[Xen-devel] Re: performance regression from c/s 21647:cfba1560054a

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 10.11.11 at 14:45, Gianluca Guida <glguida@xxxxxxxxx> wrote:
> On a personal note, thanks for asking me about shadow pagetables, I've
> been missing this kind of fun and it's a good way to celebrate my
> return to Xen (and Citrix) after exactly two years... :-)
> Stefano imported an old patch of mine, so I am the culprit for this.

So welcome back, Gianluca! And thanks for the fast response.

> On Thu, Nov 10, 2011 at 3:23 AM, Jan Beulich <JBeulich@xxxxxxxx> wrote:
>> in this c/s you did, besides the main purpose of the change, an
>> adjustment to check_for_early_unshadow() in that L2...L4 entries
>> would no longer be attempted to get unshadowed. Neither the
>> patch description nor the added comment really make clear why
>> this was done, and we now got a customer report regarding this
>> causing considerable slowdown in process creation/destruction
>> intensive workloads (e.g. shell scripts).
> 
> Unfortunately, I can't clearly remember the reason for this. I assume
> that your mentioning of shell scripts implies that this is running
> under a Linux guest. Back in the day, this patch was tailored very
> strictly against Windows guests, and that probably meant that doing so
> had a performance improvement over the way Windows handled pagetables
> and process distructions.
> 
> Perhaps (I need some time to catch up) the PV-HVM thing has made this
> patch affect Linux as well, at the point that this has became a
> problem.
> 
> Can you tell me in what guest OS this is experiencing a slow down?

It's SLE11 SP1 guests that suffered a regression after a maintenance
update (originally shipped with 4.0.0, while that patch got backported
later into 4.0.x).

> What is the order of magnitude of the performance decrease?

Factor 7...8 in their measurements.

> I have no problem in reverting that part of the change, if it makes you 
> happy.

It's not so much about making me feel good, but about understanding
what the reasons for the change were (and hence whether perhaps
some heuristic would better be used to determine whether higher
level page tables are worth getting unshadowed earlier). After all I
also don't want reverting this to result in a performance regression in
Windows guests.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel