438@muddcs.claremont.edu>
- Sender: news@muddcs.claremont.edu (The News System)
- Organization: Harvey Mudd College, Claremont, CA 91711
- References: <TMB.92Nov13104609@arolla.idiap.ch> <1992Nov13.141300.3528@clark.dgim.doc.ca>
- Date: Fri, 13 Nov 1992 19:59:26 GMT
- Lines: 25
-
- In article <1992Nov13.141300.3528@clark.dgim.doc.ca> don@mars.dgrc.doc.ca writes:
- >I believe correct results are far more important to Herman than the wrong answer
- >3 times as quickly. Therefore the folks arguing over which is more important,
- >correctness or speed, are also wasting bandwidth.
-
- Hi,
- There is one comment I'd like to make on this, which is that no one
- who's mentioned correctness and speed has discussed it properly, which is to
- say, in terms of the expected running time to find the correct answer.
-
- If there exists an algorithm which runs 6 times as fast as Herman's current
- algorithm, but gives a wrong answer 1/3 of the time, then he might expect to
- have to run it 3 times for a correct answer, which is still twice as fast as
- his current algorithm. Since factorization is easy to check, a wrong answer
- would not impose any serious time costs in that event.
-
- So you see, you should not argue over which is better, correctness or speed,
- without considering which type of algorithm you can expect to give you the
- right answer in which amount of time.
-
- I expect Herman already knows this, as most people who work with numeric
- algorithms of that sort are familiar with probabilistic algorithms and
- analyzing their running time.
-
- Ted
-