home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!mcsun!uknet!yorkohm!minster!mjl-b
- From: mjl-b@minster.york.ac.uk
- Newsgroups: comp.sys.atari.st
- Subject: Benchmarking (Was: Re: ST FORMAT DO IT AGAIN..)
- Message-ID: <721569048.8500@minster.york.ac.uk>
- Date: 12 Nov 92 11:50:50 GMT
- References: <1992Nov9.103328.24501@dcs.warwick.ac.uk> <1992Nov9.185921.9035@beaver.cs.washington.edu> <1992Nov10.152620.1302@dcs.warwick.ac.uk>
- Reply-To: mjl-b@minster.york.ac.uk (Mathew Lodge)
- Organization: Department of Computer Science, University of York, England
- Lines: 56
-
- In article <1992Nov10.152620.1302@dcs.warwick.ac.uk> leo@dcs.warwick.ac.uk (Leo Hendry) writes:
- >In article <1992Nov9.185921.9035@beaver.cs.washington.edu> bgilbert@cs.washington.edu (Ben Gilbert) writes:
- >>Just remember that MIPS stands for "Meaningless Indicator of Processor Speed"
- >>(and various similar things :-). Although the DSP is rated somewhere above
- >
- >I compiled a Drhystone test the other day which gave a rating of appoximately
- >1500 (Using Lattice C5 compiler). I was reading an Amiga newsgroup for
- >details of the A1200 and somebody quoted a figure of appox 1250 d'stones which
- >they said is 2.2 times faster than a standard Amiga 500.
-
- > Does anybody know of an accurate speed test which works across a variety of
- >machines?
-
- Now _there's_ a can of worms! Devising an accurate indication of performance
- that works across a variety of machines is extremely hard. It's the goal of
- "the general benchmark", and it may be an unattainable goal.
-
- Dhrystone is not a good general benchmark for a number of reasons:
-
- 1) It contains few loops. Studies have shown that most programs contain a
- high percentage of loops, and that code inside loops is significantly
- different to that found outside.
- 2) Large parts of the V1 benchmark can be optimised away by "dead code"
- removal.
- 3) Dhrystone was originally written in Ada, then converted to C. The
- conversion is not faithful to the Ada original, particularly when it comes
- to string operations
- 4) The poor analysis of operators lead to the inclusion of a number of
- string operations. These string operations introduce a significant amount of
- computation and looping into a program that contains few loops. Hence,
- Dhrystone results depend very much on the efficiency of the string
- operations in your prticular library.
-
- If you want to read a discussion of why general benchmarking is hard, and
- a critique of various benchmarks, you could check out the following
- (sorry, couldn't resist):
-
- MJ Lodge, "Benchmarking and Ada", 3rd year MEng Project Report, University
- of York, UK.
-
- About the best benchmarks around are the SPEC suite, a set of application
- based benchmarks. Most "high end" manufacturers (Atari isn't one of them)
- are members of SPEC, and the benchmarks tend to be written in C for Unix
- machines.
-
- In summary: determining "general" performance is hard, and error prone. If
- you do need to measure performance, then base it on the relative performance
- of a given application that is relevant to your area of interest.
-
- >- Leo
-
- Mat
-
- | Mathew Lodge | "I don't care how many times they go |
- | mjl-b@minster.york.ac.uk | up-tiddly-up-up. They're still gits." |
- | Langwith College, Uni of York, UK | -- Blackadder Goes Forth |
-