Trading Packet Headers for Packet

In high speed networks, packet processing is relatively expensive while bandwidth is cheap. This begs the question: what elds can be added to packets to make packet processing easier? By exploring this question, we devise a number of novel mechanisms to speed up packet processing. With the advent of new standards for the Data Link, Network , and Transport layers, we believe there is an opportunity to apply these techniques to improve the performance of real protocols. First, we suggest adding a data manipulation header to an easily accessible portion of each packet. This header contains pointers to elds (in various layers) required for data manipulation. This information allows implementations to eeciently combine data manipulation steps (e.g., encryption and copying) in a structured fashion. Second, we suggest adding index elds to protocol identiiers at all layers (e.g.,, connection identiiers, network addresses, DSAPs) to reduce lookup costs and generic protocol processing. Several new ideas to utilize these index elds (threaded indexing, index passing , and source hashing) are proposed. Virtual Circuit Identiiers (VCIs) have been long used to simplify lookup and packet processing in virtual circuit. In source hashing and threaded indexing, the added indices essentially serve as VCIs, but for ows in a datagram network. In source hashing, for example, the \VCI" is a consistent random label chosen by the source. Our new methods provide the beneets of normal VCIs without requiring a round trip delay for set up. Our methods can lower worst case datagram lookup times from O(log(n)) to O(1), which may be important for Gigabit routers.