One Size Fits Not All: Adapting Network Transport to Modern Workloads by Dr Ayush Mishra

Abstract

Computer networks are the backbone of modern computing, yet the fundamental primitives of network transport were designed for a vastly different era. As we shift from simple web browsing to supporting diverse applications on the Internet and large-scale distributed ML jobs in data centers, longstanding assumptions in protocol design are beginning to break. In this talk, I will argue that the "one size fits all" approach to transport is no longer viable and outline my vision for how modern transport stacks need to adapt.

First, I will discuss how application heterogeneity on the Internet is forcing a re-evaluation of flow-level fairness. I will present a framework that moves beyond traditional throughput-based fairness to accommodate the diverse throughput-delay tradeoffs required by modern traffic. Second, I will demonstrate how ML workloads are fundamentally incompatible with statistical multiplexing - a core tenet of networking. I will describe our on-going work on new transport and load-balancing mechanisms that handle the rigid, bursty patterns of ML training. I will conclude by outlining how modern applications dictate the next grand challenges in network transport - both on the Internet and within data centers.