[zfs-discuss] avgrq-sz and performance

RB aoz.syn at gmail.com
Tue Feb 14 01:23:48 EST 2012


On Mon, Feb 13, 2012 at 23:18, Manuel Amador <rudd-o at rudd-o.com> wrote:
> 1. ZFS does not coalesce contiguoous requests?
>
>
>
> 2. The deadline sched can only coalesce whatever is in the request queue.
>
>
>
> This means that the request queue needs to remain 128 or bigger for it to
> have any meaningful effect.
>
>
>
> This implies that the elevator in ZFS won't be able to reorder those
> already-submitted requests.

I think you're mistaking 'avgrq-sz' - that's "Average Request Size",
not average queue size, which in all tests I've done hovers around 20
when under full load.  It's just the requests sent to the disk for ZFS
top out at 65k, whereas for all other i/o tests I've run
(dd/btrfs/ext4) it hovers just under 128k (currently running deadline
tests included).



More information about the zfs-discuss mailing list