home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!olivea!decwrl!usenet.coe.montana.edu!rpi!uwm.edu!ogicse!borasky
- From: borasky@ogicse.ogi.edu (M. Edward Borasky)
- Newsgroups: comp.benchmarks
- Subject: Regression analysis
- Message-ID: <42239@ogicse.ogi.edu>
- Date: 1 Sep 92 20:56:48 GMT
- References: <1992Aug20.160352.13856@nas.nasa.gov> <1992Aug23.114309.3643@nosc.mil> <1992Aug26.160240.20114@murdoch.acc.Virginia.EDU> <1992Aug31.002356.24988@nosc.mil> <Btwo88.4v0@nntp-sc.Intel.COM>
- Distribution: comp.benchmarks
- Organization: Oregon Graduate Institute (formerly OGC), Beaverton, OR
- Lines: 58
-
- In article <Btwo88.4v0@nntp-sc.Intel.COM> jwreilly@mipos2.UUCP (Jeffrey Reilly) writes:
- >...
- >used a regression analysis to derive the following equation:
- > SPECmark = (0.77 x mips) + (0.70 x mflops) - 3.27.
- Question: Where do the MIPS and MFLOPS numbers come from?
-
- >SPEC's initial thought was to take the marketing philosophy that
- >there is no such thing as bad publicity, and ignore the article.
- Is this "marketing philosophy" meant to imply that those in
- marketing have this philosophy?
-
- >A SINGLE NUMBER FOR THE PRESS
- >SPEC has always acknowledged that system performance couldn't be
- >accurately represented by a single number. ...
- > ... Only in consideration of the
- >press, did SPEC generate a composite of the 10 numbers, the
- >SPECmark.
- Seems pretty inconsiderate to me to hold that performance can't
- be accurately represented by a single number, then create one
- "in consideration of the press." Single numbers are created
- because people ask for them. Averages are created because
- people don't trust peak speeds. It would be nice if SPEC could
- estimate the appropriate-dimensional Amdahl's Law surfaces for
- their benchmarks; if they ask me (REAL NICELY) I'll show them how
- to do it. In the absence of this kind of analysis, an average
- (with appropriate confidence limits, which I can also show them
- how to do) seems to be a pretty reasonable representation of relative
- computer performance, especially when dealing with commercially-
- successful systems which are, since they ARE commercially successful,
- by nature balanced for reasonably common workloads.
-
- I went through all this stuff several years ago; I went around and
- around with all of these issues and I found things that worked.
- Admittedly, I was dealing with the simpler case of supercomputers:
- CPU bound applications, performance measured in megaflops. And I
- was working in a marketing organization at the time.
-
- >REGRESSION ANALYSIS
- >A regression analysis in some fields seems to be a reasonable way
- >to estimate future values. Horse racing, weather predicting, and
- >computer performance predictions seem to fall outside the
- >predictive capabilities of this tool.
-
- 1. For sprint races, where raw speed is the most important
- factor, you can ALMOST make a profit by betting on the horse
- (or DOG) with the lowest AVERAGE time for the given race
- length. It has a considerably better expectation than the ball-
- type lotteries, Las Vegas slot machines or buying call options
- on Intel in a bear market.
-
- 2. Least squares techniques are one of many methods used in
- computational fluid dynamics, including weather forecasting.
-
- 3. I use non-linear regression to predict supercomputer performance
- by fitting Amdahl's Law curves to benchmark data. It works. It's
- a little more complicated than the above-noted formula, but it works.
- The more data points I have the better it works. I can show other
- people how to do it.
-