[e2e] TCP spoofing in overlay networks

[David Borman]

It's been done and shipping for several years with all Cray X1 systems. 
  The CNS (Cray Network Server) proxies TCP connections between the Cray 
and the outside world.  This allows the Cray <-> CNS connection to use 
64K MTUs and larger TCP windows over the fibre channel connection, and 
the CNS then deals with all the small 1500 byte packets coming from the 
outside world.  Yes, this does break the end-to-end model.  You have 
two TCP connections, one between the Cray and the CNS, and another 
between the CNS and the remote host.  The CNS mainly passes data 
between the two endpoints, and uses NAT internally so to the Cray and 
the remote host, they think they are talking directly to each other, 
when  in reality they are both talking to the CNS.  Cray has done a 
good job over the years of making the CNS as transparent as possible.  
The performance benefit outweighs any issues of the corner cases that 
occasionally pop up.  You can find documentation on the CNS by going to 
the CrayDoc website:
	http://www.cray.com/cgi-bin/swpubs/craydoc30/craydoc.cgi
and searching for "CNS".

			-David Borman


On Mar 1, 2005, at 9:31 PM, Jonathan Shapiro wrote:

> I recently had occaision to read a few papers about the practice of 
> "TCP spoofing" over satellite links---i.e inserting a proxy prior to 
> the satellite link to provide TCP feedback to the sender, effectively 
> splitting into two TCP sessions connected in tandem. I was wondering 
> if anyone had ever proposed a similar idea to improve TCP throughput 
> in overlay networks over terestrial links.
>
> /jonathan shapiro