OpenSolaris

The performance of s11 and s10 HVM guests is very slow compared to metal, dom0 
and PV guests.  This is demonstrated through the libmicro benchmark.  Below are
the apprxomate times to run this benchmark on a 4600m2 (all s11 results on
on build 89 non-debug)

	metal: 22 minutes
	dom0:  26 minutes
	domU/s11 PV: 26 minutes
	domU/s11 HVM: 152 minutes
	domU/s10u6 (Ed's ISO): 171 minutes

Attached are the output of each of these runs and a summary comparison of the min, max,
mean, stddev and percent error from each of the above runs.

All guests were run with 2G of memory and a disk slice VBD.
Can we get the same numbers on a b84 dom0
Yes - I'll get that going today.
Analysing the summary shows that the greatest performance discrepencies between bare metal performance and HVM performance (based on an S11 HVM guest) are in the following system tests:

mmap_ru128k
exit_nolibc
exit_10
mmap_ru8k
mmap_rt8k
exec
pthread_32
pthread_128
pthread_512
fork_10
fork_100
fork_1000

The mmap_ru128k test is ~ 80x slower and the fork_1000 test is about 600x slower!

It looks like the common factor to all these tests is that they are manipulating memory and it's slow. Now to find out why.
You are correct - build 84 looks better (still slow but not of the
magnitude of post-build84).

The metal, dom0 and pv results look consistent to those of build 89.
The libmicro hvm run took 51 minutes (vs.152 minutes on 89).  The summary 
report is attached as report_84.
85 is when we went from xen 3.0.x to 3.1.2? I wonder if we are getting bitten by
a changes in the hypervisor's MMU support for guests? Doesn't look like it would
have anything to do with dom0. Jurgen had indicated a problem with solaris guests.
Excessive number of minor faults for us in the fork test. Perhaps you could run
just one of these test that exhibit the worst peformance and see if minor faults
are way off. If so that would indicate where to look in xen. Just a thought.
Todd, the minf numbers don't look to be worse in HVM operation, i.e. we seem to generate approximately the same number of faults from an HVM guest as we do in dom0.

Claudia, could you run the test on an intel box? I'm exploring different areas of HVM support to see where the problem may lie and it would be very helpful to know if this is an AMD specific problem or an issue on intel as well. Just running baremetal and HVM s11 on the intel box should be enough information.

N/A

N/A