Will Xen 4.0 release use the pv-ops kernel despite it's performance hit

Will Xen 4.0 release use the pv-ops kernel despite it's performance hit

[Tom Rotenberg]

Hi All,

Will Xen 4.0 release use the pv-ops kernel, despite the fact that it
has about ~15% hit in performance?
This is a major performance hit, and releasing a new Xen version with
a 15% performance hit in the dom0 kernel compared to a previous
version, will just cause people not to upgrade to Xen 4.0 (especially
if Xen 4.0, won't support the old 2.6.18 xen kernel anymore).

Is this performance issue going to be addressed prior to Xen 4.0 release?

Tom

Re: Will Xen 4.0 release use the pv-ops kernel despite it's performance hit

[Keir Fraser]

On 20/12/2009 17:41, "Tom Rotenberg" <tom.rotenberg@gmail.com> wrote:

> Will Xen 4.0 release use the pv-ops kernel, despite the fact that it
> has about ~15% hit in performance?
> This is a major performance hit, and releasing a new Xen version with
> a 15% performance hit in the dom0 kernel compared to a previous
> version, will just cause people not to upgrade to Xen 4.0 (especially
> if Xen 4.0, won't support the old 2.6.18 xen kernel anymore).
> 
> Is this performance issue going to be addressed prior to Xen 4.0 release?

Work on both stabilisation and performance will be ongoing. But bear in mind
that upgrading to Xen 4.0 will not require users to switch to pv_ops dom0.

 -- Keir

Re: Will Xen 4.0 release use the pv-ops kernel despite it's performance hit

[Tom Rotenberg]

OK.

I just thought i saw Jeremy's mail about changing the APIC logic in
Xen to conform to the pv-ops kernel, which i thought will cause new
Xen not to work with old dom0 kernel. This isn't the case?

On Sun, Dec 20, 2009 at 11:53 PM, Keir Fraser <keir.fraser@eu.citrix.com> wrote:
> On 20/12/2009 17:41, "Tom Rotenberg" <tom.rotenberg@gmail.com> wrote:
>
>> Will Xen 4.0 release use the pv-ops kernel, despite the fact that it
>> has about ~15% hit in performance?
>> This is a major performance hit, and releasing a new Xen version with
>> a 15% performance hit in the dom0 kernel compared to a previous
>> version, will just cause people not to upgrade to Xen 4.0 (especially
>> if Xen 4.0, won't support the old 2.6.18 xen kernel anymore).
>>
>> Is this performance issue going to be addressed prior to Xen 4.0 release?
>
> Work on both stabilisation and performance will be ongoing. But bear in mind
> that upgrading to Xen 4.0 will not require users to switch to pv_ops dom0.
>
>  -- Keir
>
>
>

RE: Will Xen 4.0 release use the pv-ops kernel despite it's performance hit

[Zhang, Xiantao]

Tom Rotenberg wrote:
> OK.
> 
> I just thought i saw Jeremy's mail about changing the APIC logic in
> Xen to conform to the pv-ops kernel, which i thought will cause new
> Xen not to work with old dom0 kernel. This isn't the case?

As I know, new Xen should work with old dom0 after APIC logic changes, but new dom0 may not be able to run on old Xen. To make new dom0 work with old Xen hypervisor, has to backport two small patches Cset#20491 and Cset20509.  
Keir, 
   Do you think we should backport them to Xen3.4-testing tree ? Since logic has no any change, and just added a new hypercall, I think they should be good to xen-3.4-testing tree ?  :)
Xiantao

> On Sun, Dec 20, 2009 at 11:53 PM, Keir Fraser
> <keir.fraser@eu.citrix.com> wrote: 
>> On 20/12/2009 17:41, "Tom Rotenberg" <tom.rotenberg@gmail.com> wrote:
>> 
>>> Will Xen 4.0 release use the pv-ops kernel, despite the fact that it
>>> has about ~15% hit in performance?
>>> This is a major performance hit, and releasing a new Xen version
>>> with 
>>> a 15% performance hit in the dom0 kernel compared to a previous
>>> version, will just cause people not to upgrade to Xen 4.0
>>> (especially 
>>> if Xen 4.0, won't support the old 2.6.18 xen kernel anymore).
>>> 
>>> Is this performance issue going to be addressed prior to Xen 4.0
>>> release? 
>> 
>> Work on both stabilisation and performance will be ongoing. But bear
>> in mind that upgrading to Xen 4.0 will not require users to switch
>> to pv_ops dom0. 
>> 
>>  -- Keir
>> 
>> 
>> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xensource.com
> http://lists.xensource.com/xen-devel

Re: Will Xen 4.0 release use the pv-ops kernel despite it's performance hit

[Keir Fraser]

It's the other way round: using newer pv_ops dom0 will require a newer
version of Xen (4.0, or perhaps a newer 3.4.x if I backport the patches).
Compatibility is not broken the other way round (new Xen, old dom0), and
never should be.

