Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!linus!philabs!cmcl2!seismo!brl-smoke!smoke!mike@BRL.ARPA
From: mike@BRL.ARPA (Mike Muuss)
Newsgroups: net.unix-wizards
Subject: Re:  4.2bsd kernel auto-nicing, scheduling
Message-ID: <1027@brl-smoke.ARPA>
Date: Mon, 17-Feb-86 21:45:48 EST
Article-I.D.: brl-smok.1027
Posted: Mon Feb 17 21:45:48 1986
Date-Received: Wed, 19-Feb-86 20:30:16 EST
Sender: news@brl-smoke.ARPA
Lines: 27

A long time ago, for V6, I did a pretty snazzy scheduler that
incorporated short-term and medium-term behavior into it's
assessment of priorities.  The result was (and still is) PDP-11/70
systems with nearly instant (<200ms) response for "simple" interactions.

These days, I am firmly convinced that the proper way to implement
this would be with a user-mode daemon that collected medium-term
and long-term behavior patterns of processes, folded in issues
such as time of day, number of people on, "demo in progress" flags, etc,
and issued commands to the kernel adjusting various parameters
that would control the kernel's short-term scheduling behavior,
plus also providing per-process control (much like NICE currently,
but not user changable, and separate).

Creation of the necessary kernel mechanisms is nearly trivial.
The user-mode "Scheduling Advisor" could want to grow into a full
Expert System, but would probably be 90% satisfying if it just used
a half a dozen heuristics, most of which I believe I can articulate.

BRL currently has a low-level effort to implement this type of mechanism,
with two of our GITs (Gurus-in-training) working the problem.  If you
would like to contribute ideas, etc, please send them to <Jcst@BRL>
(Joint CS Team).  If we achieve any kind of success, we will notify
the Unix-Wizards list, so please don't just send "tell me when you are
done" messages.
		Best,
		 -Mike Muuss