Research On Internet End-to-End Congestion Control Algorithms

Congestion control is one of the key techniques in computer networks. It is also an active research area. The end-to-end congestion control of TCP is used widely by Internet.Although TCP congestion control has already succeeded in preventing Internet collapse, many new window update algorithms, using convex increment mechanism under the framework of the traditional TCP, are suggested to solve the problem of network efficiency in the high bandwidth-delay product environments. But convex increment mechanism neither enhances network efficiency essentially, nor guarantees the network stability on the condition of the large window. Additionally, the design of these algorithms also neglects the impact of the network congestion on the small timescale.This dissertation investigates the impact on network stability and efficiency of general window update algorithms. Through detailed analysis, we draw the following conclusions. Firstly, the stability condition requests for small packet-level increment and large decrement speed around the equilibrium state. Secondly, concave increment algorithms improve both aggressiveness and network efficiency, while convex increment algorithms only improve aggressiveness. Finally, convex increment algorithms result in a large numbers of dropped packets, while concave increment algorithms can evidently reduce the number of dropped packets and unnecessary retransmit timeouts. This dissertation suggests a basic design idea of concave increment algorithms, which satisfy the stability condition around the large window. On one hand, when being smaller than efficient point, the congestion window grows up as quickly as possible, and on the other hand, when being close to efficient point, congestion window grows up slowly in order to linger around efficient point as long as possible.This dissertation investigates the network congestion on the small timescale. We conclude that the packet-sending mechanism of TCP is the underlying reason of network congestion on the small timescale, and that this congestion degree depends directly on RTT level increment of general window update algorithms. Additionally, neither limitation to back-back packets nor AQM has impact on small timescale congestion. This dissertation suggests a basic design idea, reducing the congestion level on small timescale: the slow start algorithm estimates the available buffer size, which limits the increment on RTT level.Based on the design ideas of concave increment and network congestion on small timescale, this dissertation proposes ETCP congestion control under the framework of the tradition TCP: ETCP's slow start uses AQI algorithm to estimate the available buffer size, and congestion avoidance uses improved PIMD algorithm whose RTT level increment is limited by the buffer size estimation. The theoretic analysis and simulation show that ETCP can obtain sound TCP compatibility and high network efficiency at the cost of a small quantity of dropped packets in various cable networks.In economic views, this dissertation proposes a design method of the utility-based rate control algorithms, which customizes the specific rate control algorithm by selecting utility function. We conclude that the network system, where the utility-based rate control algorithms are configured, will converge to an equilibrium state, which is the optimal solution of the network's optimization problem. This dissertation analyses stable condition of the utility-based rate control algorithms, and validates network performances of the two rate control algorithms respectively according to negative reciprocal and logarithm functions.