Asri-unix.1169
net.works
utzoo!decvax!cca!sri-unix!pratt
Tue Apr  6 20:54:10 1982
Re:  WORKS Digest V2 #39
Reply-to: csd.pratt at SCORE
       Date: 4 April 1982 1011-EST
       From: Hank Walker at CMU-10A
       Subject: performance analysis and remote disks

       Benchmarks written in a high-level language and then compiled tell me
       NOTHING. ...  Assembly-code benchmarks are the
       only relevant ones for comparing architectures.

So who's comparing architectures?  The corresponding statement for automobiles
would be "Automobile road tests tell me nothing.  Bench dynamometer tests are
the only relevant ones for comparing engines."  A car customer wants to know
how the car performs, not just the engine.  Dan Lynch wanted to know the cost
of the move from timesharing to PC's.  Answering him by talking about the
performance of one component of a PC system is only relevant if it helps
predict system performance.  Dan's question is far better addressed directly
using system performance measurements than indirectly (and incompletely) with
data on the performance of just one component.

                 Anyone who thinks that a MC68000 is 70-75% of an EFFECTIVELY
       USED VAX had better think again.

You have some hard data to back this up?  (I don't know what interpretation
of "effectively" you had in mind, but it sounds like you had the rather dated
one of the machine alone executing efficiently, as opposed to the more modern
one of the human-machine combination being used efficiently.  Even so, I'd
like to see the evidence for this case.)

The benchmarks I used (there were five of them) all had the MC68000 crippled by
the C compiler.  If you want to compare the Vax to the MC68000 with the Vax
sped up using assembly code, you should permit a similar speed-up for the
MC68000.  DEC's Doug Clark by the way has extracted from several days worth of
carefully instrumented measurements on the Vax 11/780 the number 0.5 MIPS.
Even allowing for the fact that the Vax does A=B+C in one instruction, I
wouldn't be too surprised to find the MC68000 hot on its heels in assembly
language, though I wouldn't much care since no one around these parts
programs either the Vax or the 68000 in assembly language.

Thanks for the input on the cost of interposing an Ethernet between a disk and
a processor.  Your warning about TCP duly noted, this agrees with
observations of Lampson, Nelson, Popek, and others about the high cost of
general purpose protocols.  However I was startled by "the big remote disk is
undoubtedly faster than a small local Winchester," this seems to be in the
same category as the belief that a large Vax must be faster than a miniscule
MC68000.  The 10" Fujitsu Eagle Winchester transfers data at almost 2
megabytes/second, and more than one 8" Winchester runs at 1 megabyte/sec, all
quite likely to be faster than the big washing machines I bet you have in
your machine room.  Small need not be slow, for either processors or disks.

                                                       Vaughan Pratt

-----------------------------------------------------------------
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.