Anticipatory disk scheduling - soc 2008

[Nirmal Thacker]

On Tue, May 20, 2008 at 3:05 PM, Matthew Dillon
<dillon at apollo.backplane.com> wrote:
>
> :Nirmal Thacker wrote:
> :> I thought of exactly the same thing (except I was not sure where to
> :> add this)since I assume that the scheduler which I write should be
> :> accessbile as an interface (hence add an API) rather than plugging it
> :> into the NATA driver. However this would require me to modify (would
> :> it?) calls in the nata code - is this OK?
> :
> :Sure thing.  I think eventually all disk drivers which have latency due t=
> :o=20
> :seeks would use this API.
> :
> :>>    This would allow you to test your scheduler with a vkernel by also =
> :having
> :>>    the VKD driver call it (/usr/src/sys/dev/virtual/disk).
> :>=20
> :> Im not sure of what I understand about the vkernel. Is this a common
> :> feature in BSD systems or only the DragonflyBSD? Could I have more
> :> information (or hyperlinks) to this and how it is important or can
> :> help me in testing my scheduler, as you mention above.
> :
> :This is D-d-d-dragonFly exlusive(-ive-ive-ive)!  Have a look at vkernel(7=
> :).
> :
> :cheers
> :   simon
>
>    And if you really wanted to go whole-hog, create a really generic API
>    that implements multiple disk schedulers and move the elevator algorithm
>    into it as one of several schedulers, and your anticipatory algorithm
>    as another scheduler.
>
>    And if you wanted to get ultra-fancy you could implement disk-layer
>    ioctl's to set the scheduler.  Ok, that might be going too far, but
>    it's a cool idea!

Yes sounds cool! I'll try to bite off what I can chew .

Thanks Matthew and Simon- that was a lot of good advice.

I will now look at Peter Druschel's scheduler in FreeBSD 4.8 and try
to compare what could go here, what could require more addition wrt
DragonflyBSD. Of course Matthews idea of adding  kernel options seems
appropriate to have the NATA scheduler as well (for the moment).

The vkernel sounds like a very interesting concept.

Nirmal

>
>    I suspect that there is a lot of room for experimentation in this area
>    of research, not only elevator and anticipatory style schedulers but
>    also schedulers which reorder reads to improve performance, or are more
>    tag-friendly, and so on and so forth.
>
>                                        -Matt
>                                        Matthew Dillon
>                                        <dillon at backplane.com>
>