Keir Fraser wrote:
> On 26/3/08 10:00, "Keir Fraser" <keir.fraser@xxxxxxxxxxxxx> wrote:
>
>>> Keir, we checked guest installation with rhel4u3 today, we compared
>>> c/s 17284 with c/s 16720, The result shows latest c/s with mmio
>>> emulation changes is a little bit faster than before on our test
>>> system with Xeon(r) processors, about 20 seconds faster.
>>
>> That's pretty surprising! I found out that slowdown on my P4 system
>> for WinXP installation is about 15%, so not as bad as I thought. And
>> I can probably reclaim most of that performance loss.
>>
>> I find it hard to explain a performance *win* though!
>
> Well, I implemented a virtual-address to mmio-physical-address
> lookaside cache for x86_emulate(), and with that I get following
> results for install of WinXP (time is up to second reboot, after
> graphical part of install, from an auto-install CD image):
> xen 3.2: 1 hour 20 minutes 23 seconds
> xen unstable using x86_emulate(): 1 hour 33 minutes 4 seconds
> xen unstable with new optimisation: 1 hour 12 minutes 57 seconds
>
> Considering first result (Xen 3.2) as a baseline control experiment,
> basic x86_emulate() mmio performance is 16% slower, while with the
> simple extra optimisation I get a 10% speedup (so that's 22% speedup
> compared without the optimisation).
>
> Pretty nice!
>
> -- Keir
Pretty good enhancement.
Seems on your P4 system, WinXP installation can well expose this
performance issue :).
In our environment, install rhel4u3 with full packages of editors,
test-internet, authoring-and-publishing, development-tools, admin-tools
and system-tools.
c/s 17284 : 422s
c/s 16720 : 438s
-- haicheng
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel
|
Home
xen.org, 