Multicore Programming

Transcription

1 Multicore Programming Java Streams Louis-Claude Canon Bureau 414C Master 1 informatique Semestre 8 Louis-Claude Canon MCP Java Streams 1 / 124

2 Motivations Express simple iterations in a declarative way without writing the loop logic. Simple way to parallelize without writing the thread logic. Java Streams introduced in Java 8 (2014, LTS, support until 2025). Called iterators in Rust. Louis-Claude Canon MCP Java Streams 2 / 124

3 Outline Concepts Stream Operations Special Streams Collectors Parallel Data Processing Summary and References Louis-Claude Canon MCP Java Streams 3 / 124

4 Concepts Stream Definition Outline Concepts Stream Definition Difference with Collections Intermediate and Terminal Operations Stream Operations Special Streams Collectors Parallel Data Processing Summary and References Louis-Claude Canon MCP Java Streams 4 / 124

5 Concepts Stream Definition Example without Streams List<Dish> lowcal = new ArrayList<>(); for (Dish dish: menu) if (dish.getcalories() < 400) lowcaloricdishes.add(dish); Collections.sort(lowCal, new Comparator<Dish>() { public int compare(dish dish1, Dish dish2) { return Integer.compare(dish1.getCalories(), dish2.getcalories()); } }); List<String> lowcalname = new ArrayList<>(); for (Dish dish: lowcal) lowcalname.add(dish.getname()); Louis-Claude Canon MCP Java Streams 5 / 124

6 Concepts Stream Definition Example with Streams import static java.util.comparator.comparing; import static java.util.stream.collectors.tolist; List<String> lowcaloricdishesname = menu.stream().filter(d -> d.getcalories() < 400).sorted(comparing(Dish::getCalories)).map(Dish::getName).collect(toList()); Louis-Claude Canon MCP Java Streams 6 / 124

7

8

9 Concepts Stream Definition Benefits declarative (no flow control): readable and concise composable/flexible/extensible: easy to modify parallelizable Louis-Claude Canon MCP Java Streams 9 / 124

10 Concepts Stream Definition Definition A stream is a sequence of elements coming from a source and that supports data-processing operations. sequence of elements interface to access a sequenced set of values (not necessarily ordered) source data are provided by a collection, array or I/O resource data-processing operations support for database-like or functional operations Louis-Claude Canon MCP Java Streams 10 / 124

11 Concepts Stream Definition Characteristics Pipelining stream operation may return a stream Internal iteration iteration (loop logic) is implicit Louis-Claude Canon MCP Java Streams 11 / 124

12 Concepts Difference with Collections Outline Concepts Stream Definition Difference with Collections Intermediate and Terminal Operations Stream Operations Special Streams Collectors Parallel Data Processing Summary and References Louis-Claude Canon MCP Java Streams 12 / 124

13 Concepts Difference with Collections Similarity between Streams and Collections Iteration on a sequenced set of values. Louis-Claude Canon MCP Java Streams 13 / 124

14 Concepts Difference with Collections Differences Lazy vs. Eager Evaluation: values are only computed as needed. Traversable only once: a stream cannot be consumed multiple times. Internal vs. external iterations: iterations can be transparently done in parallel or in a different more optimized order. Louis-Claude Canon MCP Java Streams 14 / 124

15 Concepts Intermediate and Terminal Operations Outline Concepts Stream Definition Difference with Collections Intermediate and Terminal Operations Stream Operations Special Streams Collectors Parallel Data Processing Summary and References Louis-Claude Canon MCP Java Streams 15 / 124

16

17 Concepts Intermediate and Terminal Operations Intermediate Operations Intermediate operations: return another stream only define processing operations that will be performed (lazy evaluation) Louis-Claude Canon MCP Java Streams 17 / 124

18 Concepts Intermediate and Terminal Operations Example List<String> names = menu.stream().filter(dish -> { System.out.println("filter:" + dish.getname()); return dish.getcalories() > 300; }).map(dish -> { System.out.println("map:" + dish.getname()); return dish.getname(); }).limit(3).collect(tolist()); Louis-Claude Canon MCP Java Streams 18 / 124

