Packet collision on Samba traffic

[Richard Sharpe]

At 03:37 PM 4/5/01 -0700, Andrew Tridgell wrote:
>Okuyama-san, 
>
>> How about
>> 
>> http://www.dd.iij4u.or.jp/~okuyamak/Documents/tuning.english.html

Hmmm, well an SMB style test shows similar results. Samba never fills the
window, because Win95 uses an SMBreadBraw request with a max size of 4608.
Samba sends 4 segments, 3 with 1460 bytes, and 1 with 232 bytes. Windows
Acks and opens the window to 8760.

However, this time around, it takes Samba/Linux 16 mS to get the first
segment out in response to the SMBreadBraw, with the subsequent segments
taking 10 mS to get out the door.

In the FTP example, OTOH, Linux can pump the segments out within a mS of
each other, but it takes Windows a long time to ack, like 40mS. However, it
takes about 100mS to pump out the Ack ...

However, the behaviour for the FTP case is much more complex, with Windows
occassionaly sending an Ack for 8760 bytes of data, and then, 5mS later,
sending a window update for 8760. I have also seen Linux get very confused,
and send 8 segments with 1024 bytes of data, but I think that is a file
system problem.

Now, I am really interested to know what sort of NICs cause the behaviour
seen in the Web paper. I am sceptical ...

I look at Windows-based packed traces a lot, and I give TCP/IP courses. On
the courses we examine packet traces ... One of them is FTP from Linux to
Windows, and I have never seen this behaviour, whether on Win9X (native,
not under VMware), or WinNT to Linux.
 
Tomorrow I will check WinNT to Linux over 10 Mb/s and 100Mb/s as well as
Win98 and WinNT to Linux via 2Mb/s wireless. Mixture of versions, both SP1
and SP6.

>Cheers, Tridge
>

Regards
-------
Richard Sharpe, sharpe at ns.aus.com
Samba (Team member, www.samba.org), Ethereal (Team member, www.ethereal.com)
Contributing author, SAMS Teach Yourself Samba in 24 Hours
Author, Special Edition, Using Samba