samba - Jun 2000 - Samba performance problem over a radio link.

Hi all.

We have a performance problem over a 2 Mbits radio link.

The configuration is as follows: we have a RH Linux 6.1 (double PIII
processor, 100 mbit ethernet, raid 5), acting as a server for a network
of Windows 95, 98 and NT. On the local network I can achieve up to 40
mbits transferring large files from the server to a windows 98 machine.

This network is connected through a 2 mbits radio link to another 10
mbits network. There we have an NT workstation, an 95 and a 98 machine.

A samba read performed by the 95 or the 98 machine provides
approximately 70KB/sec. An ftp get from the same server 100KB/sec. The
same read operation, performed by the WinNT machine, achieves 100KB/sec.
Reading from an NT server on the 100MBit network provides 160KB/sec,
which is near the maximum performance.

I read through the on-line O'Reilly book about performance, and I
experimented a lot using mtuspeed410. However, I was unable to produce
any meaningful improvement. We tryied from 296 to 1500 as MaxMTU value,
and from 2x to 6x RWin. The best performance were obtained by 1152 MTU
and Rwin x4

Now I understand performance can greatly vary according to the settings
in the socket options, however, since experimenting can be a problem for
us (the network is in another town, and I have to make tests onto
windows machines also), I would like to know if there is some
"reasonable" way to find the parameters to be set, both on the windows
machine and on the samba server. I understand trial and error is an
option, but I would like to know if there is anything else which could
be used to narrow all the possible options.

Please feel free to answer me directly.

Bye
Marco Bizzarri

Marco wrote:
| We have a performance problem over a 2 Mbits radio link.
>From your message, it looks like this is what you have, in
the same units: please correct me if this isn't right, as
I'm really reduced to guessing from your description...

Description				KB/S	Kbit/S	Mbits/S	% of max
Theoretical Max. 100 Mbit/S Eth.	12500	100000	100	100
local smb read, win9x			 5000	 40000	 40	 40
local smb read, win9x to NT server	  160	 1280	  1.28	  1.28

	This can't be true...

Theoretical Maximum, 2Mbit/S radio	250	2000	2	100
remote smb read, win9x			 70	 560	0.56	 28
remote ftp read, win9x			100	 800	0.8	 40
remote smb read, NT			100	 800	0.8	 40

	This looks plausible.

	I'm going to assume you want to know why Samba 
	with win9x clients is slower than ftp and NT clients
	over the speed-limited link.

	I **think** that the win9x SMB client program is not
	as good as packing lots of no-response-required
	operations into packets as the NT SMB client, and this
	shows up strongly on the speed-limited network. The
	client send a request, and starts waiting for an ACK.

	This is called "stop-wait" processing, for obvious 
	reasons (;-)) The desired behavior is to send a bunch 
	of requests and never wait. 

	Playing with the receive window will allow the sender to 
	send as many packets, containing SMBs, as possible, but the
	SMB protocol can defeat this.  I suspect that's why the
	improvement from playing with rwin levels off suddenly.

	TCP gurus, do please chime in!

	I think you might try two experiments: tuning the TCP for 
	large file throughput (normally not recommended) and turning
	off raw read/write (ditto).

	In three separate experiments, try
		read raw = no
		write raw = no
	then 
		socket options = TCP_THROUGHPUT IPTOS_THROUGHPUT
	then 
		both

	If any causes a significant improvement, please tell
	the list: these are contra-intuitive approaches...

	Once we're past this, we'll look at SO_SNDBUF= with large values.

--dave
-- 
David Collier-Brown,  | Always do right. This will gratify some people
185 Ellerslie Ave.,   | and astonish the rest.        -- Mark Twain
Willowdale, Ontario   | //www.oreilly.com/catalog/samba/author.html
Work: (905) 415-2849 Home: (416) 223-8968 Email: davecb@canada.sun.com