We proposed an enhancement to TCP’s error recovery scheme, which we call the 
Eifel algo-rithm. It uses extra information in the TCP header to eliminate the 
problems caused by compet-ing error recovery. Our current implementation is 
based on the TCP timestamp option, and only requires changes to the TCP sender 
implementation. It does not require changes to the TCP receiver code nor to the 
protocol itself. Thus, given this backwards compatibility and the fact that it 
does not change TCP’s congestion control semantics, the new algorithm can be 
incrementally deployed.

In Chapter 4, we showed that the end-to-end performance of fully-reliable flows, 
such as those based on TCP, can only be optimized by running highly persistent 
link layer error recovery. The one missing piece, however, was a solution for 
situations where the wireless connectivity is intermittent, i.e., situations 
where spurious timeouts are likely to occur. Frequent disconnec-tions - on the 
order of seconds - without losing data are common in packet-radio networks. In 
such environments, the algorithm can improve the end-to-end throughput by 
several tens of percent, although we show that an exact quantification is highly 
dependent on the path charac-teristics. Thus, with the Eifel algorithm 
implemented in TCP and the implementation of a flow-adaptive wireless link, the 
long standing problem of "TCP over lossy links" is eliminated.

In addition, we have proposed a new retransmission timer for TCP, which we call 
the Eifel retransmission timer, that can also be incrementally deployed. It 
eliminates four major prob-lems of TCP-Lite’s retransmission timer which have 
revealed in our work. We demonstrated that the Eifel retransmission timer is a 
more precise predictor of an upper bound for the path’s RTT while reacting 
quicker to packet losses. We showed that this can increase the end-to-end 
throughput by more than 30 percent. As another alternative, we proposed an 
advanced version that becomes increasingly optimistic while adapting to the 
measured fraction of spurious time-outs. This requires the Eifel algorithm that 
opened the door to the development of a more opti-mistic retransmission timer 
because the Eifel algorithm ensures that the penalty for underesti-mating the 
RTT is minimal. In the common case, the only penalty is a single spurious 
retrans-mission. Although we studied retransmission timers in the context of 
TCP, we believe that the design principles we proposed are applicable to other 
reliable end-to-end, and link layer proto-cols.

The strength of our work related to end-to-end retransmission timers lies in its 
hybrid analysis methodology explained in Section 3.4. We developed models of 
each retransmission timer for the class of network-limited TCP bulk data 
transfers in steady state. With that model we were able to predict the problems 
of TCP-Lite’s definition of the RTO. We also used that model to develop a new 
RTO for the Eifel retransmission timer. We then validated the correctness our 
model-based analysis through measurements in a real network that yielded the 
same results.