Re: Pause intervals when copying/extracting an archive

[Ryusuke Konishi <ryusuke-sG5X7nlA6pw@public.gmane.org>]

Hi,
On Sat, 6 Jun 2009 16:49:28 +0300, nezmer-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org wrote:
> On Sat, Jun 06, 2009 at 04:56:53AM +0900, Ryusuke Konishi wrote:
> > Hi,
> > On Fri, 5 Jun 2009 22:32:56 +0300, nezmer-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org wrote:
> > > Sorry for the late reply .
> > > 
> > > I tried to use deplock but I always get this :
> > > SysRq : Show Locks Held
> > > INFO: lockdep is turned off.
> > > 
> > > sysrq t works . I hope the attached file will be helpful .
> > 
> > Thanks for the log!
> > 
> > According to you log, the performance problem seems to be caused
> > by a wait function in segment constructor of nilfs.
> > 
> > If you have a chance, please try the following coupled patch which
> > comes from two patches queued for 2.6.31.  The upper half of the patch
> > greatly improves read performance, and the lower part may have effect
> > on the write performance.
> > 
> > Regards,
> > Ryusuke Konishi
> 
> Unfortunately , the patches didn't help . I tried another flash disk. I
> tried another archive (.tar.gz this time)  and the problem is still
> there .
> 
> Is It possible the problem is related to snapshotting ? (the pauses
> happen around 7 times in 40 seconds)

I think the pause is related to garbage collection.  What will happen
if you mount the partition with -i option ?

   # mount -t nilfs2 -i /dev/xxx /mnt/media

This option skips invoking the garbage collector (cleanerd).


If it raises throughput, you can adjust by slowing down cleanerd by
rewriting /etc/nilfs_cleanerd.conf (a HUP signal must be sent to the
cleanerd to reflect the change):

   # The maximum number of segments to be cleaned at a time.
 - nsegments_per_clean     2
 + nsegments_per_clean     1

   # Cleaning interval in seconds
 - cleaning interval       5
 + cleaning interval       10


Thanks,
Ryusuke Konishi