home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.benchmarks
- Path: sparky!uunet!sun-barr!cs.utexas.edu!convex!convex!patrick
- From: Patrick F. McGehearty <patrick@convex.COM>
- Subject: Re: Surprised:gcc&dhryston benchmarks
- Originator: patrick@mozart.convex.com
- Sender: usenet@news.eng.convex.com (news access account)
- Message-ID: <1992Sep10.165714.9303@news.eng.convex.com>
- Date: Thu, 10 Sep 1992 16:57:14 GMT
- Reply-To: patrick@convex.COM (Patrick F. McGehearty)
- References: <1992Sep10.131908.17807@ntuix.ntu.ac.sg>
- Nntp-Posting-Host: mozart.convex.com
- Organization: Engineering, CONVEX Computer Corp., Richardson, Tx., USA
- X-Disclaimer: This message was written by a user at CONVEX Computer
- Corp. The opinions expressed are those of the user and
- not necessarily those of CONVEX.
- Lines: 42
-
- In article <1992Sep10.131908.17807@ntuix.ntu.ac.sg> eoahmad@ntuix.ntu.ac.sg (Othman Ahmad) writes:
- ...
- >Conclusion:
- >dhryston 1.1 is not yet useless because it actually gives a reasonable
- >processor speed indication for small programs. Sometimes this is what we
- >want.
- >
- In another posting I give some reasons which Dhrystone has the potential
- to be misleading. Those comments were for Dhrystone-2.1. Dhrystone-1.1 is
- even worse. A sufficiently advanced optimizing compiler can eliminate
- almost all of the code in Dhrystone-1.1. What do the measurements mean
- then? Nothing interesting.
-
- >My recommendations to Spec Mark consortium is to find small programs that
- >characterise a computer system characteristics, e.g. cache, RAM, disk,
- >Character, Graphic, serial, interrupt-response etc.
- >
- >Then for each software, what portions of those characteristics do they
- >consume. Cross-correlation with predicted results using those basic
- >characteristics for a software, and the actual results of running user
- >software should be done, to validate the "usefulness of those programs".
- >
- >Some of those small programs are
- >dhryston for integer small programs,
- >lorenze for floating point small programs,
- >iozone for portable I/O,
- >
- lorenze can be extremely misleading for predicting floating point
- performance. As written, there is a subroutine call for every few
- floating point operations. If you have a compiler which inlines
- that subroutine, AND you use the right options to get that inlining,
- your performance will be dramatically different (a factor of 2 on some
- machines) than if you don't inline the results. A factor of 2 is bit more
- than should be waved off as background noise. If you must use something
- that simple, try Dongarra's 100x100 Linpack benchmark (there are C versions
- floating around). Results are reported regularly for a long list of
- machines, and scrutinized by many people. Much better (but also much more
- work to get running) is to use the Perfect Club benchmarks. Another
- alternative would be to write some loops for common floating operations such
- as matrix multiply, correlation, sum reduction, etc. Do not include
- unnecessary subroutine calls in the inner loops as that just causes the
- program to be measuring something besides floating point performance.
-