[Samba] Client link utilization

[Bostjan Skufca]

This is not the whole truth. With smbclient command line tool I have
the same exact problem (max 3,8MB/s).

b.


On 9 February 2010 17:19, Jeremy Allison <jra at samba.org> wrote:
> On Tue, Feb 09, 2010 at 10:25:44AM -0500, Lennart Sorensen wrote:
>> On Tue, Feb 09, 2010 at 01:19:40AM +0100, Bostjan Skufca wrote:
>> > Then I am greedier (if I also strive for 11MB/s:) Thanks for reminding
>> > me though.
>> >
>> > Can/Did someone push it over 10MB/s (or 100MB/s with 1Gbps ethernet))?
>> >
>> >
>> > To Jeremy or someone who is involved in samba as a developer: do any
>> > particular kernel options influence performance of smbclient, that you
>> > know of? Compile time options? Compiler version?
>>
>> Well I just checked what settings my NFS mount that works great is using.  it uses:
>> rw,relatime,vers=3,rsize=1048576,wsize=1048576,namlen=255,hard,proto=tcp,timeo=600,
>> retrans=2,sec=sys,mountvers=3,mountproto=udp
>>
>> Now I tried different rsize values with cifs as well, but that doesn't
>> seem to help (at least 64k made no improvement.  Going down to 4k made
>> it slower).
>>
>> Looking at it carefully, it turns out trying to set it over 16k gets
>> ignored.  Changing the option CIFSMaxBufSize on the cifs module to 130048,
>> does allow using rsize up to that value, and the speed goes up to 8.3MB/s
>> from 7.5MB/s.  So it seems larger rsize makes it better.  Given nfs uses
>> 8 times larger yet and gets great speed may be a clue too.
>
> There are two issues here. One is the problem with the
> Linux CIFSFS which you're talking about. FYI. Samba is
> able to cope with single request read/write sizes up to
> 16MB from a UNIX extension client (smbclient can do this)
> so you might want to talk to Steve French to learn how
> to turn this on.
>
> The second issue is the one with Windows clients. As
> the redirectors are completely different, with completely
> different requests and semantics I'd rather not get the
> two confused in this thread. Please split out another
> mailing list thread for the CIFSFS client speed issue,
> so we have them clearly separated.
>
> Jeremy.
>