19 Concepts Intermediate and Terminal Operations Example Output filtering:pork mapping:pork filtering:beef mapping:beef filtering:chicken mapping:chicken Louis-Claude Canon MCP Java Streams 19 / 124

20 Concepts Intermediate and Terminal Operations Terminal Operations Terminal operations produce a result from a stream pipeline (lazy evaluation). Louis-Claude Canon MCP Java Streams 20 / 124

21 Concepts Intermediate and Terminal Operations Summary A data source (such as a collection) to perform a query on. A chain of intermediate operations that form a stream pipeline. A terminal operation that executes the stream pipeline and produces a result. Louis-Claude Canon MCP Java Streams 21 / 124

22 Stream Operations Filtering Outline Concepts Stream Operations Filtering Slicing Mapping Finding and Matching Reducing Special Streams Collectors Parallel Data Processing Louis-Claude Canon MCP Java Streams 22 / 124

23 Stream Operations Filtering Predicate A predicate is a function returning a boolean. Used to filter values. Louis-Claude Canon MCP Java Streams 23 / 124

24 Stream Operations Filtering filter Louis-Claude Canon MCP Java Streams 24 / 124

25 Stream Operations Filtering distinct Louis-Claude Canon MCP Java Streams 25 / 124

26 Stream Operations Filtering Other Special Function Names From package java.util.function: Supplier void -> T Consumer T -> void Function T -> R UnaryOperator T -> T Predicate T -> boolean Equivalences with different arity: BiConsumer (T, U) -> void BiFunction (T, U) -> R BinaryOperator (T, T) -> T BiPredicate (T, U) -> boolean Louis-Claude Canon MCP Java Streams 26 / 124

27 Stream Operations Slicing Outline Concepts Stream Operations Filtering Slicing Mapping Finding and Matching Reducing Special Streams Collectors Parallel Data Processing Louis-Claude Canon MCP Java Streams 27 / 124

28 Stream Operations Slicing Definition Slicing: skip elements by ignoring either first or last elements. With a fixed number of elements: limit, skip. With a predicate (from Java 9): takewhile, dropwhile. Louis-Claude Canon MCP Java Streams 28 / 124

29 Stream Operations Slicing limit Louis-Claude Canon MCP Java Streams 29 / 124

30 Stream Operations Slicing skip Louis-Claude Canon MCP Java Streams 30 / 124

31 Stream Operations Slicing takewhile List<Dish> slicedmenu1 = specialmenu.stream().takewhile(dish -> dish.getcalories() < 320).collect(toList()); The methods takewhile and dropwhile are more efficient than a filter when you know that the source is sorted. Louis-Claude Canon MCP Java Streams 31 / 124

32 Stream Operations Mapping Outline Concepts Stream Operations Filtering Slicing Mapping Finding and Matching Reducing Special Streams Collectors Parallel Data Processing Louis-Claude Canon MCP Java Streams 32 / 124

33 Stream Operations Mapping map Applying a function: List<String> dishnames = menu.stream().map(dish::getname).collect(tolist()); Types: 1. menu: List<Dish> 2. menu.stream(): Stream<Dish> 3. menu.stream().map(): Stream<String> 4. menu.stream().map().collect(): List<String> Louis-Claude Canon MCP Java Streams 33 / 124

34 Stream Operations Mapping flatmap Flattening streams: List<String> uniquecharacters = words.stream().map(word -> word.split("")).flatmap(arrays::stream).distinct().collect(tolist()); Apply a function that returns a stream (as map). Merge the resulting streams into a single one. Louis-Claude Canon MCP Java Streams 34 / 124

35

36 Stream Operations Finding and Matching Outline Concepts Stream Operations Filtering Slicing Mapping Finding and Matching Reducing Special Streams Collectors Parallel Data Processing Louis-Claude Canon MCP Java Streams 36 / 124

