[linux-cifs-client] Re: extremely slow throughput writing to remote

[Steven French]

linux-cifs-client-bounces+sfrench=us.ibm.com at lists.samba.org wrote on
05/23/2006 02:16:55 PM:

> > butters:~# mount -t cifs -o username=x,password=y //jesus/incoming
/mnt/
> > butters:~# dd if=/dev/zero of=/mnt/test bs=256K count=1000
> > 1000+0 records in
> > 1000+0 records out
> > 262144000 bytes (262 MB) copied, 26,9195 seconds, 9,7 MB/s
>
> The above was with 2.6.15 as in Debian testing.
>
> butters:~# dd if=/dev/zero of=/mnt/test bs=256k count=1000
> 1000+0 Datensätze ein
> 1000+0 Datensätze aus
> 262144000 Bytes (262 MB) kopiert, 9,61751 Sekunden, 27,3 MB/s
> butters:~# dd if=/dev/zero of=/mnt/test bs=256k count=1000
> 1000+0 Datensätze ein
> 1000+0 Datensätze aus
> 262144000 Bytes (262 MB) kopiert, 9,75895 Sekunden, 26,9 MB/s
>
> That's with 2.6.16 as in Debian unstable. Much better. This is about
> 210mbps, a little below half of the 540 I get with NPtcp between the
> same machines.
>
> Assuming you meant 50% of available raw bandwidth (which seems to be
> only 540mbps) and not of theoretical bandwidth (1000mbps), the
> performance is fine now.
>
> Was there a change that could explain that between 2.6.15 and .2.6.16
> in mainline, or do I get to verbally flog someone in the Debian kernel
> team? :)

It may simply be the larger write size that is helping in 2.6.16 (I don't
remember exactly when the default write moved from 4K to 56K).  You can
see the write size being used in a network analyzer or perhaps simply
estimate
it by resetting cifs stats ("echo 1 > /proc/fs/cifs/Stats") then doing the
copy and looking in /proc/fs/cifs/stats dividing bytes written by number
of write requests.

> As for the generally low bandwidth, I wonder if it's normal ... then
> again, one of the cards is only 32bit/33mhz and the traffic has to go
> through 3 D-Link switches, of all things :)
>
It makes it a lot harder if there is buffering in the route since large tcp
sends
(depending on tcp configuration) can overflow buffers, causing lost
packets, and retries slowing this a lot.


Steve French
Senior Software Engineer
Linux Technology Center - IBM Austin
phone: 512-838-2294
email: sfrench at-sign us dot ibm dot com
-------------- next part --------------
HTML attachment scrubbed and removed