[linux-cifs-client] Re: CIFS slowness & crashes

[Steven French]

I did some more testing this weekend with 100MB files, repeated write then
read tests on 2.6.13-rc3 and it did not look too bad with cifs.   On the
write test cifs was about 1/4 the speed of local, but it stayed in cache so
the read tests were about the same as local.   smbfs failed to complete the
tests (Connectathon "NFS basic" tests 5a and 5b).

> This seems quite bad. Is it waiting for packet confirmations or what?
That (TCP ack timeout) is a big chunk of the problem for the case of RedHat
server (and those who shrink the tcp send/receive bufs smaller than SMB
packet size) - but that is easy to fix in smb.conf by using the default
socket options and not overriding them to make them worse).  CIFS VFS uses
larger frames than smbfs for reads - for gige bigger frames would normally
be a lot better unless the sndbuf were artificially restricted on the
server (as some mistakenly do).   SuSE seems ok there.

Another part of the problem is getting more read requests on the wire at
one time which requires more changes to cifs_readpages and writepages.

Obviously once that is done getting a single samba server instance to
better handle more requests at one time - via async requests would be the
next task, but that Samba server work is currently held up waiting for
Linux to fix the API into the 2.6 kernel async i/o system which is nearly
unusable.

I am setting up another small gige test environment to get some numbers -
but I don't think we should have a problem getting to 50% utilization or
better with minor tweaks.


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