37 Stream Operations Finding and Matching Matching: anymatch, allmatch, nonematch Rely on a predicate: boolean ishealthy = menu.stream().allmatch(dish -> dish.getcalories() < 1000); boolean ishealthy = menu.stream().nonematch(d -> d.getcalories() >= 1000); Louis-Claude Canon MCP Java Streams 37 / 124

38 Stream Operations Finding and Matching Short-Circuiting Interrupt the processing iteration on the stream as soon as a condition is met. Similar to evaluation mechanism with and &&. Apply to limit and takewhile as well. Allow infinite stream. Louis-Claude Canon MCP Java Streams 38 / 124

39 Stream Operations Finding and Matching Finding: findany, findfirst Optional<Dish> dish = menu.stream().filter(dish::isvegetarian).findany(); findany is better for parallelization than findfirst Louis-Claude Canon MCP Java Streams 39 / 124

40 Stream Operations Finding and Matching Optional<T> An object that is either defined or null (explicit management of null references). boolean ispresent() ifpresent(consumer<t> block) T get() T orelse(t other) Louis-Claude Canon MCP Java Streams 40 / 124

41 Stream Operations Reducing Outline Concepts Stream Operations Filtering Slicing Mapping Finding and Matching Reducing Special Streams Collectors Parallel Data Processing Louis-Claude Canon MCP Java Streams 41 / 124

42 Stream Operations Reducing reduce Reducing a list of elements into a single value (also called fold): int sum = numbers.stream().reduce(0, Integer::sum); Louis-Claude Canon MCP Java Streams 42 / 124

43

44 Stream Operations Reducing No Initial Value Optional<Integer> product = numbers.stream().reduce((a, b) -> (a * b)); Louis-Claude Canon MCP Java Streams 44 / 124

45 Stream Operations Reducing Maximum and Minimum Optional<Integer> max = numbers.stream().reduce( Integer::max); Optional<Integer> min = numbers.stream().reduce( Integer::min); Louis-Claude Canon MCP Java Streams 45 / 124

46 Stream Operations Reducing Stateless vs. Stateful reduce is stateful (with a bounded state), i.e. it keeps an intermediate result (called an accumulator). sorted and distinct are even more stateful (unbounded state): they need to buffer all the elements of the stream to proceed. Louis-Claude Canon MCP Java Streams 46 / 124

47 Stream Operations Reducing Summary In addition to the distinction between intermediate and terminal operations, operations are characterized by whether they: allow short-circuiting (takewhile, limit, find, match) return an optional (for terminal operations): find, reduce, min, max are stateful (bounded or not): limit, skip, takewhile, dropwhile, distinct (un), sorted (un), (reduce) Louis-Claude Canon MCP Java Streams 47 / 124

48 Special Streams Numeric Streams Outline Concepts Stream Operations Special Streams Numeric Streams Building Streams Infinite Streams Collectors Parallel Data Processing Summary and References Louis-Claude Canon MCP Java Streams 48 / 124

49 Special Streams Numeric Streams Primitive Stream Specializations IntStream, DoubleStream and LongStream. Avoid unboxing and boxing (more efficient data representation). Additional methods for efficiency and convenience: min, max, sum, average, etc. Louis-Claude Canon MCP Java Streams 49 / 124

50 Special Streams Numeric Streams To and From a Primitive Stream maptoint: boxed: int calories = menu.stream().maptoint(dish::getcalories).sum(); Stream<Integer> stream = intstream.boxed(); Louis-Claude Canon MCP Java Streams 50 / 124

51 Special Streams Numeric Streams Specialized Optionals OptionalInt OptionalDouble OptionalLong Louis-Claude Canon MCP Java Streams 51 / 124

52 Special Streams Building Streams Outline Concepts Stream Operations Special Streams Numeric Streams Building Streams Infinite Streams Collectors Parallel Data Processing Summary and References Louis-Claude Canon MCP Java Streams 52 / 124

53 Special Streams Building Streams From Values Stream<String> stream = Stream.of("Modern ", "Java ", "In ", "Action"); Louis-Claude Canon MCP Java Streams 53 / 124

