C++ on its way to exascale and beyond

on YouTube ↗59m

" summary

C++ adds concrete parallel tools: the HPX runtime implements futures, executors and an active global address space and runs hundreds of millions of user threads; a NUMA allocator plus executor-bound parallel algorithms place data and work per domain, saturating a memory controller at ~6 cores and capping cores to save energy; benchmarks show stream-style bandwidth close to tuned OpenMP while a tiled matrix transposition outperforms OpenMP by ~30%, so the approach works but task-spawn hierarchy and tuning still cost performance.

AI-written summary · about

" chapters (19)

1. 0:00Intro
2. 0:40What is exascale
3. 2:20Why exascale
4. 8:03Challenges
5. 13:12Current contenders
6. 16:58Parallelization
7. 17:50Four horsemen
8. 20:44Current state of the art
9. 24:17HPX
10. 27:58Parallelism
11. 33:44Task Policy
12. 35:34Execution Parameters
13. 36:45SexP
14. 38:04Irregular algorithm
15. 40:30How it works
16. 44:24Benchmarks
17. 48:44Synchronous control flow
18. 51:52Dataflow extension
19. 56:13Whats beyond

" related talks

• 01h 06mAsynchronous Computation in C++Hartmut Kaiser · CppCon 2014
• 01h 02mHPX - A C++ Library for Parallelism and ConcurrencyHartmut Kaiser · CppCon 2022
• 01h 02mOverview of Parallel Programming in C++Pablo G Halpern · CppCon 2014
• 01h 16mPaying for Lunch: C++ in the ManyCore AgeHerb Sutter · CppCon 2014
• 01h 07mAnother fundamental shift in Parallelism Paradigm?Michael S Wong · CppCon 2014
• 01h 02mC++ Performance Portability - A Decade of Lessons LearnedChristian Trott · CppCon 2022
• 37mC++ for multi-accelerator, multi-vendor systemsGuy Tamir · Meeting C++ 2022
• 05mFast Parallel Programming using Modern C++Tsung-Wei Huang · CppCon 2018