Reinforcement Learning for Datacenter Congestion Control

We approach the task of network congestion control in datacenters using Reinforcement Learning (RL). Successful algorithms can dramatically improve latency and overall throughput. Until today, no such learning-based have shown practical potential this domain. Evidently, most popular recent deployments rely on rule-based heuristics that are tested a predetermined set benchmarks. Consequently, these do not generalize well to newly-seen scenarios. Contrarily, we devise an RL-based algorithm with aim generalizing different configurations real-world datacenter networks. overcome challenges as partial-observability, nonstationarity, multi-objectiveness. further propose policy gradient leverages analytical structure reward function approximate its derivative stability. show scheme outperforms alternative RL approaches, generalizes scenarios were seen during training. Our experiments, conducted realistic simulator emulates communication networks' behavior, exhibit improved performance concurrently multiple considered metrics compared deployed today real datacenters. is being productized replace some largest world.