54 Special Streams Building Streams From Nullable From Java 9: String homevalue = System.getProperty("home"); Stream<String> homevaluestream = homevalue == null? Stream.empty() : Stream.of(value); Stream<String> homevaluestream = Stream.ofNullable(System.getProperty("home")); A stream with a null value differs from an empty stream. Louis-Claude Canon MCP Java Streams 54 / 124

55 Special Streams Building Streams From Arrays int[] numbers = {2, 3, 5, 7, 11, 13}; int sum = Arrays.stream(numbers).sum(); Louis-Claude Canon MCP Java Streams 55 / 124

56 Special Streams Building Streams From a Collection List<Integer> numbers = Arrays.asList(1, 2, 1, 3, 3, 2, 4); int sum = numbers.stream().sum(); Louis-Claude Canon MCP Java Streams 56 / 124

57 Special Streams Building Streams From Files long uniquewords = 0; try (Stream<String> lines = Files.lines(Paths.get("data.txt"), Charset.defaultCharset())) { uniquewords = lines.flatmap(line -> Arrays.stream(line.split(" "))).distinct().count(); } catch (IOException e) { } Louis-Claude Canon MCP Java Streams 57 / 124

58 Special Streams Building Streams From Numeric Ranges IntStream evennumbers = IntStream.rangeClosed(1, 100).filter(n -> n % 2 == 0); range is exclusive rangeclosed is inclusive Louis-Claude Canon MCP Java Streams 58 / 124

59 Special Streams Infinite Streams Outline Concepts Stream Operations Special Streams Numeric Streams Building Streams Infinite Streams Collectors Parallel Data Processing Summary and References Louis-Claude Canon MCP Java Streams 59 / 124

60 Special Streams Infinite Streams iterate Stream.iterate(0, n -> n + 2).limit(10).forEach(System.out::println); IntStream.iterate(0, n -> n < 100, n -> n + 4).forEach(System.out::println); Needs a short-circuiting operation to terminate. No unbounded stateful operation (reduce and sorted). Louis-Claude Canon MCP Java Streams 60 / 124

61 Special Streams Infinite Streams generate Stream.generate(Math::random).limit(5).forEach(System.out::println); Requires a stateful (with a mutable state) supplier. Louis-Claude Canon MCP Java Streams 61 / 124

62 Collectors collect Summary and References Louis-Claude Canon MCP Java Streams 62 / 124 Outline Concepts Stream Operations Special Streams Collectors collect Reducing and Summarizing Grouping and Partitioning Custom Collector Parallel Data Processing

63 Collectors collect collect Collect the final result as the terminal operation. Require a Collector object as its parameter. A collector is designed to work on a stream. No relation with Collection. Several ways to define such collectors in class Collectors. stream.collect(tolist()); Louis-Claude Canon MCP Java Streams 63 / 124

64 Collectors Reducing and Summarizing Summary and References Louis-Claude Canon MCP Java Streams 64 / 124 Outline Concepts Stream Operations Special Streams Collectors collect Reducing and Summarizing Grouping and Partitioning Custom Collector Parallel Data Processing

65 Collectors Reducing and Summarizing maxby and minby Comparator<Dish> dishcaloriescomparator = Comparator.comparingInt(Dish::getCalories); Optional<Dish> mostcaloriedish = menu.stream().collect(maxby(dishcaloriescomparator)); Louis-Claude Canon MCP Java Streams 65 / 124

66 Collectors Reducing and Summarizing Summarizing: summinint, counting, averagingint and summarizingint int totalcalories = menu.stream().collect(summingint(dish::getcalories)); long count = menu.stream().collect(counting()); double avgcalories = menu.stream().collect(averagingint(dish::getcalories)); IntSummaryStatistics menustatistics = menu.stream().collect(summarizingint(dish::getcalories)); Louis-Claude Canon MCP Java Streams 66 / 124

67

68 Collectors Reducing and Summarizing joining String shortmenu = menu.stream().map(dish::getname).collect(joining(", ")); Louis-Claude Canon MCP Java Streams 68 / 124

