[mythtv-users] Raid Performance Tweaking

[Jens Axboe]

On Sat, Jul 08 2006, Blammo wrote:
> On 7/8/06, Jens Axboe <mythtv-users at kernel.dk> wrote:
> > On Fri, Jul 07 2006, Blammo wrote:
> > > While your average current disk drive, on sustained SINGLE threaded
> > > reads, sits about 40MB a second, flipping to reads/writes, or 2/3/4
> > > simult reads/writes from different parts of the disk, and now your
> > > throughput on that single spindle is going to be WAY down below that
> > > magical 40MB number, and you will start having performance issues.
> >
> > That's only really true for the 'classic' io schedulers that are work
> > conserving. At least on Linux, using 'cfq' or 'as' easily do a lot
> > better for many reading clients (or mixed read/write workload) at the
> > cost of slightly higher latency.
> 
> Agreed, that can help a whole bunch. I'm a big fan of "deadline"
> myself for my backends.

deadline is a fine simple scheduler, however it is work conserving and
will not do well throughput wise for multiple "clients".

> However, scheduler or no, random read/write speeds are always lower
> than sequential, which is a close comparison to an example of 3
> clients reading/writing to the disk at the same time, plus local
> activity.

Certainly they are lower, but they don't have to be in the 5-10% range
of disk throughput. 90% is perfectly achievable, if you tolerate a
little higher maximum latency (average latency should still be good).

-- 
Jens Axboe