Re: Benchmarks - 1.3.11

Larry McVoy (lm@neteng.engr.sgi.com)
28 Jul 1995 21:08:52 GMT


cloister bell (cloister@hhhh.org) wrote:
: > Date: Sun, 23 Jul 1995 12:47:53 -0400 (EDT)
: > From: dholland@husc.harvard.edu
: > To: linux-kernel@vger.rutgers.edu
: > Subject: Re: Benchmarks - 1.3.11
: >
: > > > Each test is run 20 times (10 for the most time-consuming). If I find the
: > > > time, I will add the calculation of the standard deviation.
: > >
: > > I would be interested in seeing that. Its all fine and well to say
: > > that task switching times or something has increased + or - 0.82% over
: > > the previous version, however when there haven't been any changes that
: > > could make a difference in such a thing, then you have to figre that
: > > you are actually detecting something else.
: >
: > Any change can make a small difference in just about anything. Suppose
: > somebody added three lines of code to the SCSI driver, and this (by
: > being slightly larger) caused the task-switch code to lie across a
: > page boundary. Presto! A few more cycles every task-switch, probably,
: > and perhaps a 1% drop in switching performance.
: >
: > Don't underestimate the effects of things like that.

: this is a good point.

This is a great point, not a good point. Another way to look at it: a
cache miss is going to cost you at least 300ns, or 30 instructions at
100Mhz. Every time you put a

if (my_new_feature_is_enabled) {
}

type of check in a critical path, add a usec or so of time to that path.
Of course, before & after measurements show no change so it is fine,
right? Uh-uh.

: it reminds me of a story i heard (actually, it was
: an internal memo that got leaked to the outside world) about the o/s
: developers at sgi. if memory serves, the sgi marketers were shouting
: "features! features!" at them, and wouldn't let them do any performance
: work.

This did indeed happen. The funny thing is that I was at Sun when I
heard about it and I was saying the exact same thing to Sun's mgmt.
I think you can correctly guess which mgmt chain cares about performance
based on my job location :-)

--
---
Larry McVoy			(415) 390-1804			 lm@sgi.com