Efficient Lock-Free Work-stealing Iterators for Data-Parallel Collections

Transcript

1. EFFICIENT LOCK-FREE WORK-STEALING ITERATORS FOR DATA- PARALLEL COLLECTIONS Aleksandar Prokopec, Dmitry Petrashko and Martin Odersky
2. implemented in ScalaBlitz scala-blitz.github.io DATA-PARALLEL hugeCollectionOfInts.sum() has low work per element
3. PERFECT WORLD
4. REGULAR WORKLOAD
6. REAL LIFE: IRREGULARITIES
7. IRREGULAR WORKLOAD collection.map(x => fun(x))
8. IRREGULAR WORKLOAD
10. WORKSTEALING SCHEDULER Threads steal work from each other. No cooperation from target thread needed, lock free. A. Prokopec and M. Odersky. Near optimal work-stealing tree scheduler for highly irregular data- parallel workloads. LCPC 2013
11. EVEN MORE REAL LIFE: OVERHEADS
12. IRREGULAR WORKLOAD
16. TYPES OF OVERHEAD Data access overhead, Scheduling overhead, Invocation overhead. Won't be covered here.
17. DATA ACCESS OVERHEAD Data consumer abstracts over data source: Inefficiencies: iteration scheme(eg iterate over tree as if over indexed collection), redundant data copying, boxing (Jvm, Python, Ruby, JavaScript).
18. HOW SEVERE IS IT? unboxed boxed a + b box(unbox(a) + unbox(a)) ~13x slowdown
19. SCHEDULING OVERHEAD Lets say we have a HashMap: Most HashMaps by nature create irregular workload. Efficient scheduling has to be data-structure aware.
20. WORK-STEALING ITERATORS Define: how to iterate; how to steal work. Transparent for user. Easy to implement for developer.
21. BENCHMARKS For cycle:
22. BENCHMARKS Irregular workload:
23. See paper :-). MORE DETAILS? Or ask questions.
24. Eliminating abstraction penalties is necessary to realistically assess the scheduling quality. High abstraction penalties could entirely mask scheduling penalties. TAKEAWAY MESSAGE
25. THANK YOU.