Dynamic Resource Management

L2S is a learning to search framework that automatically discovers high quality power managment policies for large manycore systems. L2S generalizes and goes beyond today’s manually designed non-ML methods such as linux governors for resource management by exploring power saving opportunities at minimal performance overhead across different HPC applications.

Learning to Search (L2S) framework provides the following key features:

• Automated Learning and Search Process. Bayesian Optimization enabled automated search process for finding optimal Dynamic power management policies.

Publication

• Gaurav Narang, Aryan Deshwal, Michael Kishinevsky, Raid Ayoub, Jana Doppa, Partha Pande DYNAMIC POWER MANAGEMENT IN LARGE MANYCORE SYSTEMS: A LEARNING-TO-SEARCH FRAMEWORK. Accepted in TODAES, June 2023.

• Gaurav Narang, Michael Kishinevsky, Raid Ayoub, Jana Doppa, Partha Pande [Uncertainty-aware Online Learning for Dynamic Power Management in Large Manycore systems]. Accepted in ISLPED, Vienna, Austria, Aug 2023.