Mapping packet processing applications on a systolic array network processor

Abstract

Systolic array network processors represent an effective alternative to ASICs for the design of multi-gigabit packet switching and forwarding devices because of their flexibility, high aggregate throughput and deterministic worst-case performances. However such advantages come at the expense of some limitations, given both by the specific characteristics of the pipelined architecture and by the lack of support for portable high-level languages in the software development tools, forcing software engineers to deal with low level aspects of the underlying hardware platform. In this paper we present a set of techniques that have been implemented in the Network Virtual Machine (NetVM) compiler infrastructure for mapping general layer 2-3 packet processing applications on the Xelerated X11 systolic-array network processor. In particular we demonstrate that our compiler is able to effectively exploit the available hardware resources and to generate code that is comparable to hand-written one, hence ensuring excellent throughput performances.