[Pacemaker] A question and demand to a resource placement strategy function

[Gao,Yan]

On 07/05/11 12:34, Andrew Beekhof wrote:
> On Thu, Jun 2, 2011 at 4:59 PM, Gao,Yan <ygao at novell.com> wrote:
>> On 06/01/11 18:51, Yuusuke IIDA wrote:
>>> Hi, Yan
>>>
>>> An answer becomes slow, and really I'm sorry.
>>>
>>> (2011/05/13 15:06), Gao,Yan wrote:
>>>> I understand that you think the improvement for the non-default
>>>> placement strategy makes sense to the "default" too. Though the
>>>> "default" is somewhat intended not to be affected by any "placement
>>>> strategy" so that the behaviors of existing pengine test cases and
>>>> users' deployments remain unchanged.
>>> I think that a function dispersed with the number of the start of the
>>> resource has a problem at the time of "default" setting.
>>>
>>> This problem is the Pacemaker-1.0 series, but does the same movement.
>>> If it could be settled by this correction, I thought a correction to be
>>> applicable in Pacemaker-1.0.
>>>
>>> Should not this problem be revised?
>> This would affect dozens of existing regression tests, although most of
>> the changes are just the scores of clone instances, which are due to
>> different resource allocating orders. Given 1.0 is in such a maintenance
>> state, I'm not sure we should do that for 1.0.
>>
>> Andrew, what do you think about it? Perhaps we should fix the
>> resource-number-balancing for "default" strategy in 1.1 at least?
> 
> I think for 1.1 we can do something, I'd just like to understand the
> the implications of the patch.
> It would help if there was a testcase that illustrated the negative behaviour.
> 
> Is it necessary that both parts of the old if-block are always run?
Good point. The "sort_rsc_process_order" part makes most sense. The
effect of only running it would be:

- The resources that have higher priorities and have higher scores on
the nodes (comparing the scores in the default order of the nodes) get
assigned first.

That could resolve this issue.


The basic idea of the "sort_node_weight" part is:

- The nodes that are more healthy (have higher node weights) and have
more capacities (if placement-strategy=balanced) get consumed first.

Though I'm surprised that only running the "sort_rsc_process_order" part
would increase 2 affected regression tests compared with running both.

Regards,
  Yan
-- 
Gao,Yan <ygao at suse.com>
Software Engineer
China Server Team, SUSE.

    * English - detected
    * English
    * Chinese (Simplified)

    * English
    * Chinese (Simplified)

 <javascript:void(0);> <#>