[Samba] Slow transfers from Samba server

[Sinisa Bandin]

Hello all

after browsing this list for keyword "slow" and having experienced this 
myself more than once, this is a summary of problems and possible 
resolutions as I see them:

Servers:
- new, non-brand name, but made of quality parts, MSI mobo, AMD dual 
core CPU, Seagate HDDs in raid-1 config, switch is HP ProCurve 10/100 
unmanaged
- older HP NetServer with Intel 10/100 NIC (e100 driver)


- all other transfers (scp, rsync, http, ftp) work fine, only Samba is slow
- when having more (2+) simultaneous transfers, Samba speeds up, but 
still goes slower than same number of "other" transfers
- playing with socket options can help, but only a bit, optimum seems to 
be SO_RCVBUF=32768 SO_SNDBUF=32768 TCP_NODELAY
- tried (almost) all combination of locking and other suggested 
performance related parameters, with almost no effect (blocking locks, 
oplocks, kernel oplocks, level2 oplocks, lock spin time, strict locking, 
strict sync, posix locking,  read bmpx, read raw, write raw, getwd 
cache, use sendfile)
- when I change NIC (old one is rtl8169 gigabit integrated, new one is 
rtl8139 10/100 USRobotics NIC), everything works fine
- ethtool shows autonegotiated 100 Mb/s full duplex for both NICs
- ifconfig shows no errors of any kind, ever
- samba versions 3.0.23d, 3.0.25c
- openSuSE 10.2, with kernels 2.6.18.8-5 (latest SuSE official update) 
and 2.6.22.xxxxxx from SuSE Kernel of the day project


Conclusion : there is something going wrong in communication between 
certain NIC drivers and Samba. Maybe some kernel developer should take a 
look at this.


I hope this will be helpful to someone.

With best regards,
Sinisa Bandin