It's well worth noting that "MIPS" really don't capture the performance
evolution of processors and systems. At best they only provide a vague,
fuzzy sense of what's going on over years and decades.

I'll highlight one part of the problem. Let's suppose you pick some past
system -- a System/360 Model 65 perhaps -- and then try to project that
forward to a zEnterprise EC12 produced a half century later. As mentioned,
you could pick a Drystone or other trivial program, run on it on the
System/360 Model 65, run it on intervening systems, run it on a zEnterprise
EC12, jot down some "MIPS" (or "BIPS") numbers, and try to draw value
assessments from those numbers. You could, and somebody just did.

But what have you just done? You've just artificially limited the newest
systems to a program that is reasonably conveniently executable on the
oldest system. That's certainly fair to the oldest system, but is it fair
to the newest? Probably not.

So, how about taking, say, a mixed workload of 238 batch and online
programs with 5% in 24-bit, 52% in 31-bit, and 43% in 64-bit, in a mix of
COBOL, Java, PL/I, and C with some crypto and XML parsing tossed in,
measure that on the newest processor, then measure that on the older
processors?

Ah, can't be done, right? Wrong. If you have a Turing complete machine --
or close enough, and assuming enough storage, of course -- you can run
anything. So you could run that mixed workload I just described on a 24-bit
System/360 processor. (The System/360 itself used this same concept within
its own line to define a common architecture.) Yes, it would run very, very
slowly, but it could run. Again, assuming a mythical System/360 Model 65
equipped with, say, 16GB of (bank switched or perhaps I/O-attached) core
memory and a z/Architecture emulation program. (Let's ignore disk and other
I/O in this oversimplified example.) Sure, it might take years just to get
one of the JVMs started, but it could run.

Is this second approach fair to the zEC12? Yes. Is it "fair" to the
System/360 Model 65? Maybe, maybe not. But why is relative measurement
approach #1 any better or worse than relative measurement approach #2?

Your choice of workloads for conducting such measurements matters far more
than anything else. If you constrain your universe of workloads to only
those that can "conveniently" run on the oldest system, you're already
putting your thumb on the scale in favor of that oldest system. "MIPS" does
that. Surely it should be obvious that one zEC12 MIPS is on a completely
different planet than one S360-M65 MIPS.

--------------------------------------------------------------------------------------------------------
Timothy Sipples
IT Architect Executive, zEnterprise Industry Solutions, AP/GCG/MEA
--------------------------------------------------------------------------------------------------------

E-Mail: [email protected]
----------------------------------------------------------------------
For IBM-MAIN subscribe / signoff / archive access instructions,
send email to [email protected] with the message: INFO IBM-MAIN

Reply via email to