Packet Train Model: Optimizing Network Data Transfer Performance
University of Wisconsin-Madison Department of Computer Sciences
MetadataShow full item record
Network transmission bandwidth of 100 megabits per second is now available for LANs and bandwidth over gigabits per second will soon be available for WANs. A growing number of communication applications demand the support of high speed bulk data transfers. Unfortunately, effective user level data throughput has been far below the network bandwidth due to the bottleneck effect of packet software processing. This thesis presents a new approach to reducing packet processing overhead, in particular, reducing the number of invocations of host system support and protocol processing routines. The packet train model, based on observed network traffic burstiness, provides a framework in which a large block of data can be transferred as a large packet without incurring various problems, and a proper selection of transport layer functions can be offloaded to the network front-end without overloading the front-end processor. A packet train is a special form of large data delivery over a network. On a transmission medium, a train is a sequence of consecutive packets, spaced within a maximum allowed inter-packet gap and loaded with data for the same transport layer entity. The size of a train is dynamically determined by the amount of data packetized for transmission and by interruptions of urgent data transmitted onto the medium. At data sending and receiving hosts above the medium level, a packet train is handled as one transfer unit for its protocol processing. The thesis presents measurement results which provide motivations for the packet train approach. It describes in detail designs for the major packet train mechanism components, both hardware and software. Results of a simulation evaluation of the packet train scheme vs. the traditional non-train scheme indicate that the packet train is a viable approach to improving bulk data transfer throughput under a wide range of system parameters and traffic load conditions. Finally, the thesis discusses extensions needed for the packet train approach to benefit data transfer performance in a WAN environment.