[Xen-ia64-devel] Xen/ia64 (paravirtualized dom0) overhead now at 1.7%!

["Magenheimer, Dan (HP Labs Fort Collins)" <dan.magenheimer@xxxxxx>]

Sorry again for all the email... I'm just catching up on
some things after being very focused on Xen/ia64 delivery
for the last few weeks.

I have been tracking Xen/ia64 performance since early this
year by using a "benchmark" of compiling linux-2.6.9 on dom0
(total time to build a previously built linux after a "make clean").
In January, with a privified guest and no tuning at all,
the overhead was over 130%, meaning running the benchmark on
dom0 on Xen was more than twice as slow as running it native
on the same version of linux.  By March, this was down to
about 30%.  With a paravirtualized guest, this was brought
down to about 10% in May and after the development of some
"fast paths", the overhead was down to about 4.6% in July
when Haavard Bjerke published performance numbers in his
thesis:

http://openlab-mu-internal.web.cern.ch/openlab-mu-internal/Documents/Rep
orts/Technical/Thesis_HarvardBjerke.pdf

(Xen/ia64 newbies... the above is recommended reading though
some of it is getting outdated.)

A few days ago, I checked in some more "hyperprivop" and
"hyperreflect" code, but ifdef'd default off to avoid conflicting
with the multi-domain effort.  I've been working on this
code off and on since July and had finally got it working.
With FAST_ITC, FAST_TLB_MISS_REFLECT, and FAST_TICK turned
on, I was able to reduce overhead* of Xen on the "build
linux-2.6.9" benchmark to about 2.6%.

Observing that Xen seemed to be spending a lot of time in idle,
I increased dom0 memory from 512MB to 768MB (vs. 2GB for native
Linux).  With this change, the overhead* of Xen on the benchmark
was reduced to 1.7%!  With more memory, this may go a bit lower
but bugs/constraints currently limit dom0 memory.

Some of the new code will require some additional work to support
domU... right now if the fast path sees current!=dom0, it
punts to the slow path.  Even with the fast paths fully implemented
for domU, domU performance will be slower -- perhaps significantly --
due to virtual I/O, metaphysical memory, and other factors.  I will
publish more results on domU when it becomes more stable.

Dan


_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel