Re: [PATCH 00/10] in-kernel APIC v3 (kernel side)

["Gregory Haskins" <ghaskins-Et1tbQHTxzrQT0dZR+AlfA@public.gmane.org>]

>>> On Sun, May 13, 2007 at 10:06 AM, in message
<64F9B87B6B770947A9F8391472E032160BC74642-yEcIvxbTEBqsx+V+t5oei8rau4O3wl8o3fe8/T/H7NteoWH0uzbU5w@public.gmane.org>,
"Dor Laor" <dor.laor-atKUWr5tajBWk0Htik3J/w@public.gmane.org> wrote: 
>> >
>>> Superb results, when I run the old in-  kernel apic I got much higher
> idle
>>> cpu consumtion. I know that's vmenter-  vmexit latency was drastically
>>> improved but 0-  1% cpu consumption for winxp 32 bit sounds too good
> to be
>>> true.
>>> Are you sure an ACPI HAL is used in the guest? (can be checked by
>>> looking at the computer node in the device manager).
>>
>>I did a clean install and didn't change it over to StandardPC.  The
> Control
>>Panel says "ACPI Uniprocessor PC" or something like that (recalling
> from
>>memory right now).  As a third datapoint, I am seeing the TPR changing
>>rapidly (which I would expect 0 changes if the ACPI stuff was
> disabled).
> 
> Great, I was a little skeptical (sorry for that) 

No problem.

> since several months
> ago when I tested the in- kernel- apic the idle time was much worse. After
> talking today with Avi, he said his vmentry- exit latency drastically
> improved performance and even the qemu's apic can reach these figures.

Yeah.  At first I was psyched when I saw my code running at 0-1%, but then I ran trunk and saw its not all that different :(  The good news is that windows is performing pretty well now, regardless of my patch or not.  Thats good for KVM all around.

> I'm sure that with load, you implementation will achieve performance
> advantage.

In theory, we should be able to take advantage of all the light-exit gains we have made recently since that is the primary difference for windows+ACPI now.  This means the APIC change will have an increasing impact as we improve the VMEXIT times (both in HW and SW).  So even if it doesn't make a huge diff today, it will always be contributing more and more as we push the ball forward in other areas.

But I am not disappointed.  My primary motivation is functionality, not performance. ;)  I need to be able to inject interrupts to a guest from the kernel for the work I am doing.  If I also improve performance, great!



-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/