Analysis of Linux UDP Sockets Concurrent Performance

Almost all DNS queries that traverse Internet are transported via UDP in self-contained small packages. Therefore, with no restriction of packet ordering, the intuition would say that adding thread-based parallelism to the servers will increase their performance, but it does not. This paper studies the problem of serialized access to UDP sockets, and states the problem in the way the packets are enqueued in the socket at kernel level, which introduces high levels of contention in synchronization primitives when using threads. As a naive solution, we present a multi-queue receiver network stack, which improves the performance of processing small UDP packages when multiple threads read from the same socket.