Sorting disk I/O requests... [LWN.net]

Sorting disk I/O requests...

Posted Sep 30, 2002 0:41 UTC (Mon) by Baylink (subscriber, #755)
Parent article: A new deadline I/O scheduler

Ok, I suppose the possibility exists that either

a) I don't know how to read, or

b) Everyone else involved here is missing something...

but every reference I've read on current geenration hard disk technology suggests strongly, if it doesn't say so outright, that it became pretty much useless for the kernel to try to make *any* assumptions about what was where on a drive (and therefore which requests should be filed with the drive in which order) right about the time that drives started lying to the outside world about what was going on inside them.

In particular, Scott Mueller's _Repairing and Upgrading PC's_ (I have the "Linux Edition", a tweaked version of the 11th) says this pretty much explicitly.

And yet these guys are trying to do it anyway.

Since what I read meshes with what I can figure out on my own from the facts given, I have to assume that there's *something* here that I'm missing.

Or else they're just betting that it will help enough to be worth doing, even though it's not going to work "right".

But maybe it's just me.

So many things are just me.

(Log in to post comments)

Sorting disk I/O requests...

Posted Oct 2, 2002 11:49 UTC (Wed) by axboe (subscriber, #904) [Link]

You cannot say for sure where anything is on a hard drive, but you can make some basic assumptions that generally are true. The assumptions that the deadline io scheduler makes is that:

- Generally, sector x and x+1 are contigous on media.

That is it. I say generally, because x+1 might be remapped due to defects, for instance. But you'll find that in most cases the above is true. Of course drive makers are free to do whatever they want, but general performance breakdown would occur in _all_ os's if a streamed read from LBA 0 to LBA 2048, eg, would incur lots of seeks.

deadline doesn't even attempt to say that a seek from a to b is more or less costly than a seek from x to y.

Maybe you should read the source. I'll bet you it's a lot better source than some 'repairing pcs' book.

Sorting disk I/O requests...

Posted Oct 8, 2002 2:27 UTC (Tue) by tigerand (guest, #6150) [Link]

This doesn't sound like it takes software raid (md) or filesystem I/O footprints into account. So if either of those two subsystems have any I/O scheduling type optimizations, then these optimizations are almost certainly going to clash with very unsatisfactory results.

It sounds to me like this needs a heck of a lot more thought before it should be truely adopted. Even just sorting between reads and writes could clash with optimizations in some of the fancier journaling file systems. Delayed writes and the like could cloud the algorythm. There really needs to be a way to synchronize between file system code, low level block i/o code, and md code. For instance, I can see this scheme making a huge difference on a single disk with multiple file systems and maybe even a couple of chuckleheads dd'ing to a partition (or, god forbid, a user level file system like Coda or some databases that want to use raw partitions). In that case, a file system might want to wave off it's own optimizations in favor or this. But if it's a single file system on a disk situation, the file system code might know better, especially with it's advanced knowledge about what might be coming just behind this, but isn't in the queue yet. For instance, it knows that right now there are some data writes to accomplish, but right after that there's going to be some journal writes, so save up some pending journal writes just a tad, because there's a couple more just around the corner. This is an example of special knowledge, not a conflicting optimization. But you get the idea.

Bottom line, this seems like a thing that belongs in the file systems, and at the low level is probably the wrong place to "fix" this.