69 Collectors Reducing and Summarizing reducing int totalcalories = menu.stream().collect(reducing(0, Dish::getCalories, Integer::sum)); Generalization of previous summarizing operations. First apply map function (e.g. getcalories), and then performs the reduction operation (e.g. sum). Louis-Claude Canon MCP Java Streams 69 / 124

70

71 Collectors Reducing and Summarizing reduce, reducing and collect reducing is like reduce but for a multi-level reduction. reduce must work on an immutable state (the accumulating and combining functions should have no side effect) because it can be parallelized. collect builds incrementally a mutable object (more efficient for strings because of StringBuilder). Louis-Claude Canon MCP Java Streams 71 / 124

72 Collectors Grouping and Partitioning Summary and References Louis-Claude Canon MCP Java Streams 72 / 124 Outline Concepts Stream Operations Special Streams Collectors collect Reducing and Summarizing Grouping and Partitioning Custom Collector Parallel Data Processing

73 Collectors Grouping and Partitioning Example without Streams Map<Currency, List<Transaction>> transbycurr = new HashMap<>(); for (Transaction transaction : transactions) { Currency currency = transaction.getcurrency(); List<Transaction> transactionsforcurrency = transbycurr.get(currency); if (transactionsforcurrency == null) { transactionsforcurrency = new ArrayList<>(); transbycurr.put(currency, transactionsforcurrency); } transactionsforcurrency.add(transaction); } Louis-Claude Canon MCP Java Streams 73 / 124

74 Collectors Grouping and Partitioning groupingby Map<Currency, List<Transaction>> transbycurr = transactions.stream().collect(groupingby(transaction::getcurrency)); Louis-Claude Canon MCP Java Streams 74 / 124

75

76 Collectors Grouping and Partitioning Second Example Map<Dish.Type, List<Dish>> dishesbytype = menu.stream().collect(groupingby(dish::gettype)); Louis-Claude Canon MCP Java Streams 76 / 124

77

78 Collectors Grouping and Partitioning Grouping without Classification Function public enum CaloricLevel { DIET, NORMAL, FAT } Map<CaloricLevel, List<Dish>> dishesbycaloriclevel = menu.stream().collect(groupingby(dish -> { if (dish.getcalories() <= 400) return CaloricLevel.DIET; else if (dish.getcalories() <= 700) return CaloricLevel.NORMAL; else return CaloricLevel.FAT; } )); Louis-Claude Canon MCP Java Streams 78 / 124

79 Collectors Grouping and Partitioning filtering Result: Map<Dish.Type, List<Dish>> caloricdishesbytype = menu.stream().collect(groupingby(dish::gettype, filtering(dish -> dish.getcalories() > 500, tolist()))); {OTHER=[french fries, pizza], MEAT=[pork, beef], FISH=[]} Louis-Claude Canon MCP Java Streams 79 / 124

80 Collectors Grouping and Partitioning mapping Map<Dish.Type, List<String>> dishnamesbytype = menu.stream().collect(groupingby(dish::gettype, mapping(dish::getname, tolist()))); flatmapping when generating a stream for each object instead of an object only (merge the streams). Louis-Claude Canon MCP Java Streams 80 / 124

81 Collectors Grouping and Partitioning Multilevel Grouping Map<Dish.Type, Map<CaloricLevel, List<Dish>>> res = menu.stream().collect( groupingby(dish::gettype, groupingby(dish -> { if (dish.getcalories() <= 400) return CaloricLevel.DIET; else if (dish.getcalories() <= 700) return CaloricLevel.NORMAL; else return CaloricLevel.FAT; } ) ) ); Louis-Claude Canon MCP Java Streams 81 / 124

82

83 Collectors Grouping and Partitioning collectingandthen To apply a specific function on the result of the collection operation: Map<Dish.Type, Dish> mostcaloricbytype = menu.stream().collect(groupingby(dish::gettype, collectingandthen( maxby(comparingint(dish::getcalories)), Optional::get))); Louis-Claude Canon MCP Java Streams 83 / 124

