Multidimensional Index Sets for Data Locality in HPC Applications
summary
Multidimensional index-set views in the DASH C++ library expose separate canonical and local index spaces, let algorithms compute intersections and contiguous blocks for packing halo regions, and generate optimized copy/communication schedules instead of elementwise remote accesses. The implementation targets MPI/GASPI RMA backends with a CUDA path, supports tiled, block-cyclic and irregular patterns, and demonstrates a compile-time composition that yields about a 6.2x speedup by letting the compiler optimize expression templates for multidimensional index-set views.
AI-written summary · about
chapters (6)
- 0:00Introduction and motivation
- 10:37Index spaces and data locality
- 21:05DASH library and global arrays
- 30:24Iteration spaces and patterns
- 43:08Halo exchange and practical demo
- 51:06Q&A and future directions
related talks
56m“High Performance Code 201: Hybrid Data Structures"
30mDatastore Axes: Choosing the Scalability Direction You Need
01h 03mDesigning C++ portable SIMD support
43mArchitecting for Speed & Scale
01h 19mCo-Design for High-Performance Computing Software Systems
48mDesign of data-intensive applications
01h 25mLakos'20: The "Dam" Book is Done!
15mDisjunctive Interval Analysis