LIGHT: A Compatible, high-performance and scalable user-level network stack

As the number of CPU cores and the speed of Ethernet NICs keep increasing on server machines, the network stack in the kernel has become the bottleneck for applications demanding very high throughput and ultra -low latency. Recently there is a trend towards moving the network stack out of the kernel. However, most kernel-bypass network stacks discard POSIX APIs that legacy applications have been built on, and the intricate work of transplanting applications brings the barrier to real-world deployment of kernel-bypass stacks. In this work, we propose LIGHT, a novel user-level network stack, which not only gains highly scalable performance on multi-core server, but also achieves compatibility with legacy applications. For compatibility, LIGHT realizes efficient blocking APIs in the user space, intercepts network-related APIs in a non-intrusive manner, and uses the FD space separation technique for proper API redirection. For high performance and scalability, LIGHT adopts lock-free shared queue based inter-process communication and full connection affinity to reduce overheads of system call, cache miss, etc. Experiments demonstrate that many types of legacy applications could run on LIGHT without modifying their source code. Compared with the latest kernel stack, Nginx on LIGHT achieves up to 2.86x throughput and 78.2% lower tail latency (99.9th percentile) with 14 CPU cores.

Automated summary

With the increase in CPU cores and Ethernet NICs on server machines, the network stack in the kernel has become a bottleneck for high throughput and low latency applications. Moving the network stack out of the kernel poses challenges due to the lack of POSIX API compatibility and the complexity of application transplantation. This work proposes LIGHT, a user-level network stack that achieves both high scalability and compatibility with legacy applications. LIGHT adopts efficient blocking APIs, intercepts network-related APIs, and utilizes FD space separation technique for API redirection. It also utilizes lock-free shared queue based inter-process communication and full connection affinity for high performance. Experiments show that legacy applications can run on LIGHT without modifying source code, and Nginx on LIGHT outperforms the latest kernel stack with significantly higher throughput and lower tail latency.