A Programming Model for Network Processors

Niraj Shah and William Plishker
(Professor Kurt Keutzer)

The past five years have witnessed over 30 attempts at programmable solutions for packet processing. With these architectures, network processor designers have employed a large variety of hardware techniques to accelerate packet processing, including parallel processing, special-purpose hardware, memory architectures, on-chip communication mechanisms, and use of peripherals [1]. However, despite this architectural innovation, relatively little effort has been made to make these architectures easily programmable. The current practice of program network processors is to use assembly language or a subset of C. This low-level programming language places a large burden on the programmer to understand fine details of the architecture just to implement a packet processing application, let alone optimize it. We believe the programmer should be presented with an abstraction of the underlying hardware, or programming model, which exposes just enough detail to write efficient code for that platform. This work focuses on a programming model for network processors. Our approach starts with an abstraction based on Click, a domain-specific language for packet processing. We've demonstrated the promise of our programming model by implementing an IPv4 router on the Intel IXP1200, a common network processor.

[1]
N. Shah, "Understanding Network Processors," master's thesis, UC Berkeley, September 2001.

More information (http://www-cad.eecs.berkeley.edu/~niraj) or

Send mail to the author : (niraj@eecs.berkeley.edu)


Edit this abstract