Latency Optimization in Datacenters using Adaptive Transport and Reliable Training

Datacenters perform the core computation for a wide range of large-scale applications. These applications include online data-intensive services such as web search that are user-facing and must meet stringent latency constraints, as well as distributed DNN training of sophisticated models that are time-consuming tasks. In this dissertation, we first address the performance challenges with transport protocols to ensure data from latency-sensitive applications is transferred efficiently across the datacenter network. Then we analyze the contention of different traffic patterns for switch buffer resources and propose a scheme that resolves the inherent tension between good burst absorption and high utilization. Finally, we study datacenter failures that impede the progress of distributed training jobs and waste many hours of computing resources, and to reduce the recovery time, we present a new optimization-based framework for robust gradient aggregation that allows the training task to continue in the presence of failures.