[S-News] Splus, Sun's, and Solaris 2.6 OS

Wed, 4 Feb 1998 13:19:00 -0600

A couple of weeks ago I asked questions about experiences with Splus on
a Sun multi-processor (E450) using the Solaris 2.6 OS. I appreciate the
many personal responses I got that helped us better understand our
problems. I'd like to share our current beliefs about the problems we
were having in releasing memory. These beliefs are largely the result
of a significant effort by MathSoft support to study our problem; and
I'm quite grateful for that.

The problems we saw were also observed by Mathsoft on their Solaris 2.5
machine when they repeated my code, so neither the new 2.6 OS or our
multi-processor appears to be the cause of our problem. Mathsoft
suggested, "the version of malloc() we use on Solaris (and on most unix
machines except AIX and OSF/1) appears to slow down when their are lots
of small bits of memory allocated." That helped us understand why we
didn't see a problem with our old IBM RS/6000 (AIX) while we did see it
on our new Sun E450. Mathsoft has provided us with some malloc
replacements that we can try.

My "problem function" was actually a function that called about 30 other
functions and was certainly inefficient. I'm afraid that over time I
succumbed to the belief that my time was worth more than the computer's,
and I didn't worry much about inefficient programming. Mathsoft support
pointed out several inefficiencies in my code that I called a lot.
mean.na.omit <- function(x) mean(na.omit(x))
is much less efficient than
mean.na.omit <- function(x) { bad <- is.na(x); sum(x[!bad])/sum(!bad)
In fact, Mathsoft went on to suggest an even more effiencient function
for my particular application.

I now think that my problems were more the result of "allowing a
statistician (myself) to program" than of any substantial problems with
Splus or our Sun.

I again thank the many of you who personally responded to my question of
a couple weeks ago. And I'd likely to publicly state that Mathsoft
support was exceptionally helpful with this problem.

Allen Humbolt
Quantitative Analyst
Koch Industries, Inc.
HumboltA@kochind.com e-mail
(316) 828-7911 phone
(316) 828-8674 fax

This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news