Re: stochastic fair queueing in the elevator [Re: [BENCHMARK] 2.4.20-ck3 / aa / rmap with contest]

[Andrea Arcangeli]

On Mon, Feb 10, 2003 at 03:44:35PM +1100, Nick Piggin wrote:
> Rik van Riel wrote:
> 
> >On Mon, 10 Feb 2003, Nick Piggin wrote:
> >
> >>Andrea Arcangeli wrote:
> >>
> >>
> >>>The only way to get the minimal possible latency and maximal fariness is
> >>>my new stochastic fair queueing idea.
> >>>
> >>Sounds nice. I would like to see per process disk distribution.
> >>
> >
> >Sounds like the easiest way to get that fair, indeed.  Manage
> >every disk as a separately scheduled resource...
> >
> I hope this option becomes available one day.
> 
> >
> >
> >>However dependant reads can not merge with each other obviously so
> >>you could end up say submitting 4K reads per process.
> >>
> >
> >Considering that one medium/far disk seek counts for about 400 kB
> >of data read/write, I suspect we'll just want to merge requests or
> >put adjacant requests next to each other into the elevator up to
> >a fairly large size. Probably about 1 MB for a hard disk or a cdrom,
> >but much less for floppies, opticals, etc...
> >
> Yes, but the point is _dependant reads_. This is why Andrea's approach
> alone can't solve dependant read vs write or nondependant read - while
> maintaining a good throughput.

note that for soem workloads the _dependant reads_ can have _dependant
writes_ in between. I know it's not the most common case though, but I
don't love too much to make reads that much special. But again, it makes
perfect sense to have it as a feature, but I wouldn't like it to be only
option, I mean there must always be a way to disable anticipatory
scheduling like there must be a way to disable SFQ (infact for SFQ it
makes sense to leave it disabled by default, it's only a few people that
definitely only cares to have the lowest possible latency for mplayer or
xmms, or for multiuser system doing lots of I/O, but those guys can't
live without it)

Andrea