This article describes how to do concurrent programming with Java. It covers the concepts of parallel programming, immutability, threads, the executor framework (thread pools), futures, callables and the fork-join framework.
Table of Contents
Concurrency is the ability to run several programs or several parts of a program in parallel. If a time consuming task can be performed asynchronously or in parallel, this improve the throughput and the interactivity of the program.
A modern computer has several CPU's or several cores within one CPU. The ability to leverage these multi-cores can be the key for a successful high-volume application.
A process runs independently and isolated of other processes. It cannot directly access shared data in other processes. The resources of the process, e.g. memory and CPU time, are allocated to it via the operating system.
A thread is a so called lightweight process. It has its own call stack, but can access shared data of other threads in the same process. Every thread has its own memory cache. If a thread reads shared data it stores this data in its own memory cache. A thread can re-read the shared data.
A Java application runs by default in one process. Within a Java application you work with several threads to achieve parallel processing or asynchronous behavior.
Within a Java application you work with several threads to achieve parallel processing or asynchronous behavior.
Concurrency promises to perform certain task faster as these tasks can be divided into subtasks and these subtasks can be executed in parallel. Of course the runtime is limited by parts of the task which can be performed in parallel.
The theoretical possible performance gain can be calculated by the following rule which is referred to as Amdahl's Law.
If F is the percentage of the program which can not run in parallel and N is the number of processes, then the maximum performance gain is 1/ (F+ ((1-F)/n)).
Threads have their own call stack, but can also access shared data. Therefore you have two basic problems, visibility and access problems.
A visibility problem occurs if thread A reads shared data which is later changed by thread B and thread A is unaware of this change.
An access problem can occur if several thread access and change the same shared data at the same time.
Visibility and access problem can lead to
- Liveness failure: The program does not react anymore due to problems in the concurrent access of data, e.g. deadlocks.
- Safety failure: The program creates incorrect data.
A Java program runs in its own process and by default in one thread. Java supports threads as part of the Java language via the
Thread code. The Java application can create new threads via this class.
Java 1.5 also provides improved support for concurrency with the in the
java.util.concurrent package.
Java provides locks to protect certain parts of the code to be executed by several threads at the same time. The simplest way of locking a certain method or Java class is to define the method or class with the
synchronizedkeyword.
The synchronized keyword in Java ensures:
- that only a single thread can execute a block of code at the same time
- that each thread entering a synchronized block of code sees the effects of all previous modifications that were guarded by the same lock
Synchronization is necessary for mutually exclusive access to blocks of and for reliable communication between threads.
You can use the synchronized keyword for the definition of a method. This would ensure that only one thread can enter this method at the same time. Another threads which is calling this method would wait until the first threads leaves this method.
public synchronized void critial() {
// some thread critical stuff
// here
}
You can also use the
synchronized keyword to protect blocks of code within a method. This block is guarded by a key, which can be either a string or an object. This key is called the lock. All code which is protected by the same lock can only be executed by one thread at the same time
For example the following datastructure will ensure that only one thread can access the inner block of the
add() andnext() methods.package de.vogella.pagerank.crawler; import java.util.ArrayList; import java.util.List;/** * Data structure for a web crawler. Keeps track of the visited sites and keeps * a list of sites which needs still to be crawled. * * @author Lars Vogel * */public class CrawledSites { private ListcrawledSites = new ArrayList (); private List linkedSites = new ArrayList (); public void add(String site) { synchronized (this) { if (!crawledSites.contains(site)) { linkedSites.add(site); } } } /** * Get next site to crawl. Can return null (if nothing to crawl) */public String next() { if (linkedSites.size() == 0) { return null; } synchronized (this) { // Need to check again if size has changed if (linkedSites.size() > 0) { String s = linkedSites.get(0); linkedSites.remove(0); crawledSites.add(s); return s; } return null; } } }
If a variable is declared with the volatile keyword then it is guaranteed that any thread that reads the field will see the most recently written value. The volatile keyword will not perform any mutual exclusive lock on the variable.
As of Java 5 write access to a volatile variable will also update non-volatile variables which were modified by the same thread. This can also be used to update values within a reference variable, e.g. for a volatile variable person. In this case you must use a temporary variable person and use the setter to initialize the variable and then assign the temporary variable to the final variable. This will then make the address changes of this variable and the values visible to other threads.
The Java memory model describes the communication between the memory of the threads and the main memory of the application.
