This paper argues that congestion control protocols in machine learning datacenters sit at a sweet spot between centralized and distributed flow scheduling solutions. We present MLTCP, a technique to augment today's congestion control algorithms to approximate an interleaved centralized flow schedule. At the heart of MLTCP lies a straight-forward principle based on a key conceptual insight: by scaling the congestion window size (or sending rate) based on the number of bytes sent at each iteration, MLTCP flows eventually converge into a schedule that reduces network contention. We demonstrate that MLTCP uses a gradient descent trend with a step taken at every training (or fine-tuning) iteration towards reducing network congestion among competing jobs.

This paper argues that fair-sharing, the holy grail of congestion control algorithms for decades, is not necessarily a desirable property in Machine Learning (ML) training clusters. We demonstrate that for a specific combination of jobs, introducing unfairness improves the training time for all competing jobs. We call this specific combination of jobs compatible and define the compatibility criterion using a novel geometric abstraction. Our abstraction rolls time around a circle and rotates the communication phases of jobs to identify fully compatible jobs. Using this abstraction, we demonstrate up to 1.3× improvement in the average training iteration time of popular ML models. We advocate that resource management algorithms should take job compatibility on network links into account. We then propose three directions to ameliorate the impact of network congestion in ML training clusters: (i) an adaptively unfair congestion control scheme, (ii) priority queues on switches, and (iii) precise flow scheduling.