Understanding the Packet Processing Capability of Multi-Core Servers

Compared to specialized network equipment, software routers running on commodity servers allow programmers to rapidly build and (re)program networks using the software and hardware platforms they tend to be most familiar with—that of the general-purpose computer. Unfortunately, the Achilles’ heel of software routers has been performance; commodity servers have traditionally proven incapable of high-speed packet processing thereby motivating an entire industry around the development of specialized network hardware and software. However, recent advances in PC technology promise significant speed-ups for applications amenable to parallelization; router workloads appear ideally suited to exploit these advances. This leads us to question whether it is now (or soon will be) plausible to scale software routers to current high-speed networks. As a first step toward answering this question, we study the packetprocessing capability of current commodity multi-core servers: we identify performance bottlenecks, evaluate tradeoffs between performance and programmability, and discuss what changes are needed to further scale the packet-processing capability of generalpurpose servers.