Re: [Xen-devel] [PATCH 0/7] oxenstored: improve transaction conflict handling

[Wei Liu <wei.liu2@xxxxxxxxxx>]

On Fri, Mar 18, 2016 at 10:33:35AM -0400, Konrad Rzeszutek Wilk wrote:
> On Thu, Mar 17, 2016 at 05:51:08PM +0000, Jonathan Davies wrote:
> > This patch series makes a substantial improvement to oxenstored's 
> > transaction
> > handling.
> > 
> > The original design of oxenstored assumed that a transaction would only ever
> > span a small subtree. In practice this is rarely the case and means that 
> > other
> > xenstore traffic overlapping with a transaction can cause it to fail with
> > EAGAIN. This leads to significant performance problems with toolstack 
> > operations
> > when there is a certain level of xenstore traffic coming from other sources.
> > 
> > Instead, we observe that, although the transaction may span a large 
> > subtree, it
> > is very unlikely to read or write the same keys as other xenstore traffic. 
> > This
> > observation leads to an improved transaction conflict-handling algorithm in
> > which the transaction is replayed to check it is serializable. This 
> > approach is
> > superior in performance.
> > 
> > The sixth patch is the main one that changes the way transaction conflicts 
> > are
> > handled. The first five patches prepare the way by performing some 
> > refactoring
> > and addition of infrastructure that allows for transactions to be replayed. 
> > The
> > seventh patch adds some extra logging.
> 
> All the patches have quite the Reviewed-by list already so I would think
> these can go in now?
> 

We need an ack from Dave.

But, FWIW, the result is impressive, and the changes are only internal
to oxenstored, so I think it's pretty safe interface-wise or
protocol-wise.

Wei.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel