On Thu, Aug 01 2002, Bill Davidsen wrote:
> On Thu, 1 Aug 2002, Jens Axboe wrote:
>
> > On Tue, Jul 30 2002, Bill Davidsen wrote:
>
> > > Now a request, if someone is running a database app and tests this I'd
> > > love to see the results. I expect things like LMbench to show more threads
> > > ending at the same time, but will it help a reall app?
> >
> > Note that the deadline i/o scheduler only considers deadlines on
> > individual requests so far, so there's no real guarentee that process X,
> > Y, and Z will receive equal share of the bandwidth. This is something
> > I'm thinking about, though.
>
> I'm not sure that equal bandwidth and deadline are compatible, some
> processes may need more bandwidth to meet deadlines. I'm not unhappy about
> that, it's the reads waiting in a flood of write or vice-versa that
> occasionally make an app really rot.
Correct, that's _exactly_ the case that made me attempt this type of io
scheduler. The current scheduler does well wrt optimizing for
throughput.
> > My testing does seem to indicate that the deadline scheduler is fairer
> > than the linus scheduler, but ymmv.
> >
> > > I bet it was tested briefly on IDE, my last use of IDE a week or so ago
> > > lasted until I did "make dep" and the output went all over every attached
> > > drive :-( Still, nice to know it will work if IDE makes it into 2.5.
> >
> > :/
> >
> > I'll say that 2.5.29 IDE did work fine for the testing I did with the
> > deadline scheduler, at least it survived a dbench 64 (that's about the
> > testing it got).
>
> Hum, looks as if I should plan to retest with 29 and the deadline patches.
> My personal test is a program doing one read/sec while making CD images
> (mkisofs) on a machine with large memory. It tends to build the whole
> 700MB in memory and then bdflush decides it should be written and the read
> latency totally goes to hell. I will say most of this can be tuned out at
> some small cost to total performance (thanks Andrea).
You are most welcome to test of course, I would very much like you to do
just that. 2.5.30 is out now and has the elv-queue* patch included which
included the necessary base changes for applying deadline-iosched, so
it's even easier now. I've attached the latest deadline-iosched-8 for
you.
-- Jens Axboe
This archive was generated by hypermail 2b29 : Wed Aug 07 2002 - 22:00:18 EST