PPB: a Path-based Packet Batcher to Accelerate Vector Packet Processor

Fast Data I/O (FD.io) is an open project to promote network processing based on generic hardware platforms, providing a feasible path to accelerate Network Function Virtualization (NFV). As the core of FD.io, vector packet processor (VPP) is a modularized and high-performance software framework for building network data plane applications. The core idea of VPP has two parts. Firstly, VPP processes a group of packets (packet vector) every time to reduce the instruction cache (i-cache) misses with vectorized processing model. Secondly, VPP decreases the data cache (d-cache) misses by prefetching the following packets. We observe that the received packet order has a significant impact on the performance of VPP. When the packets within a vector traverse different paths, the vector would be split into multiple small vectors. The frequency of i-cache and d-cache misses would increase.In this paper, we design a model to quantify the overhead in VPP. The key point to reduce the overhead is to decrease the number of processing paths in a vector. We propose a Path-based Packet Batcher (PPB) to accelerate VPP by adjusting the packet order. Before VPP processing packets, PPB batches the packets based on the processing paths they will traverse. We build a PPB prototype based on FPGA to evaluate the performance optimizations. The experimental results show that the reduction of i-cache and d-cache misses can be up to 57.6% and 17% respectively.