It defines the rules how changes in the memory done by threads are propagated to other threads. The Java memory model also defines the situations in which a thread re-fresh its own memory from the main memory.
It also describes which operations are atomic and the ordering of the operations.
An atomic operation is an operation which is performed as a single unit of work without the possibility of interference from other operations.
The Java language specification guarantees that reading or writing a variable is an atomic operation(unless the variable is of type
long or double). Operations variables of type long or double are only atomic if they declared with the volatile keyword. .
Assume
i is defined as int. The i++ (increment) operation it not an atomic operation in Java. This also applies for the other numeric types, e.g. long. etc).
The
i++ operation first reads the value which is currently stored in i (atomic operations) and then it adds one to it (atomic operation). But between the read and the write the value of i might have changed.
Since Java 1.5 the java language provides atomic variables, e.g. AtomicInteger or AtomicLong which provide methods like
getAndDecrement(), getAndIncrement() and getAndSet() which are atomic.
The simplest way to avoid problems with concurrency is to share only immutable data between threads. Immutable data is data which cannot changed.
To make a class immutable make
- all its fields final
- the class declared as final
- the this reference is not allowed to escape during construction
- Any fields which refer to mutable data objects are
- private
- have no setter method
- they are never directly returned of otherwise exposed to a caller
- if they are changed internally in the class this change is not visible and has no effect outside of the class
An immutable class may have some mutable data which is uses to manages its state but from the outside this class nor any attribute of this class can get changed.
For all mutable fields, e.g. Arrays, that are passed from the outside to the class during the construction phase, the class needs to make a defensive-copy of the elements to make sure that no other object from the outside still can change the data
You must protected your classes from calling code. Assume that calling code will do its best to change your data in a way you didn't expect it. While this is especially true in case of immutable data it is also true for non-immutable data which you still not expect that this data is changed outside your class.
To protect your class against that you should copy data you receive and only return copies of data to calling code.
The following example creates a copy of a list (ArrayList) and returns only the copy of the list. This way the client of this class cannot remove elements from the list.
package de.vogella.performance.defensivecopy;
import java.util.ArrayList;
import java.util.Collections;
import java.util.List;
public class MyDataStructure {
List list = new ArrayList();
public void add(String s) {
list.add(s);
}
/**
* Makes a defensive copy of the List and return it
* This way cannot modify the list itself
*
* @return List
*/
public List getList() {
return Collections.unmodifiableList(list);
}
}
The base means for concurrency are is the
java.lang.Threads class. A Thread executes an object of typejava.lang.Runnable.Runnable is an interface with defines the run() method. This method is called by the Thread object and contains the work which should be done. Therefore the "Runnable" is the task to perform. The Thread is the worker who is doing this task.
The following demonstrates a task (Runnable) which counts the sum of a given range of numbers. Create a new Java project called
de.vogella.concurrency.threads for the example code of this section.package de.vogella.concurrency.threads;/** * MyRunnable will count the sum of the number from 1 to the parameter * countUntil and then write the result to the console. * * MyRunnable is the task which will be performed * * @author Lars Vogel * */public class MyRunnable implements Runnable { private final long countUntil; MyRunnable(long countUntil) { this.countUntil = countUntil; } @Override public void run() { long sum = 0; for (long i = 1; i < countUntil; i++) { sum += i; } System.out.println(sum); } }
The following example demonstrate the usage of the
Thread and the Runnable class.package de.vogella.concurrency.threads;
import java.util.ArrayList;
import java.util.List;
public class Main {
public static void main(String[] args) {
// We will store the threads so that we can check if they are done
List threads = new ArrayList();
// We will create 500 threads
for (int i = 0; i < 500; i++) {
Runnable task = new MyRunnable(10000000L + i);
Thread worker = new Thread(task);
// We can set the name of the thread
worker.setName(String.valueOf(i));
// Start the thread, never call method run() direct
worker.start();
// Remember the thread for later usage
threads.add(worker);
}
int running = 0;
do {
running = 0;
for (Thread thread : threads) {
if (thread.isAlive()) {
running++;
}
}
System.out.println("We have " + running + " running threads. ");
} while (running > 0);
}
}
Using the
Thread class directly has the following disadvantages.- Creating a new thread causes some performance overhead
- Too many threads can lead to reduced performance, as the CPU needs to switch between these threads.
- You cannot easily control the number of threads, therefore you may run into out of memory errors due to too many threads.
The
java.util.concurrent package offers improved support for concurrency compared to the direct usage ofThreads. This package is described in the next section.
