IOsched: AS improvements (was: [Announce] [patch] Modular Scheduler Core and Completely Fair)

IOsched: AS improvements (was: [Announce] [patch] Modular Scheduler Core and Completely Fair)

[Al Boldi]

Nick Piggin wrote:
> Actually, I would still like to be able to deprecate deadline for
> AS, because AS has a tunable that you can switch to turn off read
> anticipation and revert to deadline behaviour (or very close to).

You mean antic_expire?  I tried it, and it does have an enormous effect on 
performance.  Smaller values do mimic deadline behavior in a way, as 
deadline has some strange behaviour on sync, so it's closer to noop, which 
is good for single user/proc access, while larger values are good for multi 
user/proc access.  But I always have to set it to 0, or else it starves sync 
with others reading.

I wonder if these tunables couldn't be autotuned by some form of 
history-based access pattern detection?

> It would have been nice if CFQ were then a layer on top of AS that
> implemented priorities (or vice versa). And then AS could be
> deprecated and we'd be back to 1 primary scheduler.
>
> Well CFQ seems to be going in the right direction with that, however
> some large users still find AS faster for some reason...

AS is faster no-contest.  CFQ has some deep block-io/queue problem that slows 
it down up to 25%.  It's related to max_sectors_kb and read_ahead_kb , and I 
mentioned this to Jens before, and AS show this problem sometimes too, but 
with CFQ it's much more prevalent.


Thanks!

--
Al