[Samba] 3.0.0 on RH9, throughoutput problems

[Brett Dikeman]

Greetings,

We're deploying a Samba 3.0.0 server running RedHat 9 and having 
performance problems.

The system specs are as follows:

-P4 Xeon 2.4ghz
-genuine intel server mobo(SE7501BR2) w/one gigabit, one 100baseT 
interface.  Using the gigabit interface with the kernel's e1000 driver.
-1GB/ram
-four 10k U320 drives, hardware RAID 5 on a Megaraid 320 w/64MB of 
built-in cache.  The shares are off a 200GB reiserfs partition mounted 
w/noatime.
-latest 2.4.20 SMP redhat kernel built for i686 from Redhat 9 updates
-latest glibc RPM for i686

Informal testing has shown local data transfer rates in the 100MB/sec 
ballpark.  Bonnie++ shows write speeds of over 50MB/sec for 
"intelligent" writes(getc/putc are much worse, 5MB/sec, not sure what 
that's about, but my desktop gets similar numbers save topping out at 
15MB/sec on the intelligent tests).

The network is all switched 100BaseT-FD and mostly Win2k with the 
occasional WinXP box.  The three clients I'm testing with are a Win2k 
box(latest  service packs etc) a Mandrake 9.1 box, and a 1Ghz Powerbook 
G4. Both PCs are at least P3-500 class systems.

I'm getting about 6.5-6.7MB/sec read on both systems when pulling down a 
several hundred MB, cached file. Write performance, at least from my 
Powerbook G4, is identical(I've done transfers between two powerbooks 
over a gigabit crossover at over 30MB/sec, so I know the powerbook can 
push data).  When I did a crossover connection to the server, I saw 
occasional peaks as high as 20MB/sec read and write, but then radical 
drops to 1-2MB/sec or so for several seconds; it's very inconsistent.

It gets better.  Start a copy to the server, and both upload and 
download drop to about 1.2MB/sec.  That reeks of a duplex mismatch, but 
there isn't any.

   That's not even close to acceptable performance from such a powerful 
system.  I've seen 10+MB/sec at a previous employer with a Netapp F720 
filer.  Granted, it had 7-8 FC disks- but in this case, it's moot- the 
file is cached in ram for the read tests.

CPU utilization is not an issue; combined system+user time is barely 
4-5% during a copy.  Network's fine- there are no errors according to 
ifconfig on both systems, and the card's media interface is correctly 
set to 100baseT-FD(switch lights match this).  I've double-checked all 
aspects of the interface's configuration, they are correct.  By running 
the box using a direct crossover link and gigabit ethernet, our 
network's been all but eliminated as the source of the problem.

I have tried all manner of socket options, buffer sizes, etc...including 
all defaults.  There has been little to no effect from changing the params.

Any suggestions?  I'm running out of time to deploy this system...

Thanks all,

Brett