RE: [Xen-ia64-devel] Transparent paravirtualization on Xen/ia64
> > What are you planning to benchmark? I would be very
> > interested in comparative results of the same
> > benchmark on:
> > 1) Linux 2.6.11
> > 2) Transparently paravirtualized xenlinux running as linux
> > 3) Transparently paravirtualized xenlinux running on Xen
> Sure, I'll do all of them. What about privified (not
> paravirtualized) Linux on Xen? Do you plan to continue
> support for that in the future?
Privified is an interesting datapoint though I don't expect
it to be used much in the future, other than for debugging.
(For example, yesterday I created a privified 2.6.12 to
ensure 2.6.12 would boot before I got the paravirtualized
working.) If you have time, its worth doing.
> Are there any particular aspects of performance you are
> interested in? Do you have any tips on how to isolate them? I
> was planning to just compile a Linux kernel and measure the
> wall time, but that doesn't say very much about where the
> hotspots are..
The privcnt program I posted some time ago is very useful.
I've been using it every day to get detailed statistics
on what's actually happening in Xen and using that to
guide my tuning efforts (also using "compile Linux" as
Wall time can be very misleading for benchmarking as
there can be other things going on in the background.
Make sure you minimize background activity (boot single
user mode?). Also, I suggest taking five readings (of
each measurement), throwing away the smallest and largest
and averaging the remaining three. Remember that the
first run is likely putting a lot of files into buffer
cache. And (important!) real Linux has all the memory
in the system and domain0 is currently restricted to
512MB (I think) so if all file reads are coming out
of the buffer cache in Linux but coming off of the
disk in domain0, that's not really a fair comparison.
(For compiling Linux, it should all fit in buffer cache
but if you choose a different benchmark, watch out for
this and at least vary domain0's memory size.)
The difference between 1) and 2) above is likely to be
exceedingly small, perhaps not within the error bar
of the measurement. If this proves false (reproducibly),
I have more work to do on the transparency!
Xen-ia64-devel mailing list