Mon, 12 May 2008 09:10:24 +0800, "Zhang, Jingke" wrote:
>Masaki Kanno wrote:
>> Hi Ian,
>>
>> Fri, 9 May 2008 13:07:31 +0100, "Ian Pratt" wrote:
>>
>>>>> Hi Kanno,
>>>>> We have tried cpus="^1" in the past. For we want to drop only
>>>>> CPU1 no matter how many CPUs are on the machine. The result shows:
>>>>> CPU1 can still be used by HVM and all the VCPU's affinity are "any
>>>>> cpu". So, do you think this setting is available? Or do we need
>>>>> some more comments on this? Thanks!
>>>>
>>>> Hi Zhang,
>>>>
>>>> I think that a purpose to set "cpus" is to confine CPUs which VCPUs
>>>> run to a CPU or some CPUs. If there are many CPUs on a machine, I
>>>> think that we should avoid the setting for performance.
>>>
>>> The "^1" syntax is quite useful if you're trying to reserve a CPU for
>>> use by another domain e.g. domain0, but don't want to otherwise
>>> restrict the guest.
>>
>> "cpus" is parsed from left side, and the "^1" negates the values which
>> was already parsed. So "0-3,^1,1" and "0-3" are equivalent.
>> The parsing of "cpus" cannot handle "negation" from a beginning
>> because there is not a value to negate.
>>
>> But, I'm thinking that I want to implement Zhang's demand since I get
>> your comment. So I have idea as follows.
>>
>> cpus = "0:,^1"
>>
>> The "0:" means "0 or later". How about it?
>>
>
>Hi Kanno,
> I think a "non-boundary" range may be needed to define the
>expression. How about this:
> "x-", means from CPUx to the last CPU. For example, "0-" means 0 or
>later.
> "-x", means from CPU0 to CPUx, for example, "-3" means "0,1,2,3".
> Thank you!
Hi Zhang,
Sounds good. I will make a patch with your idea.
Best regards,
Kan
>>> We need to be able to deal with setting the affinity mask for a
>>> domain (that is replicated to all VCPUs) as for individual VCPUs. I
>>> assume your patch doesn't change the behaviour of the former?
>>
>> Could you look at the following changest?
>>
>> http://xenbits.xensource.com/xen-unstable.hg/rev/5c3df1bded82
>>
>>
>>> BTW, it would also be nice to be able to specify CPUs by
>>> node.socket.core.thread as opposed to just enumerated CPU number. It
>>> should be possible to omit unused levels of the hierarchy, e.g. "0.3"
>>> could be used to refer to the 4th core of the first socket on a dual
>>> socket quad core machine.
>>>
>>> If you're looking for further improvements in this area, enable CPU
>>> groups to be defined and then allow domains/VCPUs to be assigned to a
>>> group.
>>
>> I got interest for your suggestion. Let me think for the time being.
>>
>>
>> Best regards,
>> Kan
>
>
>
>Thanks,
>Zhang Jingke
>
>_______________________________________________
>Xen-devel mailing list
>Xen-devel@xxxxxxxxxxxxxxxxxxx
>http://lists.xensource.com/xen-devel
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel
|
Home