Java 8 Parallel Stream Internals

(Part 5)

Douglas C. Schmidtd.schmidt@vanderbilt.edu

www.dre.vanderbilt.edu/~schmidt

Professor of Computer Science

Institute for Software

Integrated Systems

Vanderbilt University

Nashville, Tennessee, USA

2

• Understand parallel stream internals, e.g.

• Know what can change & what can’t

• Partition a data source into “chunks”

• Process chunks in parallel

• Configure the Java 8 parallel stream common fork-join pool

• Avoid pool starvation & improve performance w/ManagedBlocker

• Perform a reduction that combines partial results into a single result

Learning Objectives in this Part of the Lesson

join join

join

Processsequentially

Processsequentially

Processsequentially

Processsequentially

InputString1.1 InputString1.2 InputString2.1 InputString2.2

InputString1 InputString2

trySplit()

InputString

trySplit() trySplit()

See www.ibm.com/developerworks/library/j-java-streams-3-brian-goetz

3

Combining Results in a Parallel Stream

4

• After the common fork-join pool finishesprocessing chunks their partial resultsare combined into a final result

Combining Results in a Parallel Stream

join join

join

Processsequentially

Processsequentially

Processsequentially

Processsequentially

DataSource1.1 DataSource1.2 DataSource2.1 DataSource2.2

DataSource1 DataSource2

DataSource

Partial results

Final result

This discussion assumes a non-concurrent collector (more discussions follow)

5

• After the common fork-join pool finishesprocessing chunks their partial resultsare combined into a final result

• join() occurs in a single thread at each level

• i.e., the “parent”

Combining Results in a Parallel Stream

join join

join

Processsequentially

Processsequentially

Processsequentially

Processsequentially

DataSource1.1 DataSource1.2 DataSource2.1 DataSource2.2

DataSource1 DataSource2

DataSource

“Parent”

“Children”

6

• After the common fork-join pool finishesprocessing chunks their partial resultsare combined into a final result

• join() occurs in a single thread at each level

• i.e., the “parent”

Combining Results in a Parallel Stream

join join

join

Processsequentially

Processsequentially

Processsequentially

Processsequentially

DataSource1.1 DataSource1.2 DataSource2.1 DataSource2.2

DataSource1 DataSource2

DataSource

As a result, there’s typically no need for synchronizers during the joining

“Parent”

“Children”

7

• Different terminal operations combine partial results in different ways

Combining Results in a Parallel Stream

Understanding these differences is particularly important for parallel streams

8

• Different terminal operations combine partial results in different ways, e.g.

• reduce() creates a new immutable value

Combining Results in a Parallel Stream

See docs.oracle.com/javase/tutorial/essential/concurrency/immutable.html

9

• Different terminal operations combine partial results in different ways, e.g.

• reduce() creates a new immutable value

Combining Results in a Parallel Stream

See github.com/douglascraigschmidt/LiveLessons/tree/master/Java8/ex16

Processsequentially

Processsequentially

Processsequentially

Processsequentially

longs 1..2 longs 3..4 longs 5..6 longs 7..8

longs 1..4 longs 5..8

Range of longs from 1..8

long factorial(long n) {

return LongStream

.rangeClosed(1, n)

.parallel()

.reduce(1, (a, b) -> a * b,

(a, b) -> a * b);

}

2 12 30 56

10

long factorial(long n) {

return LongStream

.rangeClosed(1, n)

.parallel()

.reduce(1, (a, b) -> a * b,

(a, b) -> a * b);

}

• Different terminal operations combine partial results in different ways, e.g.

• reduce() creates a new immutable value

Combining Results in a Parallel Stream

reduce() combines two immutable values (e.g., long or Long) & produces a new one

reduce()

reduce()

Processsequentially

Processsequentially

Processsequentially

Processsequentially

longs 1..2 longs 3..4 longs 5..6 longs 7..8

longs 1..4 longs 5..8

Range of longs from 1..8

2 12 30 56

24

40,320

reduce() 1,680

11See greenteapress.com/thinkapjava/html/thinkjava011.html

• Different terminal operations combine partial results in different ways, e.g.

• reduce() creates a new immutable value

• collect() mutates an existing value

Combining Results in a Parallel Stream

12

Set<CharSequence>

uniqueWords =

getInput(sSHAKESPEARE),

"\\s+")

.parallelStream()

...

.collect(toCollection(TreeSet::new));

• Different terminal operations combine partial results in different ways, e.g.

• reduce() creates a new immutable value

• collect() mutates an existing value

Combining Results in a Parallel Stream

Processsequentially

Processsequentially

Processsequentially

Processsequentially

1st quarter of words 2nd quarter of words 3rd quarter of words 4th quarter of words

1st half of words 2nd half of words

All words in Shakespeare’s works

See github.com/douglascraigschmidt/LiveLessons/tree/master/Java8/ex14

13

Set<CharSequence>

uniqueWords =

getInput(sSHAKESPEARE),

"\\s+")

.parallelStream()

...

.collect(toCollection(TreeSet::new));

• Different terminal operations combine partial results in different ways, e.g.

• reduce() creates a new immutable value

• collect() mutates an existing value

Combining Results in a Parallel Stream

collect() collect()

collect()

Processsequentially

Processsequentially

Processsequentially

Processsequentially

collect() mutates a container to accumulate the result it’s producing

1st quarter of words 2nd quarter of words 3rd quarter of words 4th quarter of words

1st half of words 2nd half of words

All words in Shakespeare’s works

14

Set<CharSequence>

uniqueWords =

getInput(sSHAKESPEARE),

"\\s+")

.parallelStream()

...

.collect(ConcurrentHashSetCollector.toSet());

• Different terminal operations combine partial results in different ways, e.g.

• reduce() creates a new immutable value

• collect() mutates an existing value

Combining Results in a Parallel Stream

Processsequentially

Processsequentially

Processsequentially

Processsequentially

Concurrent collectors (covered later) are different than non-concurrent collectors

1st quarter of words 2nd quarter of words 3rd quarter of words 4th quarter of words

1st half of words 2nd half of words

All words in Shakespeare’s works

accumulate() accumulate()

accumulate()

ConcurrentResult Container

15

• More discussion about reduce() vs. collect() appears online

Combining Results in a Parallel Stream

See www.youtube.com/watch?v=oWlWEKNM5Aw

16

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

Combining Results in a Parallel Streamvoid buggyStreamReduce3

(boolean parallel) {

...

Stream<String> wordStream =

allWords.stream();

if (parallel)

wordStream.parallel();

String words = wordStream

.reduce(new StringBuilder(),

StringBuilder::append,

StringBuilder::append)

.toString();

See github.com/douglascraigschmidt/LiveLessons/tree/master/Java8/ex17

17

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

Combining Results in a Parallel Streamvoid buggyStreamReduce3

(boolean parallel) {

...

Stream<String> wordStream =

allWords.stream();

if (parallel)

wordStream.parallel();

String words = wordStream

.reduce(new StringBuilder(),

StringBuilder::append,

StringBuilder::append)

.toString();

Convert a list of words into a stream of works

Naturally, this call doesn’t really do any work since streams are “lazy”

18

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

Combining Results in a Parallel Streamvoid buggyStreamReduce3

(boolean parallel) {

...

Stream<String> wordStream =

allWords.stream();

if (parallel)

wordStream.parallel();

String words = wordStream

.reduce(new StringBuilder(),

StringBuilder::append,

StringBuilder::append)

.toString();

A stream can be dynamically switched to “parallel” mode!

19

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

Combining Results in a Parallel Streamvoid buggyStreamReduce3

(boolean parallel) {

...

Stream<String> wordStream =

allWords.stream();

if (parallel)

wordStream.parallel();

String words = wordStream

.reduce(new StringBuilder(),

StringBuilder::append,

StringBuilder::append)

.toString();

See mail.openjdk.java.net/pipermail/lambda-libs-spec-experts/2013-March/001504.html

The “last” call to .parallel() or .sequential() in a stream “wins”

20

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

Combining Results in a Parallel Streamvoid buggyStreamReduce3

(boolean parallel) {

...

