Parallel Processing of Packets with a PRAM

Abstract

The operation of policy tables, for access control (firewalls) and for routing (flow tables), is a major component in networks. For example, one of the most important bottlenecks in the Internet is the time taken by backbone routers to resolve packets. In practice, such routers use special hardware such as Ternary Content-Addressable Memory. How well do standard architectures for parallel computing work to speed up such operations? At first glance, it appears that firstmatch policies such as firewalls might force a serial approach, where the only mechanism for speeding up packet processing is to handle different packets on different cores. However, we show in this paper that the operation of packet-processing policies for each individual packet can in fact be sped up with a PRAM parallel computer. We go on to demonstrate how to build a policy engine for XMT (Explicit Multi Threading), a practical near-PRAM architecture and instruction set, and present some performance results.