 -- Keir

On 21/12/2009 06:16, "Tom Rotenberg" <tom.rotenberg@gmail.com> wrote:

> OK.
> 
> I just thought i saw Jeremy's mail about changing the APIC logic in
> Xen to conform to the pv-ops kernel, which i thought will cause new
> Xen not to work with old dom0 kernel. This isn't the case?
> 
> On Sun, Dec 20, 2009 at 11:53 PM, Keir Fraser <keir.fraser@eu.citrix.com>
> wrote:
>> On 20/12/2009 17:41, "Tom Rotenberg" <tom.rotenberg@gmail.com> wrote:
>> 
>>> Will Xen 4.0 release use the pv-ops kernel, despite the fact that it
>>> has about ~15% hit in performance?
>>> This is a major performance hit, and releasing a new Xen version with
>>> a 15% performance hit in the dom0 kernel compared to a previous
>>> version, will just cause people not to upgrade to Xen 4.0 (especially
>>> if Xen 4.0, won't support the old 2.6.18 xen kernel anymore).
>>> 
>>> Is this performance issue going to be addressed prior to Xen 4.0 release?
>> 
>> Work on both stabilisation and performance will be ongoing. But bear in mind
>> that upgrading to Xen 4.0 will not require users to switch to pv_ops dom0.
>> 
>>  -- Keir
>> 
>> 
>> 

Re: Will Xen 4.0 release use the pv-ops kernel despite it's performance hit

[Keir Fraser]

On 21/12/2009 06:39, "Zhang, Xiantao" <xiantao.zhang@intel.com> wrote:

>> OK.
>> 
>> I just thought i saw Jeremy's mail about changing the APIC logic in
>> Xen to conform to the pv-ops kernel, which i thought will cause new
>> Xen not to work with old dom0 kernel. This isn't the case?
> 
> As I know, new Xen should work with old dom0 after APIC logic changes, but new
> dom0 may not be able to run on old Xen. To make new dom0 work with old Xen
> hypervisor, has to backport two small patches Cset#20491 and Cset20509.
> Keir, 
>    Do you think we should backport them to Xen3.4-testing tree ? Since logic
> has no any change, and just added a new hypercall, I think they should be good
> to xen-3.4-testing tree ?  :)

Yes, go ahead. They do require backporting (and testing) -- they don't apply
as-is.

 -- Keir

RE: Will Xen 4.0 release use the pv-ops kernel despite it's performance hit

[Zhang, Xiantao]

Keir Fraser wrote:
> On 21/12/2009 06:39, "Zhang, Xiantao" <xiantao.zhang@intel.com> wrote:
> 
>>> OK.
>>> 
>>> I just thought i saw Jeremy's mail about changing the APIC logic in
>>> Xen to conform to the pv-ops kernel, which i thought will cause new
>>> Xen not to work with old dom0 kernel. This isn't the case?
>> 
>> As I know, new Xen should work with old dom0 after APIC logic
>> changes, but new dom0 may not be able to run on old Xen. To make new
>> dom0 work with old Xen hypervisor, has to backport two small patches
>> Cset#20491 and Cset20509. 
>> Keir,
>>    Do you think we should backport them to Xen3.4-testing tree ?
>> Since logic has no any change, and just added a new hypercall, I
>> think they should be good to xen-3.4-testing tree ?  :)
> 
> Yes, go ahead. They do require backporting (and testing) -- they
> don't apply as-is.

Okay, I will backport them and send it out later.  
Xiantao

Re: Will Xen 4.0 release use the pv-ops kernel despite it's performance hit

[Konrad Rzeszutek Wilk]

On Sun, Dec 20, 2009 at 07:41:31PM +0200, Tom Rotenberg wrote:
> Hi All,
> 
> Will Xen 4.0 release use the pv-ops kernel, despite the fact that it
> has about ~15% hit in performance?
> This is a major performance hit, and releasing a new Xen version with
> a 15% performance hit in the dom0 kernel compared to a previous

I am quite curious about this. Earlier in November I asked about this and
was told the same issue - but no hard numbers. Recently Jun from Intel mentioned
that they had run performance numbers that they would share in Xen-conference in Asia
(wasn't there, bummer) and which would be posted on xen-devel (Jun: hint hint).

The numbers they listed was ~5% (this is 2.6.31 dom0 compared to 2.6.18 dom0).
Thought I don't remember if baremetal 2.6.31 vs baremetal 2.6.18 has the ~5% degredation
as well?

I don't have the hard numbers, and I haven't come to the stage where I can start
tracking this down, so I was wondering if somebody else in the community had
done this? And perhaps also run the right profiling tool to figure out where the 
hit is happening?

> Is this performance issue going to be addressed prior to Xen 4.0 release?

I know that one of the issues for DomU was resolved a couple of months ago. But
besides that we are working on stabilizing the kernel first.

It goes without saying that any help in the performance department (especially
collecting profile data as this is happening) would be much appreciated.

Re: Will Xen 4.0 release use the pv-ops kernel despite it's performance hit

[Jeremy Fitzhardinge]

On 12/20/2009 10:16 PM, Tom Rotenberg wrote:
> I just thought i saw Jeremy's mail about changing the APIC logic in
> Xen to conform to the pv-ops kernel, which i thought will cause new
> Xen not to work with old dom0 kernel. This isn't the case?
>    

No.  Future pvops dom0 kernels will require Xen >= 4.0.  Xen 4.0 will 
still support older dom0 kernels.

BTW, I haven't measured pvops dom0 performance, but there's no 
particular reason to believe there'll be a big hit (the performance hit 
I was talking about refers to the overhead of fork/exec-heavy workloads, 
which is very unlike dom0's workload).

     J