Tidbit for improving Samba performance

[Mark Powell]

On Tue, 28 Aug 2007, Brooks Talley wrote:

> Right now, with net.inet.tcp.recvspace and net.inet.tcp.sendspace both 
> at 65536 and SO_RCVBUF and SO_SNDBUF also at 65536, I'm seeing about 
> 60MB/s.  Still not great, but a significant improvement.  FWIW, going 
> higher than 65536 causes the Vista workstation to lose its mapped drive 
> all the time.  And changing the MTU on both FreeBSD and Vista (to either 
> 4088 or 9014) had no discernable benefit.

I've been doing the same. I must have tried so many combinations though.
It really is a black art :(
   I agree that 65536 for all 4 options is probably about the best 
combination. Thanks.
   I'm seeing 54MB/s peak write from XP box to FBSD though and 40MB/s peak 
read.

> This is on a 7.0-CURRENT compiled without all of the debugging stuff, 
> and it's at 0.7 load and 73% idle when transferring 60MB/s, so I think 
> there's still some headroom there for improvement.

This is onto a ZFS RAIDZ1 onto 3x500GB drives which can achieve 120MB/s 
write and 160MB/s read under FBSD. This is between 2xPCI-X Intel PRO1000 
server adapters 82545 in 32bit/33MHz PCI slots. They are the only active 
devices on the PCI bus, so I would expect more. Simple tests with dd when 
FBSD is on both boxes, can easily achieve 87-90MB/s.

> Anyways, thought I'd share with the group, since I haven't seen this 
> mentioned in copious amounts of Googling.

I too found a derth of information on this matter for FBSD. I'm current 
using the following in smb.conf:

socket options = TCP_NODELAY IPTOS_LOWDELAY SO_SNDBUF=65536 SO_RCVBUF=65536
strict locking = no
use sendfile = yes

What do you have in there?
   Cheers.

-- 
Mark Powell - UNIX System Administrator - The University of Salford
Information Services Division, Clifford Whitworth Building,
Salford University, Manchester, M5 4WT, UK.
Tel: +44 161 295 4837  Fax: +44 161 295 5888  www.pgp.com for PGP key