84

85 Collectors Grouping and Partitioning partitioningby Map<Boolean, List<Dish>> partitionedmenu = menu.stream().collect(partitioningby(dish::isvegetarian)); Always generate lists for both true and false. Special map optimized for two values. Louis-Claude Canon MCP Java Streams 85 / 124

86 Collectors Custom Collector Summary and References Louis-Claude Canon MCP Java Streams 86 / 124 Outline Concepts Stream Operations Special Streams Collectors collect Reducing and Summarizing Grouping and Partitioning Custom Collector Parallel Data Processing

87 Collectors Custom Collector Collector Interface public interface Collector<T, A, R> { Supplier<A> supplier(); BiConsumer<A, T> accumulator(); Function<A, R> finisher(); BinaryOperator<A> combiner(); Set<Characteristics> characteristics(); } T stream element A accumulator element (temporary/intermediate result) R final returned element Example: tolist Louis-Claude Canon MCP Java Streams 87 / 124

88 Collectors Custom Collector Supplier Initial value for the accumulator: public Supplier<List<T>> supplier() { return ArrayList::new; } Louis-Claude Canon MCP Java Streams 88 / 124

89 Collectors Custom Collector Accumulator Takes an element and merge it with the accumulator: public BiConsumer<List<T>, T> accumulator() { return List::add; } Louis-Claude Canon MCP Java Streams 89 / 124

90 Collectors Custom Collector Finisher Convert the accumulator to the final result: public Function<List<T>, List<T>> finisher() { return Function.identity(); } Louis-Claude Canon MCP Java Streams 90 / 124

91

92 Collectors Custom Collector Combiner Combine two accumulators: public BinaryOperator<List<T>> combiner() { return (list1, list2) -> { list1.addall(list2); return list1; } } Allows the parallelization: each core processes a part of the stream, which leads to a set of accumulators that are combined. Louis-Claude Canon MCP Java Streams 92 / 124

93

94 Collectors Custom Collector Characteristics UNORDERED stream order is not important CONCURRENT can be parallelized (on all streams if UNORDERED, on unordered streams only otherwise) IDENTITY_FINISH finisher is the identity function public Set<Characteristics> characteristics() { return Collections.unmodifiableSet(EnumSet.of( IDENTITY_FINISH, CONCURRENT)); } Louis-Claude Canon MCP Java Streams 94 / 124

95 Collectors Custom Collector Alternative List<Dish> dishes = menustream.collect(arraylist::new, List::add, List::addAll); Louis-Claude Canon MCP Java Streams 95 / 124

96 Parallel Data Processing Parallel Streams Outline Concepts Stream Operations Special Streams Collectors Parallel Data Processing Parallel Streams Fork/join Framework Spliterator Summary and References Louis-Claude Canon MCP Java Streams 96 / 124

97 Parallel Data Processing Parallel Streams Thread Logic Parallelizing a processing code requires to: 1. split the data structure into chunks; 2. assign each chunk to a different thread; 3. synchronize the executions and combine the results. Parallel streams hide this logic but each step can be customized. Louis-Claude Canon MCP Java Streams 97 / 124

98 Parallel Data Processing Parallel Streams parallel and sequential A stream is either: sequential (by default or with the method sequential) or parallel (with the method parallel). The last call in the stream construction wins. LongStream.rangeClosed(1, N).parallel().sum(); Louis-Claude Canon MCP Java Streams 98 / 124

99

100 Parallel Data Processing Parallel Streams Settings the Number of Threads In the common pool, there is by default one less threads as there are processors (to let room for the main thread). With the command line: java -Djava.util.concurrent.ForkJoinPool.common. parallelism=3 In the code: System.setProperty( "java.util.concurrent.forkjoinpool.common. parallelism", "3"); ForkJoinPool forkjoinpool = new ForkJoinPool(3); forkjoinpool.submit( () -> stream.terminalop() ).get(); Louis-Claude Canon MCP Java Streams 100 / 124

