Overview of Java 8 Parallel Streams (Part 1)

Transcription

1 Overview of Java 8 Parallel s (Part 1) Douglas C. Schmidt d.schmidt@vanderbilt.edu Professor of Computer Science Institute for Software Integrated Systems Vanderbilt University Nashville, Tennessee, USA

2 Learning Objectives in this Part of the Lesson Know how aggregate operations & functional programming features are applied in the parallel streams framework Input x Aggregate operation (Function f) Output f(x) Aggregate operation (Function g) Output g(f(x)) Aggregate operation (Function h) Output h(g(f(x))) 2

3 Learning Objectives in this Part of the Lesson Know how aggregate operations & functional programming features are applied in the parallel streams framework Understand how a parallel stream works DataSource DataSource 1 DataSource 2 DataSource 1.1 DataSource 1.2 DataSource 2.1 DataSource 2.2 join join join 3

4 Overview of Java 8 Parallel s 4

5 Overview of Java 8 Parallel s A Java 8 parallel stream splits its elements into multiple chunks & uses a thread pool to process these chunks independently Input x Output f(x) Aggregate operation (behavior f) Aggregate operation (behavior g) Output g(f(x)) Aggregate operation (behavior h) Output h(g(f(x))) See docs.oracle.com/javase/tutorial/collections/streams/parallelism.html 5

6 Overview of Java 8 Parallel s A Java 8 parallel stream splits its elements into multiple chunks & uses a thread pool to process these chunks independently This splitting & thread pool are often invisible to programmers Input x Output f(x) Aggregate operation (behavior f) Aggregate operation (behavior g) Output g(f(x)) Aggregate operation (behavior h) Output h(g(f(x))) 6

7 Overview of Java 8 Parallel s A Java 8 parallel stream splits its elements into multiple chunks & uses a thread pool to process these chunks independently This splitting & thread pool are often invisible to programmers The order in which chunks are processed is likely non-deterministic Input x Aggregate operation (behavior f) Output f(x) Aggregate operation (behavior g) Output g(f(x)) Aggregate operation (behavior h) Output h(g(f(x))) i.e., programmers often have little/no 7control over how chunks are processed

8 Overview of Java 8 Parallel s A Java 8 parallel stream splits its elements into multiple chunks & uses a thread pool to process these chunks independently This splitting & thread pool are often invisible to programmers The order in which chunks are processed is likely non-deterministic This non-determinism is usually a good thing! Input x Aggregate operation (behavior f) Output f(x) Aggregate operation (behavior g) Output g(f(x)) Aggregate operation (behavior h) Output h(g(f(x))) 8

9 Overview of Java 8 Parallel s A Java 8 parallel stream splits its elements into multiple chunks & uses a thread pool to process these chunks independently This splitting & thread pool are often invisible to programmers The order in which chunks are processed is likely non-deterministic The results of the processing are likely deterministic Input x Aggregate operation (behavior f) Output f(x) Aggregate operation (behavior g) Output g(f(x)) Aggregate operation (behavior h) Output h(g(f(x))) Programmers have more control 9over how the results are presented

10 Overview of Java 8 Parallel s When a stream executes all of its aggregate operations run in a single thread List <String> <String> Search Phrases stream() map(phrase -> searchforphrase()) filter(not(searchresults::isempty)) List collect(tolist()) 10

11 Overview of Java 8 Parallel s When a stream executes in parallel, it is partitioned into multiple substream chunks that run in a common fork-join pool List <String> <String> Search Phrases parallel() map(phrase -> searchforphrase()) filter(not(searchresults::isempty)) List collect(tolist()) See docs.oracle.com/javase/8/docs/api/java/util/concurrent/forkjoinpool.html 11

12 Overview of Java 8 Parallel s When a stream executes in parallel, it is partitioned into multiple substream chunks that run in a common fork-join pool List <String> <String> Search Phrases parallel() map(phrase -> searchforphrase()) filter(not(searchresults::isempty)) List collect(tolist()) Threads in the pool process different 12 chunks in a non-deterministic order

13 Overview of Java 8 Parallel s When a stream executes in parallel, it is partitioned into multiple substream chunks that run in a common fork-join pool List <String> <String> List Search Phrases parallel() map(phrase -> searchforphrase()) filter(not(searchresults::isempty)) collect(tolist()) Intermediate operations iterate 13over & process these chunks in parallel

14 Overview of Java 8 Parallel s When a stream executes in parallel, it is partitioned into multiple substream chunks that run in a common fork-join pool List <String> <String> List Search Phrases parallel() map(phrase -> searchforphrase()) filter(not(searchresults::isempty)) collect(tolist()) A terminal operation then 14 combines the chunks into a single result

15 Overview of Java 8 Parallel s When a stream executes in parallel, it is partitioned into multiple substream chunks that run in a common fork-join pool List <String> <String> List Search Phrases parallel() map(phrase -> searchforphrase()) filter(not(searchresults::isempty)) collect(tolist()) (Stateless) Java 8 lambda expressions & method 15 references are used to pass behaviors