Stream<String> wordStream =

allWords.stream();

if (parallel)

wordStream.parallel();

String words = wordStream

.reduce(new StringBuilder(),

StringBuilder::append,

StringBuilder::append)

.toString();

This code works when parallel is false since the StringBuilder is only called in a single thread

See docs.oracle.com/javase/8/docs/api/java/lang/StringBuilder.html

21

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

Combining Results in a Parallel Streamvoid buggyStreamReduce3

(boolean parallel) {

...

Stream<String> wordStream =

allWords.stream();

if (parallel)

wordStream.parallel();

String words = wordStream

.reduce(new StringBuilder(),

StringBuilder::append,

StringBuilder::append)

.toString();

This code fails when parallel is true since reduce() expects to do

an “immutable” reduction

22

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

Combining Results in a Parallel Streamvoid buggyStreamReduce3

(boolean parallel) {

...

Stream<String> wordStream =

allWords.stream();

if (parallel)

wordStream.parallel();

String words = wordStream

.reduce(new StringBuilder(),

StringBuilder::append,

StringBuilder::append)

.toString();

There are race conditions here since there’s just one shared StringBuilder, which is not properly thread-safe..

See www.baeldung.com/java-string-builder-string-buffer

23

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

Combining Results in a Parallel Streamvoid buggyStreamReduce3

(boolean parallel) {

...

Stream<String> wordStream =

allWords.stream();

if (parallel)

wordStream.parallel();

String words = wordStream

.reduce(new String(),

(x , y) -> x + y);

This simple fix is inefficient due to string concatenation overhead

See javarevisited.blogspot.com/2015/01/3-examples-to-concatenate-string-in-java.html

24

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

Combining Results in a Parallel Streamvoid streamCollectJoining

(boolean parallel) {

...

Stream<String> wordStream =

allWords.stream();

if (parallel)

wordStream.parallel();

String words = wordStream

.collect(joining());

This is a much better solution!!

See www.mkyong.com/java8/java-8-stringjoiner-example

25

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

• Beware of issues related toassociativity & identity

Combining Results in a Parallel Streamvoid testDifferenceReduce(...) {

long difference = LongStream

.rangeClosed(1, 100)

.parallel()

.reduce(0L,

(x, y) -> x - y);

}

void testSum(long identity, ...) {

long sum = LongStream

.rangeClosed(1, 100)

.reduce(identity,

// Could use (x, y) -> x + y

Math::addExact);

See github.com/douglascraigschmidt/LiveLessons/tree/master/Java8/ex17

26

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

• Beware of issues related toassociativity & identity

Combining Results in a Parallel Streamvoid testDifferenceReduce(...) {

long difference = LongStream

.rangeClosed(1, 100)

.parallel()

.reduce(0L,

(x, y) -> x - y);

}

void testSum(long identity, ...) {

long sum = LongStream

.rangeClosed(1, 100)

.reduce(identity,

// Could use (x, y) -> x + y

Math::addExact);

This code fails for a parallel stream since subtraction

is not associative

See www.ibm.com/developerworks/library/j-java-streams-2-brian-goetz

27

• More discussion about reduce() vs. collect() appears online, e.g.

• Always test w/a parallel streamto detect mistakes wrt mutablevs. immutable reductions

• Beware of issues related toassociativity & identity

Combining Results in a Parallel Streamvoid testDifferenceReduce(...) {

long difference = LongStream

.rangeClosed(1, 100)

.parallel()

.reduce(0L,

(x, y) -> x - y);

}

void testSum(long identity, ...) {

long sum = LongStream

.rangeClosed(1, 100)

.reduce(identity,

// Could use (x, y) -> x + y

Math::addExact);

This code fails if identity is not 0L

The “identity” of an OP is defined as “identity OP value == value” (& inverse)

28

End of Java 8 Parallel Stream Internals (Part 5)