A throughput deadlock-free TCP for high-speed Internet

Abstract

Throughput deadlocks were observed when TCP was operated on high-speed networks. This deadlock problem is caused by the interaction of the sender-side and receiver-side silly window syndrome avoidance algorithms, because a TCP connection's maximum segment size is no longer small on the high-speed Internet when compared with the send and receive socket buffer sizes. We propose a new congestion-sensitive, adaptive acknowledgment algorithm (CS-AAA) to solve the deadlock problem. Unlike our previously proposed AAA, the CS-AAA is able to respond to and to recover from congestion much faster than AAA. The CS-AAA solves this problem by detecting congestion, and performing a slow-start-like mechanism. Extensive simulation results support that CS-AAA's throughput performance significantly exceeds that of AAA, especially when the send buffer size is relatively large.