Re: Checklist (sub-thread of: "Slow file sharing performance: lostpidgeons")

Les Mikesell wrote:
> On Tue, 2003-09-09 at 07:06, James Knott wrote:
> 
> 
>>I am aware the UDP is best effort and that the services using it are 
>>supposed to provide flow control if necessary.  Given that FTP, which 
>>uses TCP works fine and Samba and NFS, which use UDP have problems, what 
>>observations would you make?
>>
> 
> 
> A) The 'flow control' mechanism when going between a fast medium to
> a much slower one is to buffer some, then discard packets when the
> buffer is filled. TCP has a window size mechanism that will
> limit the unacked packets sent on a single connection. CIPE
> doesn't change this by using UDP - it just wraps the packets
> already being sent with another header during delivery.  Things that
> use UDP natively for reliable transfers (NFS, etc.) normally
> have their own windowing and error correction mechanism that
> can differ in size and backoff strategy from TCP's.

I suspect this may be the cause of the problem.
> 
> B) Samba uses TCP for file transfer, so any conclusions
> you made about UDP by observing samba are incorrect.

I don't know enough about SMB, but that's what I also thought.  However, 
there are some UDP packets as well.
> 
> C) If you want to diagnose where the problem is you will have
> to use tcpdump or some other protocol analyzer to see where the
> packets are being discarded. If you watch a tcp stream like
> samba going into the cipe interface you will see the tcp sequence
> numbers repeat as tcp's error recovery fixes things.  On the
> other end you will only see one instance come out.  Then you
> have to figure out which hardware interface dropped the cipe
> packet between them.

I posted some tcpdump files to the list yesterday.  I had sent them to 
Allan earlier.

> 
> D) Wild guess here - I haven't checked but it is possible that
> samba sets the DF (don't fragment) bit where ftp doesn't (Microsoft
> likes to do that for no particular reason...).  That means if
> the medium can't accept the whole packet it has to discard it
> instead of fragmenting so it will be rebuilt at the other end.

Shouldn't I be seeing some ICMP messages about that?  Also, I tried 
using some small MTU values and it didn't help.  Also, I looked at some 
of the packets captured with Ethereal.  They don't show the DF flag set.





--
Message sent by the [email protected] mailing list.
Unsubscribe: mail [email protected], "unsubscribe cipe-l" in body
Other commands available with "help" in body to the same address.
CIPE info and list archive: <URL:http://sites.inka.de/~bigred/devel/cipe.html>