Exit timing - results online

Exit timing - results online

[Christian Ehrhardt]

Hi,
I just added the following page to our wiki based online documentation 
to show some numbers on kvmppc as of today.
You can find that page at thew wiki entry:
  PowerPC_Exittimings 
<http://kvm.qumranet.com/kvmwiki/PowerPC_Exittimings> - an overview of 
workload dependent overhead on non hardware assisted powerpc virtualization
  (http://kvm.qumranet.com/kvmwiki/PowerPC_Exittimings)

Currently this page covers a brief introduction, workload description, 
exit timings and an example of paravirtualization improvement.

There is no "interpretation" of that data on the page yet.
I think I add some explanations and theories why you can see which 
effect there once I chatted with a few involved people.
I welcome everyone to discuss with us on the list about that too.

-- 

Grüsse / regards, 
Christian Ehrhardt
IBM Linux Technology Center, Open Virtualization

Re: Exit timing - results online

[Hollis Blanchard]

On Thu, 2008-11-06 at 12:25 +0100, Christian Ehrhardt wrote:
> Hi,
> I just added the following page to our wiki based online documentation 
> to show some numbers on kvmppc as of today.
> You can find that page at thew wiki entry:
>   PowerPC_Exittimings 
> <http://kvm.qumranet.com/kvmwiki/PowerPC_Exittimings> - an overview of 
> workload dependent overhead on non hardware assisted powerpc virtualization
>   (http://kvm.qumranet.com/kvmwiki/PowerPC_Exittimings)

Very interesting; thanks for putting these numbers together Christian.

I'm interested by the instruction emulation, which shows up as the
highest cost in every workload. That suggests that even small
improvements to the emulation path, such as using a better lookup than
nested switch statements, could have a significant effect on throughput.

It's also interesting that the average is really close to the minimum,
but the standard deviation is huge. That suggests that the vast majority
of exits take the minimum amount of time, but a few exits take a *ton*
of time. Looking through the set of instructions we emulate on 440, it
looks like the only even slightly sophisticated ones are rfi, tlbwe, and
tlbsx...

-- 
Hollis Blanchard
IBM Linux Technology Center

Re: Exit timing - results online

[Christian Ehrhardt]

Hollis Blanchard wrote:
> On Thu, 2008-11-06 at 12:25 +0100, Christian Ehrhardt wrote:
>   
>> Hi,
>> I just added the following page to our wiki based online documentation 
>> to show some numbers on kvmppc as of today.
>> You can find that page at thew wiki entry:
>>   PowerPC_Exittimings 
>> <http://kvm.qumranet.com/kvmwiki/PowerPC_Exittimings> - an overview of 
>> workload dependent overhead on non hardware assisted powerpc virtualization
>>   (http://kvm.qumranet.com/kvmwiki/PowerPC_Exittimings)
>>     
>
> Very interesting; thanks for putting these numbers together Christian.
>
> I'm interested by the instruction emulation, which shows up as the
> highest cost in every workload. That suggests that even small
> improvements to the emulation path, such as using a better lookup than
> nested switch statements, could have a significant effect on throughput.
>
> It's also interesting that the average is really close to the minimum,
> but the standard deviation is huge. That suggests that the vast majority
> of exits take the minimum amount of time, but a few exits take a *ton*
> of time. Looking through the set of instructions we emulate on 440, it
> looks like the only even slightly sophisticated ones are rfi, tlbwe, and
> tlbsx...
>
>   
Yeah I extended my patch a bit to now account several different 
subcategories of the emulation.
One issue was that a idling guest setting MSR[WE] came via the emulation 
path and recorded all the time a kvm_block waited until there was work 
for the cpu.

I updated the numbers on the wiki page on Friday.

-- 

Grüsse / regards, 
Christian Ehrhardt
IBM Linux Technology Center, Open Virtualization