Re: /proc/xen/xenbus supports watch?

[Rusty Russell <rusty@rustcorp.com.au>]

On Fri, 2005-09-23 at 10:24 +0100, Keir Fraser wrote:
> Or is the bundling simply so that you can continue to feed consistent 
> snapshotted values to the transaction from its 'shadow store', even 
> though it is ultimately a doomed transaction?

Yes that was the plan, although see other mail, I now prefer not
allowing save during transactions at all to avoid this issue.

> Do xenstored's performance problems stem from copying the store for 
> every transaction?

Yes.  Using TDB means it's now copying a single file, rather than a
whole directory tree, but if someone puts in enough domains in the store
it will become a problem and we'll need a more sophisticated
implementation.  It would be a fun problem to work on, but I don't think
I can justify spending cycles on it yet: I can create around 10,000
domain-style directories in the store on my laptop in 22 seconds (using
10,000 separate transactions).

The advantage of the "whole copy" approach are simplicity and
robustness: if you SIGKILL xenstored your chances of recovery are
excellent, as commit is done as a rename(2).

Cheers!
Rusty.
-- 
A bad analogy is like a leaky screwdriver -- Richard Braakman