fastos/fastsocket

Fastsocket is a kernel-level modification for Linux that makes the TCP networking stack scale more efficiently on multi-core servers. It was developed at SINA.COM, a Chinese internet company, in cooperation with Tsinghua University and with support from Intel. Research describing the system was published at ASPLOS 2016, a computer systems conference. The problem Fastsocket addresses is that the standard Linux kernel does not scale well for network-heavy workloads once a server has more than about four CPU cores. At 12 or more cores, the stock kernel can see throughput drop as more cores are added, because of contention around shared networking data structures. Fastsocket redesigns the socket layer so that each CPU core has its own independent path through the network stack, achieving close to linear scaling up to 24 cores. In benchmark tests, Fastsocket increased the request-handling throughput of Nginx (a web server) by 290% and HAProxy (a load balancer) by 620% on a 24-core machine compared to the base CentOS 6.5 kernel. On servers with Intel NICs supporting Flow Director, a hardware feature that steers network packets to specific CPU queues, HAProxy throughput improved by an additional 15% in proxy workloads. Enabling hyper-threading added another 20% on top of that. Fastsocket works as a custom Linux kernel combined with a small preload library. Existing applications require no code changes: a program like Nginx can use Fastsocket simply by starting with the library loaded via the LD_PRELOAD environment variable. Removing the library returns the application to standard kernel behavior, making rollback immediate. The implementation targets CentOS 6.5. The system had already been running in production at SINA.COM for HTTP load balancing since March 2014. It is released under GPLv2.