Aduke.2000
net.works
utzoo!decvax!duke!bcw
Thu Apr 8 22:51:09 1982
Re: Workstation power
From: Bruce C. Wright @ Duke University
Re: Workstation power
Recently there has been quite a bit of discussion about computer
system power as a function of MIPS. Although for some uses this
is quite appropriate, it has been quite some time since most
computing systems have been designed and purchased solely on the
basis of MIPS.
MIPS only give you the potential for CPU-bound tasks (and even
then not very well, as I'll demonstrate later). For I/O-bound
tasks (most data processing, database manipulation, editing,
compilation, linking, word processing and so forth), a much
more cogent measure can be I/O bandwidth. For example, recently
here at the Medical Center where I work, we put together an LSI-
11/23 system with large (60MB) disk drives. We had to be rather
careful which controller we purchased, because one of these high
performance disk drives will simply saturate the poor Q-bus if
there are long (multi-block) transfers. Since it wasn't really
practical to modify the software to change multi-block requests
into a series of single-block requests, we had to ensure that
any controller we got for the system would automatically buffer
the request in some way (say by breaking it up into smaller
chunks and letting the drive spin around an extra time sometimes).
The resulting system has about the CPU power of a PDP-11/34, but
I can attest that it was *noticeably* slower for most types of
computing, though we never made any performance measurements.
Even for CPU speed, MIPS can be misleading without an understanding
of the application and the hardware than such a vague number is
able to give. For example, MIPS really don't have much meaning
if the CPU doesn't have a floating point instruction set and you
need to do a lot of floating-point math; ditto for decimal and
to a somewhat lesser extent character instruction sets. Another
example is interrupts: one of the fastest machines for large
amounts of interrupt processing is the TI-990 series, since it
takes only about 1-2 instruction times to enter an interrupt,
save the registers and context, and be ready to begin processing
the interrupt service routine. Try that on your PDP-11, VAX, or
MC68000! The point is that this architecture was specially
optimized for a highly interrupt-driven environment (such as
process control) and it would take an enormous amount of hard-
ware to make those other architectures run as quickly in that
environment.
MIPS only give the barest glimpse of the actual CPU speed;
what really matters in questions of system speed is the total
match of the system to the application.
Even apart from simple performance criteria, there are a lot
of other reasons why specific machines are chosen: software,
for example. If you have specific software needs, you may have
to forego the cheaper hardware in order to lower your total
system costs.
Back in the 1940's and early 1950's, it might have had some
meaning to assign a single number to a computing system, but
this technique no longer has any particular merit.
Bruce C. Wright @ Duke University
-----------------------------------------------------------------
gopher://quux.org/ conversion by John Goerzen <
[email protected]>
of
http://communication.ucsd.edu/A-News/
This Usenet Oldnews Archive
article may be copied and distributed freely, provided:
1. There is no money collected for the text(s) of the articles.
2. The following notice remains appended to each copy:
The Usenet Oldnews Archive: Compilation Copyright (C) 1981, 1996
Bruce Jones, Henry Spencer, David Wiseman.
