Predictable Host Networking in Data Centers: Challenges, Techniques and Future Directions

Datacenter applications require robust and predictable performance guarantees from the underlying infrastructure, such as bandwidth, latency, and service time. The long-standing consensus is that the end-to-end performance is dominated by the network fabric. However, there is growing evidence that conventional host networking struggles to keep up with rapidly increasing data center access link bandwidths and ultra-low latency, making the host gradually become the bottleneck. Motivated by this trend, this article surveys recent advances in improving the predictability of host networking in data centers. We first analyze the challenges in achieving predictability at different layers of host networking. After that, we present recent advances targeting these challenges to improve predictable performance. Finally, we show the future directions and opportunities for further research.