Adaptive protocols for reliable Internet data transport.

摘要

The explosive growth of the Internet, and the advent of mobile computing, are leading to a wide-area heterogeneous Internet. In this thesis, a systematic approach is proposed for the design of end-to-end data transport mechanisms that can operate in robust fashion over such networks. Two key features of the network that are accounted for in the framework developed here are large bandwidth-delay products and noncongestion-related loss. Two broad classes of applications are considered: throughput-sensitive applications and delay-sensitive applications. The central contributions of the thesis are as follows: (a) design of methods for providing feedback regarding congestion control and error recovery in a manner that is robust to loss of acknowledgment packets, and (b) decoupling of the tasks of congestion control and error recovery, allowing robust operation in the presence of random loss of data packets. This allows application of the same congestion control principles for throughput- and delay-sensitive applications.; For throughput-sensitive applications, information sufficient for congestion control is derived from a new acknowledgment format, termed total acknowledgment (TACK), while error recovery is based on repetition-coded selective acknowledgment (SACK). The resulting acknowledgment (ACK) format is remarkably resilient to ACK loss, and is used to develop two experimental protocols. The GTACK protocol works well in a variety of situations using simulations and is a good candidate for experimental deployment over the Internet. Unlike TCP, GTACK is robust to noncongestion loss, but it shares network resources fairly with TCP in regimes of low random loss.; For delay-sensitive applications, two complementary sets of results are obtained. The first set of results are regarding the Redundancy Estimation Protocol (REP), which provides adaptive error recovery using a forward error control (FEC) scheme in which redundancy rate is adjusted, in an attempt to maintain a desired quality of service (QoS), based on feed-back from the receiver. The second set of results are regarding the delay-sensitive TACK (DSTACK) protocol, which is a modification of GTACK that provides adaptive congestion control for delay-sensitive applications. Integration of REP and DSTACK to obtain a protocol that provides adaptive error recovery and congestion control is left as a topic for future work.