Re: Playing with SATA NCQ

[Jens Axboe <axboe@suse.de>]

On Sun, May 29 2005, Jeff Garzik wrote:
> On Sun, May 29, 2005 at 09:04:21PM +0200, Jens Axboe wrote:
> > On Sun, May 29 2005, Mark Lord wrote:
> > > My basic hdparm timing test shouldn't show much of a difference
> > > with NCQ tests, becase hdparm just does a single request at a time,
> > > and waits for the results before issuing another.  Now, kernel read-ahead
> > > may result in some command overlap and a slight throughput increase, but..
> > > 
> > > Something like dbench and/or bonnie++ are more appropriate here.
> > 
> > I don't like bonnie++ very much and dbench is very write intensive. I
> > would suggest trying tiotest, find it on sf.net. It gets easily readable
> > results and they are usually fairly consistent across runs if you limit
> > the RAM to something sensible (eg 256MB and using a data set size of
> > 768MB).
> 
> As an FYI...  download Stephen Tweedie's verify-data tool at
> http://people.redhat.com/sct/src/verify-data/

Interesting, will try it tomorrow.

> Robin Miller's 'dt' is also nice to have.

Yep, have tried that in the past. I'm just recommending tiotest as an
easy and good way for people to test performance quickly. Just boot with
256MB and use eg tiobench.pl --threads 8 should be a good way to test
NCQ.

-- 
Jens Axboe