Re: [Nbd] deadlock in nbd?

["Mike Snitzer" <snitzer@...17...>]

On 10/22/07, Eric Gerlach <egerlach@...135...> wrote:
> Mike Snitzer wrote:
> > On 10/22/07, Eric Gerlach <egerlach@...135...> wrote:
> >> Peter Daum wrote:
> >>> Hi,
> >>>
> >>> Mike Snitzer <snitzer@...17...>  wrote:
> >>>  > Which IO scheduler are you using?  Please have a look at this redhat
> >>>> bug; turns out that there is definitely an issue with cfq (not
> >>>> exclussive to redhat):
> >>>> https://bugzilla.redhat.com/show_bug.cgi?id=241540
> >>>  >
> >>>> But if you're seeing this on stock kernel.org I'd imagine your just
> >>>> using anticipatory.
> >>> ... I'll do some more thorough testing to be sure that this was indeed all of it,
> >>> but at first sight it looks like this is also what is striking me - On all machines
> >>> but one (accidentally also the only one without interface bonding) I had "cfq"
> >>> as the default IO scheduler. After settting the scheduler for /deb/nbd0 to "noop",
> >>> so far it is running without a problem ...
> >> Ha ha!  It works!  Thanks a lot for the pointer.  For future reference
> >> to other people who may not want to go on a search, the secret password is:
> >>
> >> echo noop > /sys/block/nbdX/queue/scheduler (replace X as necessary).
> >>
> >> I have only one further question:  Is it safe to change the scheduler
> >> while the device is mounted?  I'd like to put that command in my
> >> rc.local, but it will only run after the device is mounted.
> >
> > Actually it was pointed out (later in the redhat bugziila, by David
> > Miller) that using the noop scheduler for nbd really isn't too smart.
> > That is because the IO requests aren't coalesced before they're sent
> > out over the wire via nbd.  As a result nbd becomes more inefficient.
> > Using cfq is fine provided you patch the kernel (like redhat has
> > done).  I've been using deadline with good success.
> >
> > And its perfectly safe to change the IO scheduler while the device is
> > mounted.  But as I said, you really do want to use one of the
> > schedulers.  So just changing the system default via elevator=<blah>
> > on the Linux kernel command line may be sufficient.
>
> It seems that there aren't any problems using the other schedulers
> (anticipatory or deadline), so should I just pick one of those? (for the
> time being until I can patch my kernel)

YMMV, but deadline has proven the most performant for my workloads.
You should benchmark both and see which is better.

Mike