Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.1 6/24/83; site bu-cs.UUCP
Path: utzoo!watmath!clyde!burl!ulysses!gamma!epsilon!zeta!sabre!petrus!bellcore!decvax!decwrl!pyramid!ut-sally!seismo!harvard!bu-cs!bzs
From: bzs@bu-cs.UUCP (Barry Shein)
Newsgroups: net.arch
Subject: Re: IOCALL results and problems
Message-ID: <835@bu-cs.UUCP>
Date: Fri, 3-Jan-86 16:36:05 EST
Article-I.D.: bu-cs.835
Posted: Fri Jan  3 16:36:05 1986
Date-Received: Sun, 5-Jan-86 01:31:53 EST
References: <380@ncr-sd.UUCP>
Organization: Boston Univ Comp. Sci.
Lines: 42
Keywords: IOCALL, Multiuser

Re: measuring multi-user system performance

The only sane thing I have heard in years is what DEC is doing with
ULTRIX (they say the method originated at AT&T) called Remote Terminal
Emulation (RTE.)

The basic idea is you put null modems on the tty mux's between two
systems (the one being measured and another, spare.) You then run
scripts through the mux from the spare machine and it records various
performance benchmarks (eg. response time, service time etc.)

Scripts can have random time distribution intervals so it needn't appear
that everyone (say, 32 terminal lines) are banging at once, one can go
so far as to simulate breaks etc by random long delays between type-ins.
Measurement could be done for hours or even days.

What is left at that point is to find scripts, one reasonable source
would be to record sessions on a current system (this would require, I
presume, backing up the system, starting recording, restoring onto the
test system and then re-running the script, it's hard, I agree, if it
is at all different then something is not going to work, and recording
things like the keystrokes to a full-screen editor would probably be
invasive, it would slow down the system being recorded and thus
probably alter users' behavior.) On the other hand, I fully believe
some reasonable compromise could be generated by simply observing
a system which could be considered analogous to the multi-user behavior
you are trying to measure.

There is obviously more to this story, I am not pretending to lay out
a methodology here, just an overview, but the idea is provocative, even
just in its obvious simplicity (wanna find out how a system performs when
it is being typed at? type at it!)

Does anyone know of or have the ATT references (I assume there are
Bell Lab TMs about this.)

	-Barry Shein, Boston University

p.s. this info derived from a session at DECUS on ULTRIX Performance
measurement by (?sorry, someone from) the ULTRIX Eng crew.

* Canonical trademark notice.