(sorry al= l for the top post... it’s hard to reply to HTML properly in my maile= r)

Hi Ben –=

On recent Linux kernels, setting “clocksour= ce=3Dtsc tsc=3Dreliable” as guest kernel parameters should be faster = than pvclock because pvclock reads the tsc to drive its algorithm.

CMOS timer is very very slow.&= nbsp; If you are not familiar with virtual time issues, you should read the= VMware paper. Xen virtual time management is a little different (bec= ause VMware can do binary translation) but most of the concepts are the sam= e and this paper also is a very good explanation of all the different clock= /timer hardware:

&nbs= p;

http://www.vmware.com/files/pdf= /Timekeeping-In-VirtualMachines.pdf

From: Bei Guan [mailto:gbtju85@gmail.com]
Sent: We= dnesday, March 28, 2012 7:45 PM
To: Dan Magenheimer
Cc:= Xen Devel; Andrei E. Warkentin
Subject: Re: [Xen-devel] Question= s on the PIT and APIC timer in Xen

2012/3/28 Dan Ma= genheimer <dan.magenheimer= @oracle.com>

Both the PIT and = APIC (as well as HPET) timers are emulated by Xen for an HVM guest. H= owever there have always been rare hard-to-reproduce problems when guests u= se them. At one time, there were workarounds that required a special = parameter “timer_mode” to be selected for HVM guests. If = I recall correctly, different timer modes overcame different limitations of= different timer emulations (or maybe mismatches between guest “bugs&= #8221; exposed by virtual time and true hypervisor emulation). These = timer_mode workarounds may or may not work with more recent Linux and Windo= ws guests. And there are still periodic bug fixes being applied for t= imers in the hypervisor emulation code.

At 4.0, full TSC emulation was ad= ded in an attempt to provide a high-resolution guaranteed monotonic timer. = However, most legacy HVM guests cannot or refuse to use TSC as = a timer. (And there may be lurking bugs there as well.)

For most PV guest OS’s, the pvclock algorit= hm is used so PIT/APIC/HPET emulation should be unnecessary (and, if used, = is likely buggy). And actually pvclock should be superceded by using = TSC, which is always faster, but for most workloads the performance differe= nce doesn’t matter.

If you are interested in a good (but difficul= t) Xen project, “virtual time” still needs work and documentati= on. I would be happy to provide lots of background documentation but = can’t volunteer to help much more than that right now.

Hi Dan,

Thank you for your detailed reply and your kindness.=

So, you mean the pvclock algorithm is v= ery faster than any Xen emulated timer, while a TSC timer is very faster th= an the pvclock for some workloads, right?

Is it possible that Xen initial its timer from the hardware no= w, such as CMOS timer? Why PV guest OS doesn't use this timer instead of us= ing the timer from the pv clock algorithm? If there is any problem with my = question, please correct me.

Thank you very m= uch.

Bei

<= /o:p>

Hope that helps,<= /span>
Dan
&= nbsp;
From: Bei Guan [mailt= o:gbtju85@gmail.com<= /a>]
Sent: Wednesday, March 28, 2012 1:32 AM
To: Xen D= evel
Cc: Andrei E. Warkentin
Subject: [Xen-devel] Quest= ions on the PIT and APIC timer in Xen

Hi,
=

I wonder whether the PIT= /APIC timer for HVM Guest OS is emulated by Xen or not.
What's difference between setting the boot PIT/APIC timer for = a PV Guest OS and HVM Guest OS?
&nbs= p;
I would appreciate any reply for my ques= tion.

-= -
Best Regards,
Bei Guan
=

--
Best Regard= s,

Bei Guan