Intelligent Node-Level Locality-Aware Load Balancing and Optimizations with Charm++'s CkLoop
DescriptionIn this talk, we discuss novel load balancers in Charm++/CharmMPI and the associated node-level library called CkLoop. We also will discuss how the design of these load balancers carefully re-uses intelligent heuristics from inter-node load balancers of Charm++. Furthermore, we will show how inter- and intra-node load balancers can be synergized: for example, the frequency of Charm++/CharmMPI’s inter-node load balancing can be tuned in conjunction with more frequent node-level load balancing in a hierarchical scheme. Other techniques include selectively turning cores off to optimize performance and energy simultaneously, and automatically-sized teams to trade-off locality and inter-node communication frequency. Such integrated load balancing strategies bring down overall costs of load balancing for simulations for science and engineering run on modern supercomputers, and thus reduce time-to-solution for such simulations.
TimeWednesday, June 2814:30 - 15:00 CEST
Computer Science, Machine Learning, and Applied Mathematics