Low-Level HPC System Software for Task Parallelism
DescriptionThis talk will discuss the process of enhancing node-level dynamic scheduling for MPI+X applications, expressed in higher-level models for task parallelism, such as OpenMP, and capable of targeting both CPU and GPU processors. The main focus of the talk will be on low-level HPC system software implemented in LLVM using a blend of compiler and runtime support for enablement of applications expressed in higher-level models. The LLVM tool chain for OpenMP, initially developed with a focus on CPU processors, has undergone significant extension in recent years to support GPU processors as well. The talk will also discuss current best practices for debugging and profiling applications that are executed on top of this low-level system software layer.
TimeWednesday, June 2815:30 - 16:00 CEST
Computer Science, Machine Learning, and Applied Mathematics