> "I'm not convinced it's even valid there, if there is
> any type of virtualization (LPAR or VM) active and
> there are shared resources."
>
> Alan's right - bogomips is a red flag! And the
> assumption that all people take their numbers for VM
> instances in production environments is interesting
> (and knee jerk). I think I should change my handle to
> "Lonely_Lpar_in_a_controlled lab_guy".

Jim,

Read the comment again. "if" signals a conditional expression.  *If* you
have your LPARs defined with shared processors or *IF* you have VM present
and *IF* there happens to be a demand for CPU from another LPAR or virtual
machine that is sharing those processors, THEN you will not get repeatable
results for the bogomips calculation, and the numbers WILL vary widely. *If*
you can afford to dedicate physical processors to single Linux instances,
then you are a) considerably more fortunate than most of us, b) have a very
unusual workload, and/or c) in a lab situation.

You've argued your case for running in LPARs here before -- no need to
rehash it again. Perhaps we don't agree on terminology: I consider CPU
sharing between LPARs a form of virtualization -- note I did not say VM, I
said virtualization -- because in that case, there is not a one-to-one
correspondence between physical hardware and logical hardware at all times.
Do you agree? If not, then that might be the disconnect here.

> So, rather grousing about bogomips, what standard
> measure do you have that can measure the relative
> speed of the processors!

What kind of instruction mix do you want to sample? You can use the
"standard" benchmarks, but they're pretty much just as useless on zSeries as
they are on every other platform.  You could use LSPR numbers, but they're
pretty much measuring z/OS workloads, which don't tell you much of anything
useful about Linux. You could use a set of representative applications
(which is what we do when we do performance comparisons), but without a
large database of performance data on other platforms, nobody agrees with
you or pays any attention to your tests.

The person you want to ask about zSeries processor performance is Bob Rogers
in POK. He designed the bloody things -- ask him what he uses to measure
them. You may also find the IBM Systems Journal issue on the z900 helpful;
there was some detailed discussion of CPU internals in those articles.

-- db

Reply via email to