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 03:58:26PM +1100, Nick Piggin wrote:
> Jakob Oestergaard wrote:
> 
> >On Sun, Feb 09, 2003 at 08:33:43PM -0800, Andrew Morton wrote:
> >
> >>David Lang <[e-mail]> wrote:
> >>
> >>>note that issuing a fsync should change all pending writes to 
> >>>'syncronous'
> >>>as should writes to any partition mounted with the sync option, or writes
> >>>to a directory with the S flag set.
> >>>
> >>We know, at I/O submission time, whether a write is to be waited upon. 
> >>That's in writeback_control.sync_mode.
> >>
> >>That, combined with an assumption that "all reads are synchronous" would
> >>allow the outgoing BIOs to be appropriately tagged.
> >>
> >
> >This may be a terribly stupid question, if so pls. just tell me  :)
> >
> >I assume read-ahead requests go elsewhere?  Or do we assume that someone
> >is waiting for them as well?
> >
> >If we assume they are synchronous, that would be rather unfair
> >especially on multi-user systems - and the 90% accuracy that Rik
> >suggested would seem exaggerated to say the least (accuracy would be
> >more like 10% on a good day).
> >
> Remember that readahead gets scaled down quickly if it isn't
> getting hits. It is also likely to be sequential and in the
> track buffer, so it is a small cost.
> 
> Huge readahead is a problem however anticipatory scheduling
> will hopefully allow good throughput for multiple read streams
> without requiring much readahead.

the main purpose of readahead is to generate 512k scsi commands when you
read a file with a 4k user buffer, anticipatory scheduling isn't very
related to readahead.

Andrea