VCPU cores & sockets

VCPU cores & sockets

[Ben Guthro]

Before I get to entrenched in the ACPI code - I thought I'd ping the 
community to make sure nobody else has investigated this already -

Can we present cores vs sockets for guest VCPUs?
Right now, the number of total processors can be specified, but can we 
specify cores per socket?

I see this potentially being an hvmloader function, as part of its acpi 
build code.
Could this be accomplished more generally?

Ben

Re: VCPU cores & sockets

[Jean Guyader]

Ben Guthro wrote:
> Before I get to entrenched in the ACPI code - I thought I'd ping the 
> community to make sure nobody else has investigated this already -
> 
> Can we present cores vs sockets for guest VCPUs?
> Right now, the number of total processors can be specified, but can we 
> specify cores per socket?
> 
> I see this potentially being an hvmloader function, as part of its acpi 
> build code.
> Could this be accomplished more generally?
> 

Ben,

This already exists, It's supported via a CPUID configuration.
Right now there are just some comments in the xmexample.hvm file.

I think it would be better to have an entry in the config file like you 
suggest. I will work on that.

Cheers,

-- 
Jean Guyader

Re: VCPU cores & sockets

[Ben Guthro]

so...if I read your patch, and the intel manual properly -
CPUID.4:EAX[31:26] + 1 is the number of cores...

...
#           '4:eax=0x3,ebx=0x0,ecx=0x0,edx=0x0',
#  '0x80000000:eax=0x3,ebx=0x0,ecx=0x0,edx=0x0']
...

This would set the EAX of leaf 4 to 3 - indicating a dual core processor?

Am I understanding that config file correctly?

Jean Guyader wrote on 06/17/2008 10:44 AM:
> Ben Guthro wrote:
>> Before I get to entrenched in the ACPI code - I thought I'd ping the 
>> community to make sure nobody else has investigated this already -
>>
>> Can we present cores vs sockets for guest VCPUs?
>> Right now, the number of total processors can be specified, but can 
>> we specify cores per socket?
>>
>> I see this potentially being an hvmloader function, as part of its 
>> acpi build code.
>> Could this be accomplished more generally?
>>
>
> Ben,
>
> This already exists, It's supported via a CPUID configuration.
> Right now there are just some comments in the xmexample.hvm file.
>
> I think it would be better to have an entry in the config file like 
> you suggest. I will work on that.
>
> Cheers,
>

Re: VCPU cores & sockets

[Keir Fraser]

On 17/6/08 15:44, "Jean Guyader" <jean.guyader@eu.citrix.com> wrote:

>> Can we present cores vs sockets for guest VCPUs?
>> Right now, the number of total processors can be specified, but can we
>> specify cores per socket?
>> 
>> I see this potentially being an hvmloader function, as part of its acpi
>> build code.
>> Could this be accomplished more generally?
> 
> This already exists, It's supported via a CPUID configuration.
> Right now there are just some comments in the xmexample.hvm file.

Yes, and I don't think CPU topology needs to have any effect on the ACPI
tables.

> I think it would be better to have an entry in the config file like you
> suggest. I will work on that.

I argue against further redundant domain config options.

 -- Keir

Re: VCPU cores & sockets

[Ben Guthro]

> I argue against further redundant domain config options.
>
>  -- Keir
>   
While I would normally agree with this statement...
knowing that you must set CPUID.4:EAX[31:26] to cores + 1 (but only if 
CPUID.1:EDX[28]==1) is not exactly user friendly...

Re: VCPU cores & sockets

[Jean Guyader]

Ben Guthro wrote:
> so...if I read your patch, and the intel manual properly -
> CPUID.4:EAX[31:26] + 1 is the number of cores...
> 
> ...
> #           '4:eax=0x3,ebx=0x0,ecx=0x0,edx=0x0',
> #  '0x80000000:eax=0x3,ebx=0x0,ecx=0x0,edx=0x0']
> ...
> 
> This would set the EAX of leaf 4 to 3 - indicating a dual core processor?
> 
> Am I understanding that config file correctly?
> 

My previous patch has nothing to do with that.
What you are looking for is just above in this file.

-- 
Jean Guyader

Re: VCPU cores & sockets

[Keir Fraser]

On 17/6/08 16:56, "Ben Guthro" <bguthro@virtualiron.com> wrote:

>> I argue against further redundant domain config options.
>> 
>>  -- Keir
>>   
> While I would normally agree with this statement...
> knowing that you must set CPUID.4:EAX[31:26] to cores + 1 (but only if
> CPUID.1:EDX[28]==1) is not exactly user friendly...

I might consider having 'xm create' or the domain-config-cooking bit of xend
grind a new cores_per_socket config option into the canonical cpuid option.
I don't think it should be part of the persistent vm configuration state
that xend keeps in xenstore and saves/restores. And this could be done
equally well in higher-level tools.

 -- Keir

Re: VCPU cores & sockets

[Samuel Thibault]

Ben Guthro, le Tue 17 Jun 2008 11:56:30 -0400, a écrit :
> 
> >I argue against further redundant domain config options.
> >
> > -- Keir
> >  
> While I would normally agree with this statement...
> knowing that you must set CPUID.4:EAX[31:26] to cores + 1 (but only if 
> CPUID.1:EDX[28]==1) is not exactly user friendly...

Maybe we could add a little tool that computes the CPUIDs you want?

Samuel

Re: VCPU cores & sockets

[Ben Guthro]

my mistake - I was looking at a rather old version of the file

Jean Guyader wrote on 06/17/2008 11:57 AM:
> Ben Guthro wrote:
>> so...if I read your patch, and the intel manual properly -
>> CPUID.4:EAX[31:26] + 1 is the number of cores...
>>
>> ...
>> #           '4:eax=0x3,ebx=0x0,ecx=0x0,edx=0x0',
>> #  '0x80000000:eax=0x3,ebx=0x0,ecx=0x0,edx=0x0']
>> ...
>>
>> This would set the EAX of leaf 4 to 3 - indicating a dual core 
>> processor?
>>
>> Am I understanding that config file correctly?
>>
>
> My previous patch has nothing to do with that.
> What you are looking for is just above in this file.
>