New Techniques for Making TCP Robust to Corruption-Based Loss

		  Wesley M. Eddy, Shawn Ostermann
			  Ohio University
		 {weddy,ostermann}@irg.cs.ohiou.edu

			    Mark Allman
	   BBN Technologies / NASA Glenn Research Center
			  mallman@bbn.com


One of the longstanding problems with using the Transmission Control Protocol
(TCP) in space-based networks is the prevalence of packet loss due to
non-negligible bit-error rates.  TCP's congestion control algorithms are based
on the assumption that packet losses are caused by the aggregate load being
placed on a network exceeding its capacity.  Therefore, when TCP detects packet
loss, the protocol reduces its sending rate in an attempt to relieve the
congestion on the network.  In wireless networks (and, specifically space-based
networks), packet loss is not necessarily a signal that the network path is
congested.  Therefore, corruption-based losses cause TCP to needlessly reduce
the sending rate.

Several classes of mitigations to the problems of non-congestion- based losses
have been suggested in the literature.  One possible technique is for a gateway
to inform one of the endpoints of a connection when its packets are dropped due
to corruption.  However, if the host and/or port addresses lie among the
corrupted bits, this may not be possible.  Another class of solutions employs
smart middleboxes on either side of a wireless link.  These boxes attempt to
isolate the wireless portions of the path and locally buffer packets to
transparently recover from corruption.  Such middleboxes sometimes break the
end-to-end model of TCP and may be rendered useless in the face of encrypted
traffic (e.g., if IPsec is employed).

The scheme we will present in this talk is different from the schemes
previously researched.  The scheme, known as Cumulative Explicit Transport
Error Notification (CETEN), was originally introduced in [1].  In this talk, we
present a system that shares the spirit of [1], but significantly differs in
its approach.

CETEN allows routers along a network path to update a field in the packet
header based on the error rate of the directly connected link on which the
packet arrived.  When a packet reaches its destination, the probability of any
packet being lost due to corruption along the path can be determined by
examining the field that was updated as the packet passed through the network.
The data receiver echoes this information back to the sender.  Meanwhile, the
sender uses an algorithm developed for this project, called LEAST, to assess
the overall loss rate experienced by the connection.  By examining the overall
loss rate and the aggregate corruption rate, the TCP sender can determine the
fraction of its traffic being lost due to congestion and the fraction being
lost due to corruption.  Armed with this information, the TCP sender adjusts
the factor by which it reduces the congestion window on each loss event.  If
all loss is caused by congestion, the factor is 0.5 (the IETF's currently
specified factor for TCP).  On the other hand, if all loss is due to
corruption, the factor is set to 1.0, meaning that the TCP sender will repair
loss without reducing its sending rate.  Given a mix of congestion and
corruption loss, the factor will be between 0.5 and 1.0 (based on the ratio of
congestion loss to corruption loss).

In the talk we will outline the CETEN framework for dealing with
corruption-based loss.  Additionally, we will show preliminary simulation
results of the CETEN system and outline the next steps in this area of
research.


References

[1] Rajesh Krishnan, Mark Allman, Craig Partridge, James
    P.G. Sterbenz.  Explicit Transport Error Notification (ETEN) for
    Error-Prone Wireless and Satellite Networks.  Technical Report
    TR-8333, BBN Technologies, March 2002.