Bug#309405: workaround

[Martin Koeppe <mkoeppe@gmx.de>]

On Wed, 24 Dec 2008, Moritz Muehlenhoff wrote:

> reassign 309405 linux-2.6
> thanks
>
> On Sat, Oct 01, 2005 at 02:50:48PM +0200, Martin Koeppe wrote:
> > Hello,
> >
> > in between I managed to work around the problem.
> > With cifs 1.34 compiled into 2.6.11 (and probably 2.6.12, too),
> > a acceptable performance can be achieved with mount option "direct".
> >
> > For further plans and (hopefully) the final fix for the problem see:
> > http://lists.samba.org/archive/linux-cifs-client/2005-September/000989.html
> > http://lists.samba.org/archive/linux-cifs-client/2005-September/000991.html
> >
> > In between this bug should remain open.
>
> Has this been resolved in current kernels?

I don't use linux-cifs any more since 2 years, instead I use MS NFS 
server on the server, which is much more reliable for me.

But I have now tested this problem, and it is still there:

Client:  Debian lenny, 2.6.26, cifs 1.56
Server:  Windows 2003 x64 SP2
Gigabit LAN

# mount //server/scratch /mnt -o serverino,user=...,pass=...

With this mount I still get decreasing performance while copying a 
large file (600MB) from client to server (i.e. write to the share). (I 
used Midnight Commander (mc) for copying, and its throughput 
measurement.) Apparently the file is cached locally first, and this 
cache is then swapped out, later swapped in, and then transferred, 
causing increasing delays while copying. No problem when adding 
"directio". Disk activity for the non-directio case is very high on 
the client, compared to the directio case.

NFS read   18 MB/s
NFS write  15 MB/s

CIFS directio read   14 MB/s
CIFS directio write  14 MB/s

CIFS non-directio read   17 MB/s
CIFS non-directio write   9 MB/s *

* begins with 19 MB/s and then constantly decreases to 9 MB/s for the 
whole transfer in average.


Martin


> Cheers,
>        Moritz