Robust Data Compression of Network Packets

This paper describes an approach for compressing data packets that enables inter-packet compression without the drawback of multiplying the effect of packet loss. By adding an acknowledgment scheme, the sender can limit the history state used by the compression algorithm to those packets that have been correctly received. A vector identifying the packets used as history is included in the compressed packet, enabling the receiver to reconstruct the history state required to decompress the packet. This approach improves the compression achieved compared to stateless schemes, while retaining robustness in the face of packet loss or reordering. The paper reports several simulation experiments using our approach. We first describe an implementation based on the publicly available zlib implementation of the popular deflate compression format. We then describe the implementation of a Lempel-Ziv ’77 variant called thwack that is more efficient at handling the unpredictable history state used to compress and decompress packets. Introduction We are interested in improving the performance of packet networks. When available computational resources are large compared to the network bandwidth, it can be beneficial to compress data packets before transmitting them. We must consider both the speed of a compression algorithm and its ability to compress. Too slow an algorithm may reduce performance, while insufficient compression limits any potential gain. As an additional complication, many networks are unreliable -they drop or reorder packets. The Internet is an example of such a network. If compression introduces dependencies between packets, it may be impossible for the receiver to decompress a packet when a previous packet is lost. Existing solutions to this problem include: 1) Make the network reliable. Even unreliable networks typically provide a reliable endto-end transport service; on the Internet TCP is such a service. Compressing packets at the transport level is feasible, but must be applied end-to-end, and often requires cooperation by the application. Also, this approach is not appropriate for transparent compression over a link in the middle of the network, a case in which we are interested. 2) Stateless compression: Compress each packet independently. Since each packet is independent, it can be always be decompressed by the receiver. Unfortunately this independence fundamentally interferes with the effectiveness of compression. Compression is achieved by finding the commonality of between various parts of the data and exploiting this commonality to send less information. Stateless compression only examines the data in a single packet, which reduces the compression ratio. In particular,