A Halo abstraction for distributed n-dimensional structured grids within the C++ PGAS library DASH.

Area of Science:

• Computer Science
• Parallel Computing
• Scientific Computing

Background:

• The Partitioned Global Address Space (PGAS) library DASH offers C++ containers for distributed N-dimensional structured grids.
• Parallelizing structured grids efficiently requires robust support for stencil operations and halo areas.

Purpose of the Study:

• To present enhancements to the DASH library for improved parallelization of structured grids.
• To introduce and explain a novel halo abstraction concept and associated software components.
• To evaluate the performance benefits of these enhancements compared to traditional Message Passing Interface (MPI) implementations.

Main Methods:

• Developed multiple stencil operators and automatic halo size derivation within the DASH library.
• Implemented efficient halo data exchange mechanisms and communication-hiding optimizations.
• Created example code to compare DASH-based implementation with a pure Message Passing Interface (MPI) approach.

Main Results:

• The enhanced DASH library provides convenient and efficient support for stencil operations and halo areas.
• The halo abstraction and wrapper components are detailed.
• Comparative analysis shows reduced code complexity and improved runtime performance using the enhanced DASH library over MPI.

Conclusions:

• The presented enhancements significantly improve the parallelization of structured grids using the DASH library.
• The halo abstraction offers a more manageable approach to handling halo data in distributed computing.
• The DASH library with these enhancements presents a compelling alternative to MPI for structured grid computations.