Evaluating Network Processors in IP Forwarding

Abstract:

January 31, 2001: Updated Section 4.4.3 to correct arithmetic error
and include measurements for 128-byte packets.

This paper evaluates the performance of emerging network
processors---in particular, designs that employ multiple hardware
contexts to hide memory latency---in constructing IP routers. Such
processors are designed to forward minimum-sized IP packets at line
speeds, with the advantage (over ASIC-based solutions) of being
programmable. However, programming such network processors involves
two challenges. The first is how to effectively employ the multiple
contexts in a way that fully utilizes the memory bandwidth. The second
is how to allow the network processor to be programmed dynamically (so
it can support new functionality) without violating the processor's
tight timing constraints. This paper addresses both of these
challenges on a prototype board that uses the IXP1200 network
processor. We demonstrate that it is possible to support
8 x 100 Mbps ports with enough headroom to access up to
224 bytes of state information for each minimum-sized IP packet.