[Xen-devel] rdtsc strangeness on upstream kernel

[Dan Magenheimer <dan.magenheimer@xxxxxxxxxx>]

Hi Jeremy --

In trying to understand some strange behavior I was
seeing on a RHEL6b2 guest, I ran across an interesting
anomaly, and it seems to be true on different upstream
(pvops) kernels.  At first I thought it was a result
of the xen_sched_clock->xen_clocksource_read patch
you recently posted, but after some testing this
appears to be unrelated.

The number of rdtsc/second goes up dramatically when
there is CPU-intensive load on an upstream kernel!
I know we both observed some cases where rdtsc/sec
was very high, but I don't think we ever were able to
reproduce this consistently.

First, this is a single vcpu, 64-bit 2.6.32 (RHEL6b2)
kernel.

I am observing ~300 rdtsc/sec on an idle VM.  When
I run a load of:

main() {while(1);}

I am observing about 10000 rdtsc/sec!!

This is a CONFIG_HZ_1000=y kernel, so I would
expect 1000 rdtsc/sec, or maybe 2000 rdtsc/sec,
but 10000 makes me wonder if there is some hidden
bug.

Changing to vcpus=2, the above numbers roughly
double... ~600 goes up to 20000 rdtsc/sec.
So clearly it is some kind of per-vcpu activity.

Any idea what kernel parts might be doing this?
I can instrument the xen rdtsc code to collect
some more information, but thought I'd get ideas
from you first.

Thanks,
Dan

P.S. Lest anyone panic about this, a rate of
10000 rdtsc/sec/vcpu should translate to ~1%
performance.  Not a catastrophe but still worth
understanding.

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