Improving Throughput in High Bandwidth-Delay Product Networks with Random Packet Losses

It is well known that the standard TCP has become a performance bottleneck in the networks with large bandwidth-delay products. The situation gets worse if high-speed wireless links are part of the networks, due to the frequent random losses over the wireless links. This is because the standard TCP increments its congestion window too slowly in the absence of packet losses and decrements it too drastically in response to packet losses. A natural solution is to make TCP more aggressive. This approach has been exercised in the recent TCP development such as HSTCP and MulTCP which have the capability of using a single logical connection to emulate the behaviour of a set of multiple standard TCP connections. In the meantime, TCP Parallelisation uses a set of parallel TCP connections to transfer data for an application process. Then, a question arises - can the single-connection based approach achieve the similar performance as TCP Parallelisation in the environments where random packet losses prevail. Our analysis shows that TCP Parallelisation has the better performance and is more efficient for performance improvement.