Polytasks: A Compressed Task Representation for HPC Runtimes

Abstract

The increased number of execution units in many-core processors is driving numerous paradigm changes in parallel systems. Previous techniques that focused solely upon obtaining correct results are being rendered obsolete unless they can also provide results efficiently.

This paper dives into the particular problem of efficiently supporting fine-grained task creation and task termination for runtime systems in shared memory processors.

Our contributions are inspired by our observation of High Performance Computing (HPC) programs, where it is common for a large number of similar fine-grained tasks to become enabled at the same time.

We present evidence showing that task creation, assignment of tasks to processors, and task termination represent a significant overhead when executing fine-grained applications in many-core processors.

We introduce the concept of the polytask, wherein the similarity of tasks created at the same time is exploited to allow faster task creation, assignment and termination. The polytask technique can be applied to any runtime system where tasks are managed through queues.

The main contributions of this work are:

• 1 The observation that task management may generate substantial overhead in fine-grained parallel programs for many core processors.

• 2 The introduction of the polytask concept: A data structure that can be added to queue-centric scheduling systems to represent groups of similar tasks.

• 3 Experimental evidence showing that the polytask is an effective way to implement fine-grained task creation/termination primitives for parallel runtime systems in many-core processors.

We use microbenchmarks to show that queues modified to handle polytasks perform orders of magnitude faster than traditional queues in some scenarios. Furthermore, we use microbenchmarks to measure the amount of time spent executing tasks. We show situations where fine-grained programs using polytasks are able to achieve efficiencies close to 100% while their efficiency becomes only 20% when not using polytasks. Finally, we use several applications with fine granularity to show that the use of polytasks results in average speedups from 1.4X to 100X depending on the queue implementation used.

This research was, in part, funded by the U.S. Government. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government.