 |
Center for Nonlinear Studies |
As parallel computation enters the exascale era where applications may run on millions to billions of processors concurrently, all aspects of the computational model need to undergo a transformation to meet the challenges of scaling impaired applications. One class of models aimed towards exascale computation is the task-based parallel computational model. Task-based execution models and their implementations aim to support parallelism through massive multi-threading where an application is split into numerous tasks of varying size that execute concurrently. In task-based systems, scheduling tasks onto resources can incur large overheads that vary with the underlying hardware. In this work, the goal is to dynamically control task grain size to minimize these overheads. Performance studies are used to determine overheads and metrics to use for dynamically tuning task granularity for improved performance. The performance studies and ensuing dynamic adaptation example use HPX, the first implementation of the ParalleX execution model. HPX is a C++ general runtime system that employs asynchronous fine-grained tasks and asynchronous communication for parallel and distributed applications. The performance studies give an understanding of task scheduling overheads and indication of task duration when scaling on both multi-core (Intel Ivy Bridge systems) and many-core (Xeon Phi) systems. The knowledge gained can be applied to understanding loss of efficiency in parallel and distributed applications due to both task scheduling and overheads caused by parallelization on the underlying hardware.
Host: Louis Vernon