From Concurrent to Parallel

on YouTube ↗01h 01m

" summary

Java SE 8 parallel streams and the fork/join runtime target data-parallel work, not IO concurrency, by using recursive divide-and-conquer on a common ForkJoinPool; benchmarks show near 8× speedup on an 8-core box for a 1,000,000-element int sum but a ~5× slowdown for boxed Integer due to poor locality. A practical rule is N×Q > 10,000 (N items × work per item) before parallelism pays, and parallelism is limited by splitting cost, merge cost, encounter-order ops, and shared mutable state.

AI-written summary · about

" chapters (16)

1. 0:00Introduction to the talk
2. 0:40The end of Moore's Law
3. 2:36Challenges in using multi-cores
4. 4:23History of Java concurrency
5. 7:31Concurrency vs. parallelism
6. 11:07Requirements for parallelism
7. 13:23Exploitable parallelism
8. 17:17Parallel decomposition patterns
9. 23:12Divide and conquer technique
10. 29:06Sources of parallel overhead
11. 32:52The Fork/Join framework
12. 34:18Introduction to Java streams
13. 37:20Stream performance modeling
14. 44:33Optimizing stream performance
15. 48:57Conclusion and best practices
16. 50:27Q&A session

" related talks

• 50mFrom Concurrent to ParallelBrian Goetz · Jfokus 2016
• 46mMake Your CPU Cores Sweat with Parallel StreamsLukasz Pater · Oracle OpenWorld 2017
• 45mThe Power and Perils of Parallel StreamsVenkat Subramaniam · Oracle OpenWorld 2017
• 48mParallel Streams, CompletableFuture, and All That: Concurrency in Java 8Kenneth Kousen · Oracle OpenWorld 2017
• 02h 35mParallel and Asynchronous Programming with Streams and CompletableFutureVenkat Subramaniam · Devoxx Belgium 2017
• 41mParallel Programming, Fork Join, and ReducersDaniel Higginbotham · Clojure/west 2016
• 01h 00mLet’s Get to the Rapids: Java 8 Stream PerformanceMaurice Naftalin · Philly ETE 2015
• 58mTurbo Charge CPU Utilization in Fork/Join Using the ManagedBlockerDr Heinz M. Kabutz · JDD 2017