Time spent handling interrupts, context switches

[Dominic Dunlop]

In article <36715 at bbn.COM> ellard at BBN.COM (Dan Ellard) writes:
>Has anyone ever done any testing to measure the percentage
>of time the UNIX kernel spends handling interrupts (i.e. in the interrupt
>context), or in context switching for moderately loaded systems? 

Firstly, /usr/group's technical committee has a working group concerned
with performance measurement.  I'm sorry to say that I can't put my hands
on an email address for a contact person.  Can somebody in netland please
post one?  Although those who attend (typically bimonthly) meetings get the
most out of these working groups, some have associated mailing lists which
aid participation by those who cannot attend.  Again, I don't know if this
is the case with performance measurement.  Clarification please, somebody?

Secondly, well, there I was thumbing through AT&T Technical Journal,
Vol 67 #5 (Sept-Oct 88) yesterday (I was waiting for a thirty-one page
fax to arrive, but that's another story).  In it I found

An Analytical Model for UNIX Systems
Gopalakrishnan Ramamurthy

``This paper presents an analytical model for a single-processor
interactive computer system running under the UNIX operating system.  The
model is a closed multichain, multiclass priority-queueing network model.
The model solution is an approximation and is based on mean value analysis.
The validity of the model has been established by using it with a number of
workloads and workload mixes and with a number of different computer
systems.  mean response times predicted by the model are within 10 percent
of measured values.  The model can be extended to tightly-coupled
multiprocessor systems and multiple UNIX systems running with network file
systems over local-area networks.''

and also

A Performance Study of the UNIX System V Fork System Call using Casper
Ronald E Barkley and Curt F Schimmel

``In this paper, we describe a general-purpose timing and tracing package,
called CASPER, for UNIX System V.  We also present the results of a study
of the behaviour of the UNIX System V fork system call using this package.
CASPER is designed to support high-resolution timing and tracing of
software under UNIX System V.  With CASPER, we were able to obtain detailed
and accurate information on the behavoiur of the fork system call, which
led to minor algorithmic changes that reduced the time to execute small
forks by about 15 percent.''
>
>I'm mostly interested in data about BSD-flavored UNIX or MACH, but
>info about any other flavor of UNIX or other operating systems like
>VAX/VMS, OS/360, MULTICS, etc. would be useful too.

These papers are specifically about UNIX System V release 2 (Ramamurthy)
and release 3 (Barkley & Schimmel).  The latter paper cites 

CASPER the Friendy Daemon
R E Barkley, D Chien
Proceedings of the 1988 Summer Usenix Conference

which might also be worth a look (must get around to ordering a copy of
those proceedings).

The math in the Ramamurthy paper, although not extensive, is not for
amateurs (I'm definitely an amateur in this context); readers of the
Barkley & Schimmel paper would benefit from having access to System V,
release 3 source code (I don't).

Sales message: I'd say that the AT&T Technical Journal is well worth $40
per year (US) or $50 per year (foreign), although potential subscribers
should be warned that it covers all manner of technical topics
(telecommunications, materials science, economics, human factors...), not
just UNIX.  The subscription address is AT&T Bell Laboratories
Circulation Group, Room 1K-424, 101 J F Kennedy Parkway, Short Hills,
NJ 07078.  (Disclaimer: I am not assocated with \*(oR other than as
\*(rE...)

-- 
Dominic Dunlop
domo at sphinx.co.uk  domo at riddle.uucp