TCP implementations and false Time Out detection in OBS networks

This paper compares Reno, New-Reno and Selective Acknowledgements (SACK), the three most common TCP implementations today in (future) optical burst switched (OBS) networks. In general, SACK, which considers multiple Triple Duplicated ACKed (TD) losses in one round, is found to perform best in OBS networks, while New-Reno, which improves Reno in packet switched networks by fast retransmission in responding to partial ACKs, may however perform worse than Reno. All three TCP implementations react to a Time Out (TO) loss in the same way (i.e. using Slow Start). In OBS networks, where a burst may contain all packets from one round, and a burst loss occurs mainly due to contention instead of buffer overflow, such a TO event may no longer imply heavy congestion, or in other words, it may be a false TO or FTO. Such FTOs, which may be common in OBS networks especially for fast TCP flows, can significantly degrade the performance of all existing TCP implementations. Accordingly, we also propose a new TCP implementation called Burst TCP (BTCP) which can detect FTOs and react properly, and as a result, improve over the existing TCP implementations significantly.