Re: Q: least-square fit for performance model

Henry Spencer <henry@zoo.toronto.edu>
7 Dec 1997 22:10:00 -0500

          From comp.compilers

Related articles
Q: least-square fit for performance model chunghsu@paul.rutgers.edu (Chung-Hsing Hsu) (1997-12-05)
Re: Q: least-square fit for performance model henry@zoo.toronto.edu (Henry Spencer) (1997-12-07)
| List of all articles for this month |

From: Henry Spencer <henry@zoo.toronto.edu>
Newsgroups: comp.benchmarks,comp.compilers,comp.arch
Date: 7 Dec 1997 22:10:00 -0500
Organization: SP Systems, Toronto
References: 97-12-039
Keywords: benchmarks, performance

Chung-Hsing Hsu <chunghsu@paul.rutgers.edu> wrote:
>Is there any literature discussing how suitable it is to use
>least-square fitting to get the performance model of a set of
>benchmarks? ...


A small comment on statistical methods: there are good reasons *not*
to use least-squares fitting if the number of data points is small and
the data is noisy. Least-squares fitting is not robust; even with
large numbers of points, it can give badly distorted answers if most
of the data follows an obvious pattern but there are a few "outlier"
points far outside the pattern. With small numbers of points the
answers can be almost meaningless.


Professional statistical advice is important for such attempts to
extract meaning from very limited data. You should seek it *before*
you start work, not after you've gathered your data -- it may well
influence how much data you gather, and how you gather it. Such
advice is best sought in a Statistics or Applied Math department, not
in a book or on Usenet or the Web, and you'll probably have to pay for
it. It's worth it.
--
| Henry Spencer
| henry@zoo.toronto.edu
--


Post a followup to this message

Return to the comp.compilers page.
Search the comp.compilers archives again.