Trading packet headers for packet processing

Reviewer: Robert Bruce McLaughlin

It has long been known that one can speed the processing time of a packet by a node by adding information to the packet header. Until recently, however, bandwidth was more expensive than the processing power of the node. As the bandwidth of links increases, node processing is becoming the more limited resource. As a result, it seems logical to add bits to the packet to speed up node processing. These bits could be added for each layer that processes the packet—all the OSI layers, for example—greatly speeding the packet on its way. In this paper, the authors propose additions to the packet header, namely a source hashing scheme, threaded indices, and a data manipulation layer (DML). Source hashing is similar to hashing in a database. A somewhat randomized pointer is created by the source and translated toward the destination by the interceding nodes. It does not require a path setup, so it is connectionless in nature. Threaded indices have a per-hop index for each source-destination pair, or they know the route the packet will take through the network. The idea of DML is to allow extensions of the data processing that can be done at each layer, such as encryption. The problem with any packet routing network is knowledge of paths. It is always likely that the source will know the destination. It is unlikely, especially in a connectionless network, that the source will ever know the path to the destination. Often the nodes know only the next closest node to the destination. In an ideal network, all the nodes would know all the paths at all times. Generally, the overhead for such information is considered too expensive in light of its being time-delayed. It is not clear that, in such networks, either the threaded indexing or the DML schemes proposed would improve performance. The proposed hashing scheme would improve performance in connectionless networks. I think the authors envision the use of DML and threaded indices in a connection-oriented or virtual connection-oriented network such as asynchronous transfer mode (ATM) permanent virtual circuit s. In that context, they would have value, assuming that the network is such that a circuit can be reserved at call setup, not a trivial problem in a large ATM network. Understanding what path to take through a connectionless network at any given instance is not easy. People who are trying to combine the statistical gain of a connectionless network with the services that depend on a guaranteed connection are finding that this is a hard problem. The authors may have an interesting piece of the solution.