Thread pools manage a pool of worker threads. The thread pools contains a work queue which holds tasks waiting to get executed.
A thread pool can be described as a collection of
Runnable objects (work queue) and a connections of running threads. These threads are constantly running and are checking the work query for new work. If there is new work to be done they execute this Runnable. The Thread class itself provides a method, e.g. execute(Runnable r) to add a new Runnable object to the work queue.
The Executor framework provides example implementation of the java.util.concurrent.Executor interface, e.g. Executors.newFixedThreadPool(int n) which will create n worker threads. The ExecutorService adds life cycle methods to the Executor, which allows to shutdown the Executor and to wait for termination.
Create again the Runnable.
package de.vogella.concurrency.threadpools;/** * MyRunnable will count the sum of the number from 1 to the parameter * countUntil and then write the result to the console. * * MyRunnable is the task which will be performed * * @author Lars Vogel * */public class MyRunnable implements Runnable { private final long countUntil; MyRunnable(long countUntil) { this.countUntil = countUntil; } @Override public void run() { long sum = 0; for (long i = 1; i < countUntil; i++) { sum += i; } System.out.println(sum); } }
Now you run your runnables with the executor framework.
package de.vogella.concurrency.threadpools;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
public class Main {
private static final int NTHREDS = 10;
public static void main(String[] args) {
ExecutorService executor = Executors.newFixedThreadPool(NTHREDS);
for (int i = 0; i < 500; i++) {
Runnable worker = new MyRunnable(10000000L + i);
executor.execute(worker);
}
// This will make the executor accept no new threads
// and finish all existing threads in the queue
executor.shutdown();
// Wait until all threads are finish
executor.awaitTermination();
System.out.println("Finished all threads");
}
}
In case the threads should return some value (result-bearing threads) then you can use the
java.util.concurrent.Callable class.
The code examples for this section are created in a Java project called de.vogella.concurrency.callables.
The executor framework presented in the last chapter works with
Runnables. Runnable do not return result.
In case you expect your threads to return a computed result you can use
java.util.concurrent.Callable. TheCallable object allows to return values after completion.
The
Callable object uses generics to define the type of object which is returned.
If you submit a
Callable object to an Executor the framework returns an object of typejava.util.concurrent.Future. This Future object can be used to check the status of a Callable and to retrieve the result from the Callable.
On the
Executor you can use the method submit to submit a Callable and to get a future. To retrieve the result of the future use the get() method.package de.vogella.concurrency.callables;
import java.util.concurrent.Callable;
public class MyCallable implements Callable {
@Override
public Long call() throws Exception {
long sum = 0;
for (long i = 0; i <= 100; i++) {
sum += i;
}
return sum;
}
}
package de.vogella.concurrency.callables;
import java.util.ArrayList;
import java.util.List;
import java.util.concurrent.Callable;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.Future;
public class CallableFutures {
private static final int NTHREDS = 10;
public static void main(String[] args) {
ExecutorService executor = Executors.newFixedThreadPool(NTHREDS);
List> list = new ArrayList>();
for (int i = 0; i < 20000; i++) {
Callable worker = new MyCallable();
Future submit = executor.submit(worker);
list.add(submit);
}
long sum = 0;
System.out.println(list.size());
// now retrieve the result
for (Future future : list) {
try {
sum += future.get();
} catch (InterruptedException e) {
e.printStackTrace();
} catch (ExecutionException e) {
e.printStackTrace();
}
}
System.out.println(sum);
executor.shutdown();
}
}
Java 5.0 provides supports for additional atomic operations. This allows to develop algorithm which are non-blocking algorithm, e.g. which do not require synchronization, but are based on low-level atomic hardware primitives such as compare-and-swap (CAS). A compare-and-swap operation check if the variable has a certain value and if it has this value it will perform this operation.
Non-blocking algorithm are usually much faster then blocking algorithms as the synchronization of threads appears on a much finer level (hardware).
