[Qemu-devel] QEMU on windows timing problems

[Qemu-devel] QEMU on windows timing problems

[Daniel Ionita]

Hello.

I have a problem with time drifting inside the QEMU machine (running
XP SP2). This is more visible as the load inside the VM increases
(heavy hard-disk activity, almost 100% CPU usage). However, I need to
control some actions inside the VM machine which depend on time. Any
idea how I can increase QEMU's time resolution, to make it more exact?

I need something on the order of seconds. For example, on the latest
tests that I've done, in a total run of 7 minutes QEMU's clock lagged
by 4 minutes (so it's ~50% slower than real time).

Host machine x86, Windows, with kqemu installed. I mention that the
problem is much worse when I run QEMU in a SMP environment, but it is
fixable by setting process affinity to specific CPUs.

Thanks a lot.

-- 
Best regards,
 Daniel                          mailto:dionita@bitdefender.com

RE: [Qemu-devel] QEMU on windows timing problems

[Dor Laor]

>Hello.
>
>I have a problem with time drifting inside the QEMU machine (running
>XP SP2). This is more visible as the load inside the VM increases
>(heavy hard-disk activity, almost 100% CPU usage). However, I need to
>control some actions inside the VM machine which depend on time. Any
>idea how I can increase QEMU's time resolution, to make it more exact?
>
>I need something on the order of seconds. For example, on the latest
>tests that I've done, in a total run of 7 minutes QEMU's clock lagged
>by 4 minutes (so it's ~50% slower than real time).
>
>Host machine x86, Windows, with kqemu installed. I mention that the
>problem is much worse when I run QEMU in a SMP environment, but it is
>fixable by setting process affinity to specific CPUs.
>

I checked the problem with/without KVM and got identical results.
The problem source is that qemu relays on SIGALARM for qemu - timer, not
only the signal does not arrive on time, several signals may unit.
Relaying on PIT source clock makes the problem worse because several PIT
interrupts might got unified too.

Solutions:
- (Works foe second granularity) Use dyn-tick enabled kernel (the latest
i386 kernel or patched 2.6.22 for 64 bit).
- (Simple, not elegant workaround) Measure the PIT frequency and make
sure that X PIT irqs are acknowledged 
- (best solution) Use dyn-tick kernel + implement qemu dyn-tick timer.
HTH, 	Dor.