The Research Of DPDK-based Layer-3 Forwarding Technology

With the emergence of Intel QuickAssist Technology, PCI-E, QPI, etc., the I/O capacity of general purpose servers has a significant improvement. At the same time, the network adaptor's speed increases from 1000Mb/s to 10Gb/s as well as 40Gb/s. Even some research institutions and enterprises are trying to do some research on the 100Gb/s technology. The study of network applications with commodity servers become an advanced research hotspot. The current kernel-based solution for network packet processing has several bottlenecks, which limit the performance of servers. Considering this, Intel develops a tool called Data plane Development Kit to bypass the Linux kernel network stack so that the programmer can develop high performance network applications by using DPDK.This paper is based on DPDK and aims at designing a L3 forwarding system based on DPDK. At first this paper will introduce the background of why DPDK is developed and some researches related to DPDK from home and abroad in detail. Then we will analysis the Linux kernel network stack and find out some performance limiting factors on it. After that we introduce DPDK and analysis its architecture, and also introduce the features of DPDK. We will then implement the L3 forwarding system named D-L3 Forwarding for fast IPv4 packets forwarding. D-L3 Forwarding system implements multiple receive and send queues for each NIC port and bind each of them to a single logic core for parallel IPv4 packets processing. We also use SSE and batch processing to improve system's performance. After that we will analysis some critical factors that may have effect on system's performance.Finally we run D-L3 Forwarding system and do tests on it. The tests has shown that D-L3 Forwarding gets a very fast packet forwarding in case of 10Gb/s NIC. Tests also prove that NUMA can have a critical effect on system's performance, while huge page size and the number of memory channels can't influence system's performance.