Re: [Qemu-devel] [RESEND][PATCH 0/3] Fix guest time drift under heavy load.

[Gleb Natapov]

On Thu, Nov 06, 2008 at 08:40:09AM -0600, Anthony Liguori wrote:
> Paul Brook wrote:
>>> You see time drift with 100hz timers in the guest??  That makes very
>>> little sense as even without hr timers, the host should have no problem
>>> delivering a 10ms timer.
>>>     
>>
>> If your host has HZ=100 and no HR timers than a 10ms interval is fairly 
>> borderline, and it's not that hard to end up missing an interrupt, 
>> especially under heavy load.
>>   
>
> If the host is otherwise idle, then the guest should have ample time to  
> run.  With /dev/rtc, you should be able to get a reliably 1ms timer.
And only one guest.

BTW I am saying that guest timer is 100HZ since I assume that windows2003
behaves like XP in this regard, but may be default timer frequency of 2003
is different. I'll check this.

>
> Gleb: are you perhaps using a qcow2 file in conjunction with -snapshot?   
I am using qcow2, but without -snapshot.

> If that's the case, and you're doing an IO intensive workload, then you  
> would be expanding the qcow2 file and updating the metadata is  
> synchronous.  This could cause lost time for the guest.
>
I think that's what happens when I see a huge drift, yes.

> If you are using qcow2, can you still reproduce with a raw file and not  
> using -snapshot?
I'll try next week to run with a raw file.

>
>> If part of qemu gets swapped out then all bets are off, and you can 
>> easily stall for significant fractions of a second. No amount of host 
>> high resolution time support will help you there.
>>   
>
> Running a steady workload, you aren't going to be partially swapped.
>
We want to oversubscribe host as much as possible, and workload will
vary during a lifetime of the VMs.

--
                        Gleb.