Parcae: Proactive, Liveput-Optimized DNN Training on Preemptible Instances
Offered By: USENIX via YouTube
Course Description
Overview
Explore a groundbreaking approach to cost-effective deep neural network (DNN) training in this 16-minute conference talk from NSDI '24. Delve into Parcae, an innovative system that leverages preemptible cloud instances to significantly reduce training costs for large DNNs. Learn how Parcae's proactive strategy optimizes 'liveput,' a novel metric combining throughput and robustness, to adapt to predicted resource changes before instance preemptions occur. Discover the system's key features, including lightweight instance migration and an availability predictor, which enable it to outperform existing spot-instance DNN training systems by up to 10 times. Gain insights into Parcae's ability to achieve near-optimal performance for training large DNNs under frequent preemptions, a scenario where current approaches struggle to make progress.
Syllabus
NSDI '24 - Parcae: Proactive, Liveput-Optimized DNN Training on Preemptible Instances
Taught by
USENIX
Related Courses
Software as a ServiceUniversity of California, Berkeley via Coursera Software Defined Networking
Georgia Institute of Technology via Coursera Pattern-Oriented Software Architectures: Programming Mobile Services for Android Handheld Systems
Vanderbilt University via Coursera Web-Technologien
openHPI Données et services numériques, dans le nuage et ailleurs
Certificat informatique et internet via France Université Numerique