followup to dumb question (or, VAX11-750 MIPS, but which MIP?)

[Paul De Bra]

In article <6572 at zodiac.UUCP> pkahn at ads.com (Phil Kahn) writes:
>OK.  So the general conceensus is that a VAX11-750 is about .6
>VAXMIPS.  A person noted that VAXMIPS are not equal to "regular" MIPS.
>I was hoping to compare it (generally) to a SUN 4/260 which SUN rates
>at 10 MIPS.  Doesn't that make the 4/260 about 16.7 times faster than
>a 750? I understand the difficulty with these terms, but I want to
>ballpark them right for a paper in order to place some algorithm
>benchmarks we ran into perspective (i.e., when using a "modern" machine). 
>

Let me dug out some test results to show that the comparison is not at
all that easy. The VAX is a CISC machine, meaning it has a large set of
instructions, requiring complex instruction decoding which takes time.
The Sun-4 is a RISC machine, meaning it has a small set of instructions
requiring very little time, but more complex operations require several
instructions...
I ran a number of loops for the well known sieve, taking say 15 million
instructions. The 750 needs 25.0 seconds, which means it runs at 0.6
MIPS. The SUN 4/260 needs 1.4 seconds, which means 10.7 MIPS.
Now, lets run a increment loop, from 0 to 10 million. This is an increment,
test and jump, thus 30 million instructions. The 750 needs 50.1 seconds,
which means 0.6MIPS. The Sun 4/260 needs 1.7 seconds, which means 17.6
MIPS. At very simple tasks the Sun is suddenly faster...
Now, let's run something very complicated: a number of shell scripts
in parallel, all running things like od, grep, wc, etc. The used files
are very small, so there is no need for disk access which would make
the result rather disk-dependent. The 750 needs 113 seconds, and the
Sun 4/260 19 seconds. Still assuming that the 750 runs at 0.6 MIPS
(confirmed by the previous 2 tests) the Sun-4 runs at 3.5 MIPS.

You see that there is NO WAY to say the Sun-4 runs this-many-times
faster than the 750. It depends on what you try to run. Especially in
multitasking situations performance gains obtained by large register sets,
cache memory and other tricks is lost because of context switching.
Note that the loss in performance gain is not specific to the Sun-4
versus the Vax. A Vax 8550 for instance is 11 times faster than the 750
in the sieve, 12 times faster in the loop, but only 4.2 times faster
in the test with the shell scripts in parallel. So in large multiuser
applications faster computers do not boost performance as much as their
"MIPS" value suggests.

Paul.
-- 
------------------------------------------------------
|debra at research.att.com   | uunet!research!debra     |
------------------------------------------------------