I once heard a story that cmos machines are always
in lpar mode, even if one specifies lpar=no on the
iocds. I do not know if it is true, most people
run lpar anyway, so the sie part would always be an
issue, however I guess that lpar uses more sie assist
type features then vm would. Again all model or even
microcode patchlevel dependent.
Benchmarking on a specific instruction mix is the way
to go, but harder to optimize on.
I guess the level of sie would also impact on the
efficiency of the tlb etc.  I remember looking at aix/esa
using vector processors under VM v=f vs v=v, which made
a hughe difference in performance.

Jan Jaeger.

Rob van der Heij wrote:
>
> > Also, I think there is a misunderstanding about the performance
> > counters: they (the ones I ment) are CPU architecture specific and
> > neither gcc nor gprof knows about them. They are implemented as "extra"
>
> fwiw: I never have seen anything like that published (or even mentioned).
> Things like the number of clock cycles to execute a specific sequence of
> instructions have never been published afaik. The RPP ratings of a model are
> determined by running well-defined instruction mix benchmarks (and these
> probably are aimed more towards running z/OS than Linux). Most likely
> processor design was done to favor common instruction sequences (e.g.
> an XC instruction with memory operands may not really fetch the operand
> if both addresses are the same).
> I would not be surprised if such low level instrumentation were only present
> in the hardware simulators they use for processor design.
>
> The particular model number is not shown by Q CPUID. The processor type
> for all S/390 (9672 in your case) and Freeway machines (2064) is the same.
> Changes in clock speed, cache size, look aside tables, extra instruction sets
> etc tend to mark 'generations' (like G5 vs G6).
>
> And since you run in a virtual machine you probably care about CP overhead.
> A virtual machine runs at native speed under control of SIE until it touches any
> of the architecture items marked for interception (at that point the control
> program will further complete the virtualisation or deal with the intercept 
>otherwise).
> The smaller the number of SIE intercepts, the closer the virtual machine gets to
> native speed.
>
> Rob

Reply via email to