SRC01. Scheduling Strategies and Bounds for Cholesky Factorization on Heterogeneous Platforms
Student: Suraj Kumar (Inria)
Supervisor: Olivier Beaumont (Inria)
Abstract: We provide an analysis and comparison of different dynamic strategies for task graph scheduling on platforms consisting of heterogeneous and unrelated resources, such as GPUs and CPUs. Parallelism inside processing nodes makes static scheduling strategies difficult to precisely predict the performance of both communications and computations, due to shared resources and co-scheduling effects. Recently, to cope with this limitation, many dynamic task-graph based runtime schedulers (StarPU, StarSs, QUARK, PaRSEC) have been proposed. Dynamic schedulers decisions depend on information such as the set of available tasks, the location of data, the state of the resources and task priorities. Our analysis is deep, but we concentrate on a single kernel, namely Cholesky factorization of dense matrices on heterogeneous platforms. We analyze different dynamic strategies and propose a set of intermediate strategies, by adding more static features into dynamic strategies. We also compute theoretical upper bounds on task graphs performance.
Two-page extended abstract: pdf