Improving TCP throughput performance on high-speed networks with a receiver-side adaptive acknowledgment algorithm

A drastic TCP performance degradation was reported when TCP is operated on the ATM networks. This deadlock problem is `caused' by the high speed provided by the ATM networks. Therefore this deadlock problem is shared by any high-speed networking technologies when TCP is run on them. The problems are caused by the interaction of the sender-side and receiver-side Silly Window Syndrome (SWS) avoidance algorithms because the network's Maximum Segment Size (MSS) is no longer small when compared with the sender and receiver socket buffer sizes. Here we propose a new Receiver-Side Adaptive Acknowledgment Algorithm (RSA 3) to eliminate the deadlock problems while maintaining the SWS avoidance mechanisms. Unlike the current delayed acknowledgment strategy, the RSA 3 does not rely on the exact value of MSS and the receiver's buffer size to determine the acknowledgment threshold. Instead the RSA 3 periodically probes the sender to estimate the maximum amount of data that can be sent without receiving acknowledgment from the receiver. The acknowledgment threshold is computed as 35% of the estimate. In this way, deadlock-free TCP transmission is guaranteed. Simulation studies have shown that the RSA 3 even improves the throughput performance in some non-deadlock regions. This is due to a quicker response taken by the RSA 3 receiver. We have also evaluated different acknowledgment thresholds (35%, 50%, 75%). It is found that the case of 35% gives the best performance when the sender and receiver buffer sizes are large.