Tunable Tasking in PARSEC for Numerical Linear Algebra Routines Used in HPC
DescriptionThis talk will show that a specialization of node-level load balancing for each numerical linear algebra routine of the SLATE library is needed due to intricacies of each routine. In particular, it will show an auto-tuned low-overhead loop scheduling strategy with customized data layouts speeds up dense matrix factorizations on a heterogeneous node by a factor of 1.6x and reduce performance variability across executions by up to 23.5%, without a significant impact to numerical error. The customized techniques are incorporated in the SLATE library and its associated PARSEC runtime.
TimeWednesday, June 2814:00 - 14:30 CEST
Event Type
Computer Science, Machine Learning, and Applied Mathematics