[dm-devel] Re: clustered snapshots

[Jonathan Brassow]

I've just looked at the last group of patches briefly.  So far, I like  
what I see.  I especially like the concept of locking locally first,  
and then switching to cluster.  If I add cluster lock caching - this  
will further speed things up.

I'd prefer to have the UUID that is passed in on the CTR line be more  
general - leaving the possibility for future features/flags, rather  
than just plopping an extra argument on the end.  I also need to look  
over the 'reread_exceptions' stuff, as this seems a little light...  I  
assume you didn't just take my patch to allow rereads because by  
adding an additional function to the API, you simplify some of the  
logic?

Also, any progress on adding the cluster locking to the your shared  
snapshots?

  brassow


On Sep 29, 2009, at 4:39 PM, Mikulas Patocka wrote:

> Hi
>
> I uploaded a new version there. It has selective re-read and it has
> optimized locking --- it does no remote communication if the chunk  
> already
> exists, so it should be as efficient as Jon's approach.
>
> Jon, please look at it and review the patches, if you find out that  
> this
> approach can't work at all, we can drop it and go with your  
> approach. If
> it is OK, we can consider adopting it, it looks simpler than making
> clustered-exception-store.
>
> Mikulas
>
>
> On Mon, 28 Sep 2009, Mikulas Patocka wrote:
>
>> Hi
>>
>> I uploaded my test clustered snapshots to
>> http://people.redhat.com/mpatocka/patches/kernel/clustered-snapshots-preview/
>>
>> My patches take different approach from Jon's patches. My patches
>> basically replace down_write(&s->lock) and up_write(&s->lock) with
>> clusterized locking.
>>
>> If there are pending exceptions, the cluster lock must be held  
>> while the
>> local lock is unlocked. The cluster lock is droppen when all pending
>> exceptions are reallocated and the local lock is dropped.
>>
>> The patches are based on my & Mike's merging.
>>
>> These patches are less invasive than Jon's, they area small, they  
>> don't
>> change so much logic and most importantly, they leave merging as it  
>> is.
>>
>> Note that it was never tried in a cluster because I don't have a  
>> cluster!!
>> So there may be a silly bug that makes it not work at all. The  
>> purpose of
>> the patches is to show different simpler approach to clustering. You
>> should test it and debug it.
>>
>> TODO:
>> - implement lock caching (Jon's task for his dm-lock module)
>> - once we implement it, we can implement selective re-read --- i.e.  
>> don't
>> reread the exceptions if the cluster lock was not taken by any  
>> other node
>> - in a few cases we could optimize it to use readlock or only local  
>> lock
>> - implemented cluster merging
>>
>> Mikulas
>>