EECS Publication
Fully Dynamic Scheduler for Numerical Computing on Multicore Processors
Jakub Kurzak and Jack Dongarra
The dataflow model is gaining popularity as a paradigm for programming multi-core processors and multi-socket systems of such processors. This work proposes a programming interface and an implementation for a dataflow-based scheduler, which dispatches tasks dynamically at runtime. The scheduler relies on data dependency analysis between tasks in a sequential representation of an algorithm, which provides productivity and facilitates rapid prototyping for developers. Also, through the application of dataflow principles, it ensures efficient scheduling with provisions for data reuse. Although designed with generality in mind, the scheduler is mainly intended for handling computational tasks in dense linear algebra in the PLASMA framework. Some customizations specific to that area are presented. Similarities and differences with other existing solutions are also discussed and preliminary performance results are given.
Published 2009-06-05 04:00:00 as ut-cs-09-643 (ID:75)