7 Chapter - Packet processing on a SIMD stream processor

Abstract

This chapter explores the application of stream architectures to packet processing tasks, IPv4 Forwarding and AES encryption in particular. Both applications were run on generic stream architecture (Imagine), and experiments were conducted to characterize the performance of both applications for different configurations of this architecture. For a system clock of 500 MHz, the throughput of the AES encryption in ECB mode varies from 2.02 Gb/s (96-block packets) to 0.8 Gb/s (8-block packets). The IPv4 Forwarding application, with a configuration of one Imagine with 8 clusters, delivered a worst-case performance of around 67 cycles per packet, for a packet trace constructed from the MAE-WEST routing table. Hence, the forwarding engine was able to support packet traffic coming at a rate of OC-48 assuming a clock frequency of 500 MHz. The LRF and memory bandwidth characteristics of the two packet processing applications confirm that they are at two different ends of the application spectrum, with one being memory-intensive and the other being computation intensive. The low SRF characteristics for the two packet processing applications, compared to the Depth media application, is due to the fact that the processing of the packets for both applications is done primarily in one kernel. This results in reduced trips to the SRF between kernels, decreasing the SRF bandwidth utilization. The packet processing applications have comparable metrics to that of media applications indicating that this architecture could be as useful for network applications as it is for media applications.