Hydro - Surrogate-Based Hyperparameter Tuning Service in Datacenters
Explore a conference talk from OSDI '23 that introduces Hydro, an innovative surrogate-based hyperparameter tuning service designed for datacenters. Learn about the challenges of hyperparameter tuning for large-scale deep learning models and how Hydro addresses these issues through job-level and cluster-level optimizations. Discover the two key components of Hydro: the Tuner, which generates and optimizes surrogate models, and the Coordinator, which improves tuning efficiency and resource utilization. Gain insights into the experimental results that demonstrate Hydro's ability to significantly reduce tuning makespan without compromising quality. Access the source code and delve into the potential impact of this technology on deep learning model development and datacenter resource management.