[Samba] samba 3 performance issues

[Collen Blijenberg]

Beside the story Greg Folkert wrote (witch make sense)
remember this about GB network carts:

You'll never get a full 1000 mb/s !
i saw you have a celeron processor (witch ain't the fastest in performance)
coz' gb nic's tent to use a lot processor overhead, also raid cards use 
(a little) processor time.
hdd through put is an issue, all together makes the performance.

I've tested 2 marvel yukons (pci-x) between 2 xp clients, with only 
memory transfers and got (after some windows tweaking, coz windows isn't 
gb lan ready by default!!!!) 500mb/s (so that is only 50%)
and had 99% proc. load. (pentium M 1.8)
on realtek gb lan cards it was even worse.

so a big processor(s) and real fast hdd's might do you some good!

so put this together with Greg's story, and you'll get the fact's.

Cheers, and good luck with testing/tweaking

Collen.

Greg Folkert wrote:
> On Thu, 2006-03-30 at 13:52 -0500, Rohit Kumar Mehta wrote:
>> I believe I have some hardware or configuration related performance 
>> issues running samba 3.0.14a-3sarge.
>>
>> Our server is an Intel Celeron 2 Ghz with 512 MB of RAM and a 3ware
>> card using SATA disks in a RAID 5 configuration (3ware controller card). 
>>   We have a gigabit network and are using Intel Gigabit ethernet cards 
>> e1000).
>>
>> When copying large files to the samba shares on the system, the transfer
>> rate maxes out near 100 mb/s.  We tested with nttcp and were able to get
>> speeds of nearly 800mb/s.  So I think it is safe to conclude this is not 
>> a network issue.
>>
>> Various tools like top, xosview and mpstat convinced us that we are 
>> bound in the CPU. Stopping the samba file transfer and the cpu idle time 
>> exceeds 90%.  We are convinced that our CPU is the bottleneck,
>> but not sure why.
>>
>> #cat /proc/cpuinfo
>> processor       : 0
>> vendor_id       : GenuineIntel
>> cpu family      : 15
>> model           : 2
>> model name      : Intel(R) Celeron(R) CPU 2.00GHz
>> stepping        : 9
>> cpu MHz         : 1996.920
>> cache size      : 128 KB
>> fdiv_bug        : no
>> hlt_bug         : no
>> f00f_bug        : no
>> coma_bug        : no
>> fpu             : yes
>> fpu_exception   : yes
>> cpuid level     : 2
>> wp              : yes
>> flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge 
>> mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe cid
>> bogomips        : 3956.73
>>
>> Does anyone have any advice on how to speed up our file transfers? We 
>> regularly have to 18 GB worth of files to this system, and it would be 
>> very good if we could speed it up.  At current speeds, we get no 
>> advantage at all from even having gigabit network cards!
>>
>> Please feel free to ask me any other questions about our system setup. 
>> Thanks in advance for any advice,
> 
> Have you done *ANY* system caching parameters, filesystem tuning, or
> Samba Config tuning?
> 
> What have you done besides verify it is not the network itself?
> 
> Have you tested throughput for the 3ware card?
> 
> I can tell you this, if you have the RAID-5 setup not-optimally to work
> with the block sizing on your Filesystem you'll never get excellent
> throughput.
> 
> I always tend to use largest blocking factors with the 3ware cards for
> RAID-5. This (for me at least) has proven the fastest and least latency
> ridden settings for me. But then I am using XFS on all of my 3ware
> RAID-5 setups.
> 
> For Mirroring, I typically let the defaults work. Defaults have been by
> far the best setup for most filesystems. If you still believe you are
> suffering from CPU overload, I'd suggest sending it to the RAID-5 array
> with over compressed scp (with mild compression of 4 or 5) and then
> without compression. See what you get.
> 
> I am betting the real problem comes from multiple bus-mastering cards
> conflicting or colliding. The Intel-E1000 and the 3ware card are
> definitely both bus-mastering.
> 
> There are a couple of things on the Samba side you can do. Turn off
> Logging (you don't need it really), change the read and send buffer
> sizes, change the TCP setting it uses to be more in line with Gigabit,
> move to using Jumbo frames, get a TOE (TCP Offload Engine) NIC.
> 
> Then if you still have issues, turn on logging for the stuff you are
> worried about (auth would be 0, etc...) and then add a sniffer to you
> connection. You'll definitely find something. My gut reaction is that
> since this is a Celeron Processor, you really need to goto 64-bit slots
> on the mother board. Getting a PCI-X capable motherboard would greatly
> help your problems.
> 
> One last thing, any of the 95xx cards from 3ware are 3.3V only and are
> PCI2.3 compliant, they will function incorrectly possibly even be ruined
> or not recognized by a 5V or Auto-detect 5v/3.3v slot. The 9xxx, 8xxx
> and 7xxx cards can be used in either a 5V or 3.3V PCI slot. 
> 
> Good luck.
>