Automatic Parallelism Management

on YouTube ↗25m

" summary

Fine-grained Fork–Join Parallel ML gets a nearly zero-cost call convention that adds two return addresses and copies the stack frame so a continuation can be promoted into a task without allocating a full closure. A runtime uses an interrupt timer (every 500 μs) and token-based amortized spending to drive a provably efficient promotion algorithm. Implemented in Maple, it cuts a worst-case 46× slowdown down to about 2–3× on 64 cores while preserving work and span.

AI-written summary · about

" chapters (14)

1. 0:00Automatic parallelism intro
2. 0:26Fine-grained task parallelism
3. 3:13Granularity control problem
4. 5:20Automatic parallelism overview
5. 7:24Compilation pipeline & PEcall
6. 10:51Dynamic promotion scheduling
7. 12:35Experimental results
8. 13:59Summary & implementation
9. 16:28Q&A overhead & promotions
10. 18:32When promotions occur
11. 20:36LLVM integration
12. 22:19Compatibility with futures
13. 23:21Promotions & cache locality
14. 24:22Final Q&A & conclusion