Pktpit: separating routing and packet transfer for fast and scalable software routers.

We propose an architecture to achieve fast and scalable software routers by separating the packet forwarding process into routing by the CPU and direct packet transfer over PCI Express (PCIe) between Network Interface Cards (NICs). Performance improvement of software routers in throughput and latency is demanded due to increased traffic volume and diversification of service forms such as Network Functions Virtualization. As a promising approach to accelerate software routers, offloading packet forwarding to NICs has been achieved by previous studies; however, the approach has a limitation that it cannot be used for IP forwarding with a large number of routes because of the insufficient amount of device memory in the NICs. The Pktpit architecture, which we propose, addresses this limitation by imposing Longest Prefix Matching (LPM) for routing on the CPU while achieving high throughput and low latency by offloading packet forwarding to NICs. The evaluation shows that the Pktpit architecture achieves 59.08% higher throughput and 2.76% lower latency than an implementation of current software routers while processing LPM with over 840,000 routes.