Measurement Method Study Of TCP Throughput

The TCP transport protocol is one of two important transport protocols in TCP/IP protocol stack. A significant amount of today's Internet traffic is carried by TCP. As an important end-to-end internet performance parameter, TCP throughput is an important index of study and analyzing TCP behavior and performance. Measuring TCP throughput will be helpful for adjusting QoS at applicational layer so as to improve the quality of services. Besides that, the study on Measuring TCP throughput will also promote the studies of QoS routing and end-to-end multicast congestion control protocol.Some methods and tools have been developed to measure TCP throughput. However, these tools will produce large measurement traffic. If being used incorrectly, these methods will have large impact on network. So they are not suitable for high speed network or on-line measurement environment. In this thesis, we introduce and analyze relevant measurement methods and survey tools in detail. We also distinguish such terms extremely apt to obscure as the bandwidth and throughput, etc. To avoid disadvantages of direct measurement method, a new method is proposed, which measures TCP throughput indirectly with a TCP throughput analysis model. Based on detailed analysis for some TCP throughput analysis models, this dissertation uses the Gilbert two-state model to represent the packet lost behavior of end-to-end internet path and builds a model for TCP congestion control process according to TCP congestion control mechanism. Furthermore, a new measurement algorithm is presented based on the proposed analysis model. The experiments of NS2 Simulation indicate that the new model can measure TCP throughput and its change trend more precisely than those old models. The proposed model in this paper has three characteristics as following: (1) It doesn't need the cooperation of destination nodes. The algorithm only requires source nodes to measure some metrics of end-to-end internet path, such as RTT, packet loss rate, maximum announced window by receivers, MSS and so on. So the measurement only needs to run the measurement tool on source nodes, and needn't the cooperation of destination nodes. (2) It's suitable for IP networks with various banswidth. Because this algorithm uses the analysis model based on TCP congestion control mechanism to measure TCP throughput indirectly, it need not to send many measurement data packets. By capturing the ACK packets in the TCP connectionestablished phase, the algorithm can support all kinds of TCP windows announced by receivers in different bandwith IP network. (3) The algorithm is also suitable for a long period on-line measurement. Due to only measuring some end-to-end internet path parameters such as RTT and packet loss rate, it doesn't produce large measurement traffic. So it has little impact on network and can support on-line measurement in a long period.Besides these, we have developed a large-scale distributed network measurement and analysis system (NetTurbo).This system integrats active mearsurement and passive measurement techonlogies. It uses a frame of distributed collection and centralized analysis to support traffic and performace analysis of 2-7 layer in various time scales. And a TCP throughput measurement tool using the proposed algorithm has been realized for the system.