Re: [PATCH -mm] [3/3] Add the Elevator I/O scheduler

[Nate Diller]

On 8/4/06, Al Boldi <a1426z@xxxxxxxxx> wrote:
> Nate Diller wrote:
> > This is the Elevator I/O scheduler.  It is a simple one-way elevator,
>
> Thanks for yet another attempt to achieve an efficient 2.6 elevator.

hopefully it's useful for you, I'd love to see performance results one
way or the other.

> > +static inline char contig_char(struct el_request *e)
>
> You probably meant el_req.  Check the rest.
>
> > +static void print_queue(struct request_queue *q, struct el_data *el)
> > +{
> > +       struct el_req *e;
> > +
> > +       printel(el);
>
> Should be print_el_data.
>
> Applied against 2.6.17, it boots with this:
>         io scheduler elevator registered (default)
>         elevator: forced dispatching is broken (nr_sorted=13), please report this

damn, i thought i fixed that.

> > +In pure form its largest weakness is starvation of other processes due to
> > +one process writing a very large number of contiguous requests (e.g.
> > +tarring a very large tar file while other processes are trying to run).
>
> cat /dev/hda > /dev/null starves the rest of the system.

yeah, that can be worse on some systems than others, it seems to vary
a bit with file system and device driver.  some setups cause more
"breaks" in the stream of requests.

> > +The max_contig and max_write tunables are two (imperfect) solutions. They
> > +
> > +These two tunables are still under construction, but they have proven
> > +somewhat useful in practice.  Usually, max_contig should be the same size
> > +(in bytes) as ra_pages.
>
> It's 0 by default.  Setting it to ra_pages prints this:
>         nate 1977: max_contig went backwards

hmmm, haven't tested this recently.  dunno why that would have broken...

> > +
> > +  SSTF
> > +
> > +The SSTF feature was added on a whim.  It ignores the tunables, and
> > probably +breaks tracing.  I didn't ever see it perform better than SCAN,
> > but who +knows?
>
> Starves too.

oh, yeah, definitely, this algorithm is only suitable for a workload
where you want latency over fairness.  suppose i should mention that.

Thanks for your comments and testing

NATE
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/