A New Model for TCP Packet Loss.
Daniel Genin and Kevin Mills
ITL/ANTD, NIST Gaithersburg, MD 20899.
TCP is the dominant protocol used in today's Internet to reliably transmit data between end-systems. Studies have shown that TCP packets account for over 80% of all Internet traffic by volume. The current version of TCP attempts to minimize congestion in and maximize utilization of a network by reducing transmission speed when packets are lost (due to buffer overflow) and increasing transmission speed when packets are delivered successfully. Because this congestion control mechanism is executed independently by all end-systems the resulting collective behavior is very hard to predict even for simple networks, let alone for the entire Internet with its thousands of routers and millions of end-systems. Nevertheless, stable and efficient operation of the Internet under increasing user and application demands depends upon thorough understanding of the network traffic dynamics.
Clearly, what is really needed are faithful and simultaneously computationally tractable mathematical models. Simulation, while permitting limited exploration of the network dynamics, cannot reveal the relationships between network parameters and quantities of interest such as throughput, packet loss, etc. In spite of the great efforts devoted to modeling network dynamics under TCP congestion control in the last 15 years the results are still very meager. For instance, there is still no mathematical model relating throughput of a TCP connection and such fundamental network parameters as bandwidth, delay, and buffer size. This unsatisfactory state of affairs is an indication of the magnitude and difficulty of the problem.
The bulk of research on the topic has focused on fluid approximation models. Fluid approximation models are continuous deterministic limits of discrete stochastic packet level models, applicable when the number of packets sent in a round trip (and hence the bandwidth-delay product) is large. A major stumbling point for the fluid approximation models is the stochastic nature of packet loss. While packet loss is sensitive to higher moments of the inter-arrival time distribution these statistics are not captured by the simplistic fluid approximation models. Consequently, since throughput of a TCP connection is essentially determined by packet loss the accuracy of equilibrium prediction of the resulting models is mediocre at best and their power of prediction of network dynamics is non-existent.
Our project focuses on understanding the relationship between packet loss and the fundamental network parameters -- bandwidth, delay and buffer size. By studying the statistics of the packet arrival process we derive a model relating packet loss to these parameters. In the future, our packet loss model should lead to more accurate models of TCP network dynamics by improving the accuracy of existing fluid approximation models.
Mentor's Name: Kevin Mills
Email: kmills at nist dot gov
Address: Bldg. 222, Rm. 218, MS8920
Division: ANTD, Lab: ITL
Author's Name: Daniel Genin
Email: dgenin at nist dot gov
Division: ANTD, Lab: ITL
Address: Bldg. 222, Rm. 214, MS 8920
Is your mentor a Sigma Xi Member? No