101 Parallel Data Processing Parallel Streams Pitfalls Necessary to understand the semantics to predict how the stream will be parallelized. Performance issue: Stream.iterate(1L, i -> i + 1).limit(n).parallel().reduce(0L, Long::sum); Correctness issue: Accumulator accumulator = new Accumulator(); LongStream.rangeClosed(1, n).foreach(accumulator::add); Louis-Claude Canon MCP Java Streams 101 / 124

102 Parallel Data Processing Parallel Streams Measuring Performance Critical to test performance (even if code is well understood). Speed-up: ratio of sequential execution time over parallel execution time. With n cores, the ideal speed-up is n. Micro-benchmarking is difficult: Cold-start (warm-up). Variability (multiple iterations) Java Microbenchmark Harness (JMH). Louis-Claude Canon MCP Java Streams 102 / 124

103 Parallel Data Processing Parallel Streams General Advices Avoid boxing (prefer primitive stream specialization). limit, findfirst rely on the order and require synchronization. findany and unordered may solve this issue. Avoid parallelizing small streams with few data or processing. Prefer underlying data source that can be split efficiently (e.g. ArrayList, rangeclosed, and streams that convey their sizes). Be cautious of costly combiner method. Measure and test. Louis-Claude Canon MCP Java Streams 103 / 124

104 Parallel Data Processing Fork/join Framework Outline Concepts Stream Operations Special Streams Collectors Parallel Data Processing Parallel Streams Fork/join Framework Spliterator Summary and References Louis-Claude Canon MCP Java Streams 104 / 124

105 Parallel Data Processing Fork/join Framework ForkJoinPool Parallel streams rely on the class ForkJoinPool, which is an implementation of ExecutorService. Split tasks and distribute them on a pool of threads. Louis-Claude Canon MCP Java Streams 105 / 124

106

107 Parallel Data Processing Fork/join Framework Work-Stealing Creation of more tasks than available threads because tasks may be heterogeneous. Each thread starts by splitting tasks, putting them in their execution queue. Once a task is small enough, it is executed. When a thread finishes all tasks in its execution queue, it steals waiting tasks from the other queues. Louis-Claude Canon MCP Java Streams 107 / 124

108

109 Parallel Data Processing Spliterator Outline Concepts Stream Operations Special Streams Collectors Parallel Data Processing Parallel Streams Fork/join Framework Spliterator Summary and References Louis-Claude Canon MCP Java Streams 109 / 124

110 Parallel Data Processing Spliterator Interface Splitable iterator: define how to split the stream that is iterated. public interface Spliterator<T> { boolean tryadvance(consumer<? super T> action); Spliterator<T> trysplit(); long estimatesize(); int characteristics(); } Louis-Claude Canon MCP Java Streams 110 / 124

111 Parallel Data Processing Spliterator trysplit Logic for splitting data stream. public Spliterator trysplit() { if (currentsize < MIN_SIZE) return null; long half = currentsize / 2; currentsize -= half; return new Spliterator(half); } Louis-Claude Canon MCP Java Streams 111 / 124

112

113 Parallel Data Processing Spliterator tryadvance Feeds the consumer with the next stream element. public boolean tryadvance(consumer action) { action.accept(nextvalue()); currentsize--; return currentsize!= 0; } Louis-Claude Canon MCP Java Streams 113 / 124

114 Parallel Data Processing Spliterator estimatesize The size can be estimated (used for balancing work). public long estimatesize() { return currentsize; } Louis-Claude Canon MCP Java Streams 114 / 124

115 Parallel Data Processing Spliterator characteristics Information to control or optimize the usage of the spliterator. public int characteristics() { return ORDERED + DISTINCT + SORTED + CONCURRENT + SIZED + SUBSIZED + NONNULL + IMMUTABLE; } ORDERED the order of the elements must be enforced DISTINCT all elements are distinct SORTED elements are sorted CONCURRENT the modification of the stream source is thread-safe SIZE the estimated size is precise SUBSIZED all created spliterators are SIZED NON-NULL all elements are non-null IMMUTABLE the stream source cannot be modified Louis-Claude Canon MCP Java Streams 115 / 124

