COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	.require "memo.pub[let,jmc]" source
C00009 00003	.skip 1
C00010 ENDMK
C⊗;
.require "memo.pub[let,jmc]" source;
.turn off "{"
.cb HARDWARE PERFORMANCE PROBLEMS OF TIME-SHARED COMPUTERS


	Time-shared computers have never performed as efficiently
as their designers have hoped.
There are many problems in the operating systems, but the thesis
of this article is that no present hardware admits really efficient
time-sharing systems.
We will try
to identify some problems and suggest possible solutions.

	The machines we shall consider are the PDP-1D, the PDP-6,
the PDP-10 (in its successive KA-10, KI-10, KL-10, 2040 and
2050 versions), and the IBM 7090, the IBM 360/67, the IBM 370/168,
and the Honeywell 645.  I think other machines have had similar
problems for the reasons I will advance, but I know less about them.

	The symptoms are simple.  An advertised increase in computer
power looks real at first but doesn't translate into a proportionate
increase in the number of users the machine can handle at a given
level of service.  Moreover, once installed,
a given machine seems to get worse with time!

	The general diagnosis is that the machine and time-sharing
system designers have neglected the following facts:

	1. A large part of the use of time-sharing systems is in
programs that do an extremely small amount of computation per
interaction with the user.  A consequence is that the computation
involved in scheduling and other overhead
can dominate the computation the user gets.
	
	2. As applications develop, features are added to programs,
and each feature increases the amount of memory with which a user
can potentially interact - even when the number of actual memory
references per interaction does not increase.

Reactions to this situation include the following:

	1. Paging.

	2. Cache memories.

	3. Dedicated mini and micro computers.

	4. Systems of mini and micro computers.

	5. People shouldn't write that kind of program.
.skip 1
.begin verbatim
John McCarthy
Artificial Intelligence Laboratory
Computer Science Department
Stanford University
Stanford, California 94305

ARPANET: MCCARTHY@SU-AI
.end

.turn on "{"
%7This draft of

PUBbed at {time} on {date}.%1