Evaluating the impact of network I/O on ultra-low delay packet switching

Low latency is a crucial requirement for demanding conferencing applications, such as Networked Music Performance (NMP), the collaboration of musicians in real time. Modern conferencing systems employ a Selective Forwarding Unit (SFU) to transparently duplicate and forward media streams between participants. Since an SFU does not process the media streams, so as to reduce delay, its latency is mainly determined by the underlying network I/O mechanism that moves packets to/from the network hardware and user space. Such mechanisms are usually based on POSIX sockets, which were not designed for high performance networking. We designed and implemented pktswitch, a minimal, socket-based SFU and measured its performance. We then modified pktswitch to employ netmap, a framework for fast packet I/O, to overcome the performance bottlenecks imposed by the socket-based design. The modified implementation handles packets in user space, with minimal kernel interaction. We describe and contrast the two implementations and then compare their performance in terms of packet processing overhead and delay. Our results show that the netmap-based implementation reduces packet processing overhead by 76% and delay by 89% compared to the socket-based implementation, thus allowing an SFU to host much higher loads (e.g. more users with more media streams and higher bit rates) without introducing delays.