Re: Eradic disk access during reads

[Willy Tarreau <willy@w.ods.org>]

On Sun, Sep 18, 2005 at 07:16:46AM +0300, Al Boldi wrote:
> Willy Tarreau wrote:
> > On Sat, Sep 17, 2005 at 01:32:53PM +0300, Al Boldi wrote:
> > > Willy Tarreau wrote:
> > > > On Sat, Sep 17, 2005 at 07:26:11AM +0300, Al Boldi wrote:
> > > > > Monitoring disk access using gkrellm, I noticed that a command like
> > > > >
> > > > > cat /dev/hda > /dev/null
> > > > >
> > > > > shows eradic disk reads ranging from 0 to 80MB/s on an otherwise
> > > > > idle system.
> 
> New synonym: eradic=erratic :)

I've noticed from the first mail too :-)

> > Denis' tool seems clearly more suited to analyse your problem.
> 
> The problem seems to be a multi-access collision in the queue, which forces a 
> ~50% reduction of thruput, which recovers with another multi-access 
> collision.  Maybe?!

do you have anything else connected to the same controller ? I've encountered
a problem a long time ago with SATA drives (it was around 2.4.25, with the
beginning of SATA support), where I could not use RAID on my drives because
simultaneous accesses to the drives was not possible. I then realized that
every time I accessed a drive during a read from or a write to another one
would stall. I don't remember if CDROM access caused the same trouble. This
was with a SATA-patched ata-piix driver, but fortunately this is fixed by now.

Regards,
Willy