Re: [Ext2-devel] disk throughput

[Jens Axboe <axboe@suse.de>]

On Mon, Nov 05 2001, Mike Fedyk wrote:
> On Mon, Nov 05, 2001 at 08:18:36AM +0100, Jens Axboe wrote:
> > On Mon, Nov 05 2001, Jens Axboe wrote:
> > > 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.
> > 
> > Ok here it is. It's not as efficient as the 2.5 version since it
> > proceeds to scan the entire queue for a merge, but it should suffice for
> > your testing.
> > 
> 
> Does the elevator still favor blocks that are on the outside of the platter
> over all others?  If so, I think this should be removed in favor of a
> timeout just like the other seek requests...

It doesn't quite favor outside LBA's (lower numbers), it's a combination
of sector and device. It's hard to do this right in 2.4 because request
sectors are not absolute but a combination of partion + offset. 2.5 will
have this fixed, generic_make_request remaps buffer heads (well actually
bio's, but same deal) before submitting so the I/O scheduler can be a
bit smarter.

> I've been able to put a swap partition on the outside of my drive, and get
> utterly abizmal performance, while on another similar system, with swap on
> the inside of the drive performance was much better during a swap storm...

That sounds very odd, swap should be much faster on the outside.

-- 
Jens Axboe