Impact study of data locality on task-based applications through the Heteroprio scheduler.

Area of Science:

• Computer Science
• High-Performance Computing
• Parallel Computing

Background:

• Modern heterogeneous computing nodes require effective task-based approaches for parallel application development.
• Dynamic schedulers are critical for task distribution and load balancing in these systems.
• The DAG scheduling problem involves selecting optimal processing units for each task.

Purpose of the Study:

• To extend the Heteroprio scheduler for improved performance on multi-GPU nodes.
• To incorporate data locality considerations into task distribution strategies.
• To evaluate performance improvements without analyzing task dependencies.

Main Methods:

• Extended the Heteroprio scheduler to consider data locality.
• Implemented distinct task-lists for different memory nodes.
• Investigated locality affinity evaluation methods independent of task dependencies.
• Evaluated the approach on linear algebra and stencil codes.

Main Results:

• Simple heuristics based on data locality yield significant performance improvements.
• The enhanced scheduler reduced total memory transfer by over 50%.
• Demonstrated effectiveness on diverse scientific computing applications.

Conclusions:

• Integrating data locality into task scheduling is a viable strategy for heterogeneous computing.
• The improved Heteroprio scheduler offers substantial performance gains and memory transfer reductions.
• This method provides a practical approach to optimizing parallel applications on modern hardware.