Re: wrong percent with vsync+randrw

From: Jens Axboe <jens.axboe_at_oracle.com>
Date: Mon, 7 Apr 2008 10:54:56 +0200

On Mon, Apr 07 2008, Zhang, Yanmin wrote:
> On Mon, 2008-04-07 at 10:20 +0200, Jens Axboe wrote:
> > On Mon, Apr 07 2008, Zhang, Yanmin wrote:
> > > On Mon, 2008-04-07 at 10:04 +0200, Jens Axboe wrote:
> > > > On Mon, Apr 07 2008, Zhang, Yanmin wrote:
> > > > > Jens,
> > > > >
> > > > > I ran a testing with vsync+randrw. I expected the both read and write's precent are
> > > > > 50%. But vmstat showed write was more than 99%.
> > > >
> > > > The problem here is that the window service is too small, because of
> > > > your big queue depth. I modified the fio algorithm to only look into ios
> > > > issued, I think that is the fair metric. With that, you should have a
> > > > nice distribution between reads and writes - can you pull the update and
> > > > see that it works for you?
> > > It's hard to pull in company as IT configures firewall to filter most git
> > > access. Would you like to send me a patch against the lastest tarball? Or just
> > > upload the latest tarball on the snapshot website?
> >
> > No problem, I kicked the snapshot script manually, so the current one
> > should have it.
> I downloaded it but got below info when untar it. I hit the similar issue
> before. Although I have no md5sum, I think perhaps something is wrong about
> your tarball, or the snapshot scripts?
>
>
>
>
> fio/HOWTO
>
> gzip: stdin: decompression OK, trailing garbage ignored
> fio/blktrace_api.h
> tar: Child returned status 2
> tar: Error exit delayed from previous errors

There's a MD5SUMS file in the same directory, please compare the sum
with what you downloaded. It looks fine from here.

-- 
Jens Axboe
Received on Mon Apr 07 2008 - 10:54:56 CEST

This archive was generated by hypermail 2.2.0 : Mon Apr 07 2008 - 11:00:02 CEST