Identification of Architectural Bottlenecks in Programming Network Processors

Chidamber Kulkarni1, Christian Sauer2, and Matthias Gries3
(Professor Kurt Keutzer)
GSRC

Network processors exploit task and packet-level parallelism to achieve large throughputs. This has resulted in a huge diversity of architectures for similar applications [1]. From an ease of programmability perspective, these architectures can be classified along three main axes, namely in the type and organization of processing elements, different scheduling schemes between processing elements and other resources, like I/O interface buffers, and communication between different resources, like memory operations as well as bus configurations. Each of the above aspects significantly influence the ease of programmability.

A primary goal of this work is to implement one or more benchmarks on different network processors and understand the complex inter-play between different architectural components and programmability. In the process, we hope to define a broader abstraction for programming such heterogeneous application-specific multi-processors. Such an abstraction can be seen as a hardware abstraction layer across multiple network processors or as a starting point for exploring new heterogeneous ASIPs. It is important to note that this work differs from a typical programming model where the emphasis is on capturing application requirements as compared to ease of mapping or ease of porting across heterogeneous mutliprocessor ASIPs. In this work, we have currently implemented a 16-port IPv4 benchmark [1] on an Intel IXP1200 and Motorola C-PortC-5 and are currently investigating the different architectural influences for both of the implementations.

[1]
M. Tsai, C. Kulkarni, C. Sauer, N. Shah, and K. Keutzer, "A Benchmarking Methodology for Network Processors," Network Processors 2002: Design Principles and Practice, Morgan Kaufmann Publishers Inc., 2002.
1Postdoctoral Researcher
2Visiting Industrial Fellow, Infineon Technologies, CPR-ST, Munich
3Postdoctoral Researcher

More information (http://www.gigascale.org/mescal) or

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


Edit this abstract