[Sheepdog] [PATCH] sheep: disable object cache by default

[Liu Yuan]

On 05/11/2012 04:06 AM, MORITA Kazutaka wrote:

> At Fri, 11 May 2012 00:31:49 +0800,
> Liu Yuan wrote:
>>
>> On 05/10/2012 10:37 PM, Christoph Hellwig wrote:
>>
>>>> What did you mean by en/disable it on a per VDI basis? I think object
>>>>> cache does have this attribute, in terms of QEMU command line option. Or
>>>>> I miss some context?
>>> Have an attribute on the sheepdog_inode that the object cache is used
>>> for access to this VDI, without having to coordinate all possible users
>>> of it, e.g. a way to specify it when creating the vdi and having any
>>> user simpliy do the right thing.
>>
>> I don't think it make any sense. This just complicate the things, for
>> now we have a global switch and per VDI switch already, any attempt to
>> introduce yet another switch looks really crazy to me. Maybe it would
> 
> If we allow multiple cache features in future, I think it makes sense
> to have such attribute because we cannot specify the detailed cache
> control in the qemu command line options.  It enables us to provide
> different cache levels according to the customer demand.
> 


I am not against introducing another cache mechanism if considered
necessary. Let's find the best default option when other cache feature
is actually implemented.

Thanks,
Yuan

>> help your use case, but I think you should be aware:
>>
>> 1) Sheepdog originally targets for mixing computing and storage in one
>> cluster to maximize the resource usage.
>> 2) Sheepdog isn't inherently optimized for users who has extremely
>> high-end and fast network.
> 
> Yes, Sheepdog aimed at that.  However, currently there are already
> some users (one of them belongs to our NTT group) who want to use
> Sheepdog without mixing computing and storage resources or try
> Sheepdog with non-commodity hardware like Fusion-io or Infiniband.
> As a sheepdog maintainer, I'd like to accept these requirements if it
> doesn't bring much complexity to Sheepdog design and implementation.
> 
> Anyway, it's okay to me to enable the object cache by default if it
> fits for 1) and 2) (I think we should document this restriction more
> clearly for users not to enable it wrongly), but want to see other
> cache features which fit for other requirements, too.
>