16 Overview of Java 8 Parallel s The same aggregate operations can be used for sequential & parallel streams See docs.oracle.com/javase/8/docs/api/java/util/stream/.html 16

17 Overview of Java 8 Parallel s The same aggregate operations can be used for sequential & parallel streams e.g., SearchGang uses the same aggregate operations for both SearchWithSequentials & SearchWithParallels implementations Search Phrases stream() vs. parallel() map(phrase -> searchforphrase()) filter(not(searchresults::isempty)) collect(tolist()) See github.com/douglascraigschmidt/livelessons/tree/master/searchgang 17

18 Overview of Java 8 Parallel s The same aggregate operations can be used for sequential & parallel streams Java 8 streams can thus treat parallelism as an optimization & leverage all available cores! Search Phrases parallel() map(phrase -> searchforphrase()) filter(not(searchresults::isempty)) collect(tolist()) See qconlondon.com/london-2017/system/files/presentation-slides/concurrenttoparallel.pdf 18

19 Overview of Java 8 Parallel s The same aggregate operations can be used for sequential & parallel streams Java 8 streams can thus treat parallelism as an optimization & leverage all available cores! Naturally, behaviors run by these aggregate operations must be designed carefully to avoid accessing unsynchronized shared state.. parallel() Search Phrases map(phrase -> searchforphrase()) filter(not(searchresults::isempty)) collect(tolist()) See henrikeichenhardt.blogspot.com/2013/06/why-shared-mutable-state-is-root-of-all.html 19

20 Overview of How a Parallel Works 20

21 Overview of How a Parallel Works A Java 8 parallel stream implements a map/reduce variant optimized for multi-core processors Map Reduce Partition See en.wikipedia.org/wiki/mapreduce 21

22 Overview of How a Parallel Works A Java 8 parallel stream implements a map/reduce variant optimized for multi-core processors It s actually more like the split-apply-combine data analysis strategy DataSource 1 DataSource 2 DataSource DataSource 1.1 DataSource 1.2 DataSource 2.1 DataSource 2.2 join join join See 22

23 Overview of How a Parallel Works Split-apply-combine works as follows: 1. Split Recursively partition a data source into independent chunks CollectionData CollectionData 1 CollectionData 2 CollectionData 1.1 CollectionData 1.2 CollectionData 2.1 CollectionData 2.2 See en.wikipedia.org/wiki/divide_and_conquer_algorithm 23

24 Overview of How a Parallel Works Split-apply-combine works as follows: 1. Split Recursively partition a data source into independent chunks Spliterators are defined to partition collections in Java 8 CollectionData CollectionData 1 CollectionData 2 CollectionData 1.1 CollectionData 1.2 CollectionData 2.1 CollectionData 2.2 See docs.oracle.com/javase/8/docs/api/java/util/spliterator.html 24

25 Overview of How a Parallel Works Split-apply-combine works as follows: 1. Split Recursively partition a data source into independent chunks Spliterators are defined to partition collections in Java 8 You can also define custom spliterators InputString InputString 1 InputString 2 InputString 1.1 InputString 1.2 InputString 2.1 InputString 2.2 Input Strings to Search Search Phrases 45,000+ phrases See github.com/douglascraigschmidt/livelessons/tree/master/searchspliterator 25

26 Overview of How a Parallel Works Split-apply-combine works as follows: 1. Split Recursively partition a data source into independent chunks Spliterators are defined to partition collections in Java 8 You can also define custom spliterators Parallel streams perform better on data sources that can be split efficiently & evenly InputString InputString 1 InputString 2 InputString 1.1 InputString 1.2 InputString 2.1 InputString 2.2 See 26

27 Overview of How a Parallel Works Split-apply-combine works as follows: 1. Split Recursively partition a data source into independent chunks 2. Apply chunks independently in a pool of threads InputString InputString 1 InputString 2 InputString 1.1 InputString 1.2 InputString 2.1 InputString 2.2 Splitting & applying run simultaneously 27 (after certain limit met), not

28 Overview of How a Parallel Works Split-apply-combine works as follows: 1. Split Recursively partition a data source into independent chunks 2. Apply chunks independently in a pool of threads Programmers have some control over how many threads are in the pool InputString InputString 1 InputString 2 InputString 1.1 InputString 1.2 InputString 2.1 InputString

29 Overview of How a Parallel Works Split-apply-combine works as follows: 1. Split Recursively partition a data source into independent chunks 2. Apply chunks independently in a pool of threads 3. Combine Join partial results into a single result InputString InputString 1 InputString 2 InputString 1.1 InputString 1.2 InputString 2.1 InputString 2.2 join join join 29

30 Overview of How a Parallel Works Split-apply-combine works as follows: 1. Split Recursively partition a data source into independent chunks 2. Apply chunks independently in a pool of threads 3. Combine Join partial results into a single result Performed by terminal operations like collect() & reduce() join InputString InputString 1 InputString 2 InputString 1.1 InputString 1.2 InputString 2.1 InputString 2.2 join join See 30

31 End of Overview of Java 8 Parallel s (Part 1) 31