Re: [Ext2-devel] disk throughput

Mike Fedyk (mfedyk@matchmail.com)
Sun, 4 Nov 2001 23:14:47 -0800


On Mon, Nov 05, 2001 at 08:06:35AM +0100, Jens Axboe wrote:
> On Sun, Nov 04 2001, Andrew Morton wrote:
> > The meaning of the parameter to elvtune is a complete mystery, and the
> > code is uncommented crud (tautology). So I just used -r20000 -w20000.
>
> It's the number of sectors that are allowed to pass a request on the
> queue, because of merges or inserts before that particular request. So
> you want lower than that probably, and you want READ latency to be
> smaller than WRITE latency too. The default I set is 8192/16384 iirc, so
> go lower than this -- -r512 -w1024 or even lower just to check the
> results.
>

Does the elevator do better with powers of two?

> > This was based on observing the request queue dynamics. We frequently
> > fail to merge requests which really should be merged regardless of
> > latency. Bumping the elvtune settings fixed it all. But once the
> > fs starts writing data out contiguously it's all academic.
>
> Interesting, the 2.5 design prevents this since it doesn't account
> merges as a penalty (like a seek). I can do something like that for 2.4
> too, I'll do a patch for you to test.
>

I'd be very interested in this patch. Can you post it pubically?

> --
> Jens Axboe
>

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