116 Summary and References Outline Concepts Stream Operations Special Streams Collectors Parallel Data Processing Summary and References Louis-Claude Canon MCP Java Streams 116 / 124

117 Summary and References Stream Concepts A stream is a sequence of elements from a source that supports data-processing operations. Streams make use of internal iteration: the iteration is abstracted away through operations such as filter, map, and sorted. There are two types of stream operations: intermediate and terminal operations. Intermediate operations such as filter and map return a stream and can be chained together. They are used to set up a pipeline of operations but do not produce any result. Terminal operations such as foreach and count return a non-stream value and process a stream pipeline to return a result. The elements of a stream are computed on demand ( lazily ). Louis-Claude Canon MCP Java Streams 117 / 124

118 Summary and References Stream Operations The Streams API lets you express complex data processing queries. You can filter and slice a stream using the filter, distinct, takewhile (Java 9), dropwhile (Java 9), skip, and limit methods. You can extract or transform elements of a stream with map and flatmap. You can find elements in a stream using findfirst and findany. You can match a given predicate in a stream using the allmatch, nonematch, and anymatch methods. These methods make use of short-circuiting: a computation stops as soon as a result is found; there is no need to process the whole stream. You can combine all elements of a stream iteratively to produce a result using the reduce method, for example, to calculate the sum or find the maximum of a stream. Some operations such as filter and map are stateless: they do not store any state. Some operations such as reduce store state to calculate a value: they are stateful. Louis-Claude Canon MCP Java Streams 118 / 124

119 Summary and References Special Streams There are three primitive specializations of streams: IntStream, DoubleStream, and LongStream. Their operations are also specialized accordingly. Streams can be created not only from a collection but also from values, arrays, files, and specific methods such as iterate and generate. An infinite stream has an infinite number of elements (for example all possible strings). This is possible because the elements of a stream are only produced on demand. You can get a finite stream from an infinite stream using methods such as limit. Louis-Claude Canon MCP Java Streams 119 / 124

120 Summary and References Collectors collect is a terminal operation that takes as argument various recipes (called collectors) for accumulating the elements of a stream into a summary result. Predefined collectors include reducing and summarizing stream elements into a single value, such as calculating the minimum, maximum, or average. Predefined collectors let you group elements of a stream with groupingby and partition elements of a stream with partitioningby. Collectors compose effectively to create multilevel groupings, partitions, and reductions. You can develop your own collectors by implementing the methods defined in the Collector interface. Louis-Claude Canon MCP Java Streams 120 / 124

121 Summary and References Parallel Data Processing Internal iteration allows you to process a stream in parallel without the need to explicitly use and coordinate different threads in your code. Behavior and performance of parallel code can be counterintuitive, it is always necessary to check and measure them. Parallel execution can improve performance, especially when the number of elements is large or the processing is time consuming. From a performance point of view, using the right data structure (e.g. primitive streams) is more important than trying parallelization. The fork/join framework is used to recursively split streams and distribute them with a work-stealing approach. Spliterator s define how a parallel stream can split the data it traverses. Louis-Claude Canon MCP Java Streams 121 / 124

122 Summary and References Official Documentation Documentation of package stream Documentation of interface Stream Documentation of interface Collector Documentation of class Collectors Documentation of interface Spliterator Louis-Claude Canon MCP Java Streams 122 / 124

123 Summary and References Modern Java in Action, 2nd edition, 2018 Covers recent Java additions (Java 8, 9 and 10). Presents streams, API improvements, lambdas, reactive programming and functional-style Java programming. Louis-Claude Canon MCP Java Streams 123 / 124

124 Summary and References Demonstration Compute the sum of all square numbers between 1 and n. IntStream.rangeClosed(1, n).filter(i -> { var x = Math.floor(Math.sqrt(i)); return x * x == i; }).sum() Louis-Claude Canon MCP Java Streams 124 / 124