For example this created a non-blocking counter which always increases. This example is contained in the project called de.vogella.concurrency.nonblocking.counter.
package de.vogella.concurrency.nonblocking.counter;
import java.util.concurrent.atomic.AtomicInteger;
public class Counter {
private AtomicInteger value = new AtomicInteger();
public int getValue(){
return value.get();
}
public int increment(){
return value.incrementAndGet();
}
// Alternative implementation as increment but just make the
// implementation explicit
public int incrementLongVersion(){
int oldValue = value.get();
while (!value.compareAndSet(oldValue, oldValue+1)){
oldValue = value.get();
}
return oldValue+1;
}
}
And a test.
package de.vogella.concurrency.nonblocking.counter;
import java.util.ArrayList;
import java.util.HashSet;
import java.util.List;
import java.util.Set;
import java.util.concurrent.Callable;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.Future;
public class Test {
private static final int NTHREDS = 10;
public static void main(String[] args) {
final Counter counter = new Counter();
List> list = new ArrayList>();
ExecutorService executor = Executors.newFixedThreadPool(NTHREDS);
for (int i = 0; i < 500; i++) {
Callable worker = new Callable() {
@Override
public Integer call() throws Exception {
int number = counter.increment();
System.out.println(number);
return number ;
}
};
Future submit= executor.submit(worker);
list.add(submit);
}
// This will make the executor accept no new threads
// and finish all existing threads in the queue
executor.shutdown();
// Wait until all threads are finish
while (!executor.isTerminated()) {
}
Set set = new HashSet();
for (Future future : list) {
try {
set.add(future.get());
} catch (InterruptedException e) {
e.printStackTrace();
} catch (ExecutionException e) {
e.printStackTrace();
}
}
if (list.size()!=set.size()){
throw new RuntimeException("Double-entries!!!");
}
}
}
The interesting part is how the
incrementAndGet() method is implemented. It uses a CAS operation.public final int incrementAndGet() {
for (;;) {
int current = get();
int next = current + 1;
if (compareAndSet(current, next))
return next;
}
}
The JDK itself makes more and more use of non-blocking algorithms to increase performance for every developer. Developing correct non-blocking algorithm is not a trivial task.
For more information on non-blocking algorithm, e.g. examples for a non-blocking Stack and non-block LinkedList, please see http://www.ibm.com/developerworks/java/library/j-jtp04186/index.html
Java 7 introduce a new parallel mechanism for compute intensive tasks, the fork-join framework. The fork-join framework allows you to distribute a certain task on several workers and then wait for the result.
E For Java 6.0 you can download the package (jsr166y) from Download site
For testing create the Java project "de.vogella.performance.forkjoin". If you are not using Java 7 you also need to
jsr166y.jar to the classpath.
Create first a
algorithm package and then the following class.package algorithm; import java.util.Random;/** * * This class defines a long list of integers which defines the problem we will * later try to solve * */public class Problem { private final int[] list = new int[2000000]; public Problem() { Random generator = new Random(19580427); for (int i = 0; i < list.length; i++) { list[i] = generator.nextInt(500000); } } public int[] getList() { return list; } }
Define now the
Solver class as shown in the following example coding.package algorithm;
import java.util.Arrays;
import jsr166y.forkjoin.RecursiveAction;
public class Solver extends RecursiveAction {
private int[] list;
public long result;
public Solver(int[] array) {
this.list = array;
}
@Override
protected void compute() {
if (list.length == 1) {
result = list[0];
} else {
int midpoint = list.length / 2;
int[] l1 = Arrays.copyOfRange(list, 0, midpoint);
int[] l2 = Arrays.copyOfRange(list, midpoint, list.length);
Solver s1 = new Solver(l1);
Solver s2 = new Solver(l2);
forkJoin(s1, s2);
result = s1.result + s2.result;
}
}
}
Now define a small test class for testing it efficiency.
package testing;
import jsr166y.forkjoin.ForkJoinExecutor;
import jsr166y.forkjoin.ForkJoinPool;
import algorithm.Problem;
import algorithm.Solver;
public class Test {
public static void main(String[] args) {
Problem test = new Problem();
// check the number of available processors
int nThreads = Runtime.getRuntime().availableProcessors();
System.out.println(nThreads);
Solver mfj = new Solver(test.getList());
ForkJoinExecutor pool = new ForkJoinPool(nThreads);
pool.invoke(mfj);
long result = mfj.getResult();
System.out.println("Done. Result: " + result);
long sum = 0;
// check if the result was ok
for (int i = 0; i < test.getList().length; i++) {
sum += test.getList()[i];
}
System.out.println("Done. Result: " + sum);
}
}
A concurrent application has the risk of a deadlock. A set of processes are deadlocked if all processes are waiting for an event which another process in the same set has to cause.
For example if thread A waits for a lock on object Z which thread B holds and thread B wait for a look on object Y which is hold be process A then these two processes are looked and cannot continue in their processing.
