[sheepdog] [PATCH v2] sheep/recovery: multi-threading recovery process

[Liu Yuan]

On Wed, Jan 29, 2014 at 04:28:35PM +0900, Hitoshi Mitake wrote:
> At Tue, 28 Jan 2014 18:01:42 +0800,
> Liu Yuan wrote:
> > 
> > Rationale for multi-threaded recovery:
> > 
> > 1. If one node is added, we find that all the VMs on other nodes will get
> >    noticeably affected until 50% data is transferred to the new node.
> > 
> > 2. For node failure, we might not have problems of running VM but the
> >    recovery process boost will benefit IO operation of VM with less
> >    chances to be blocked for write and also improve reliability.
> > 
> > 3. For disk failure in node, this is similar to adding a node. All
> >    the data on the broken disk will be recovered on other disks in
> >    this node. Speedy recoery not only improve data reliability but
> >    also cause less writing blocking on the lost data.
> > 
> > Our oid scheduling algorithm is intact and simply add multi-threading onto top
> > of current recovery algorithm with minimal changes.
> > 
> > - we still have ->oids array to denote oids to be recovered
> > - we start up 2 * nr_disks threads for recovery
> > - the tricky part is that we need to wait all the running threads to
> >   completion before start next recovery events for multiple nodes/disks events
> > 
> > This patch passes "./check -g md -md" on my local box
> 
> On my box, at least 32 and 33 failed. I'm seeking the root cause now
> but this patch seems to be a little bit dangerous.
> 

Yes, this shouldn't go to stable-0.8, but master is okay and at least we need to
pass all the tests before it can goto master.

Thanks
Yuan