You can have a data race (also named race condition) in your application when you have two or more tasks writing a shared variable outside a critical section—that's to say, without using any synchronization mechanisms.

Under these circumstances, the final result of your application may depend on the order of execution of the tasks. Look at the following example:

package com.packt.java.concurrency;

public class Account {

  private float balance;

  public void modify (float difference) {

    float value=this.balance;
    this.balance=value+difference;
  }

}

Imagine that two different tasks execute the modify() method in the same Account object. Depending on the order of execution of the sentences in the tasks, the final result can vary. Suppose that the initial balance is 1000 and the two tasks call the modify() method with 1000 as a parameter. The final result should be 3000, but if both tasks execute the first sentence at the same time and then the second sentence at the same time, the final result will be 2000. As you can see, the modify() method is not atomic and the Account class is not thread-safe.

There is a deadlock in your concurrent application when there are two or more tasks waiting for a shared resource that must be free from the other, so none of them will get the resources they need and will be blocked indefinitely. It happens when four conditions happen simultaneously in the system. They are Coffman's conditions, which are as follows:

There exist some mechanisms that you can use to avoid deadlocks:

A livelock occurs when you have two tasks in your systems that are always changing their states due to the actions of the other. Consequently, they are in a loop of state changes and unable to continue.

For example, you have two tasks—Task 1 and Task 2—and both need two resources: Resource 1 and Resource 2. Suppose that Task 1 has a lock on Resource 1, and Task 2 has a lock on Resource 2. As they are unable to gain access to the resource they need, they free their resources and begin the cycle again. This situation can continue indefinitely, so the tasks will never end their execution.

Resource starvation occurs when you have a task in your system that never gets a resource that it needs to continue with its execution. When there is more than one task waiting for a resource and the resource is released, the system has to choose the next task that can use it. If your system has not got a good algorithm, it can have threads that are waiting for a long time for the resource.

Fairness is the solution to this problem. All the tasks that are waiting for a resource must have the resource in a given period of time. An option is to implement an algorithm that takes into account the time that a task has been waiting for a resource when it chooses the next task that will hold a resource. However, fair implementation of locks requires additional overhead, which may lower your program throughput.

Priority inversion occurs when a low-priority task holds a resource that is needed by a high-priority task, so the low-priority task finishes its execution before the high-priority task.

Our starting point to implement a concurrent algorithm will be a sequential version of it. Of course, we can design a concurrent algorithm from scratch, but I think that a sequential version of the algorithm will give us two advantages:

In this step, we are going to analyze the sequential version of the algorithm to look for the parts of its code that can be executed in a parallel way. We should pay special attention to those parts that are executed most of the time or that execute more code because, by implementing a concurrent version of those parts, we're going to get a greater performance improvement.

Good candidates for this process are loops where one step is independent of the other steps or portions of code that are independent of other parts of the code (for example, an algorithm to initialize an application that opens the connections with the database, loads the configuration files, initialize some objects. All the previous tasks are independent of each other).

Once you know what parts of the code you are going to parallelize, you have to decide how to do that parallelization.

The changes in the code will affect two main parts of the application:

You can take two different approaches to accomplish this task:

Another important point to keep in mind is the granularity of your solution. The objective of implementing a parallel version of an algorithm is to achieve improved performance, so you should use all the available processors or cores. On the other hand, when you use a synchronization mechanism, you introduce some extra instructions that must be executed. If you split the algorithm into a lot of small tasks (fine-grained granularity), the extra code introduced by the synchronization can provoke performance degradation. If you split the algorithm into fewer tasks than cores (coarse-grained granularity), you are not taking advantage of all the resources. Also, you must take into account the work every thread must do, especially if you implement a fine-grained granularity. If you have a task longer than the rest, that task will determine the execution time of the application. You have to find the equilibrium between these two points.

The next step is to implement the parallel algorithm using a programming language and, if it's necessary, a thread library. In the examples of this book, you are going to use Java to implement all the algorithms.

After finishing the implementation, you have to test the parallel algorithm. If you have a sequential version of the algorithm, you can compare the results of both algorithms to verify that your parallel implementation is correct.

Testing and debugging a parallel implementation are difficult tasks because the order of execution of the different tasks of the application is not guaranteed. In Chapter 11, Testing and Monitoring Concurrent Applications, you will learn tips, tricks, and tools to do these tasks efficiently.

The last step is to compare the throughput of the parallel and the sequential algorithms. If the results are not as expected, you must review the algorithm, looking for the cause of the bad performance of the parallel algorithm.

You can also test different parameters of the algorithm (for example, granularity or number of tasks) to find the best configuration.

There are different metrics to measure the possible performance improvement you can obtain parallelizing an algorithm. The three most popular metrics are:

In this section, you learned some important issues you have to take into account when you want to parallelize a sequential algorithm.

First of all, not every algorithm can be parallelized. For example, if you have to execute a loop where the result of an iteration depends on the result of the previous iteration, you can't parallelize that loop. Recurrent algorithms are another example of algorithms that can be parallelized for a similar reason.

Another important thing you have to keep in mind is that the sequential version of an algorithm with better performance can be a bad starting point to parallelize it. If you start parallelizing an algorithm and you find yourself in trouble because you don't easily find independent portions of the code, you have to look for other versions of the algorithm and verify that the version can be parallelized in an easier way.

Finally, when you implement a concurrent application (from scratch or based on a sequential algorithm), you must take into account the following points:

The basic classes of the Java concurrency API are:

The Java concurrency API includes different synchronization mechanisms that allow you to:

The following mechanisms are considered to be the most important synchronization mechanisms:

The executor framework is a mechanism that allows you to separate thread creation and management for the implementation of concurrent tasks. You don't have to worry about the creation and management of threads, only about creating tasks and sending them to the executor. The main classes involved in this framework are:

The Fork/Join framework defines a special kind of executor specialized in the resolution of problems with the divide and conquer technique. It includes a mechanism to optimize the execution of the concurrent tasks that solve these kinds of problems. Fork/Join is specially tailored for fine-grained parallelism as it has a very low overhead in order to place the new tasks into the queue and take queued tasks for execution. The main classes and interfaces involved in this framework are:

Streams and Lambda expressions are maybe the two most important new features of the Java 8 version. Streams have been added as a method in the Collection interface and other data sources and allow processing all elements of a data structure, generating new structures, filtering data and implementing algorithms using the map and reduce technique.

A special kind of stream is a parallel stream which realizes its operations in a parallel way. The most important elements involved in the use of parallel streams are:

Normal data structures of the Java API (ArrayList, Hashtable, and so on) are not ready to work in a concurrent application unless you use an external synchronization mechanism. If you use it, you will be adding a lot of extra computing time to your application. If you don't use it, it's probable that you will have race conditions in your application. If you modify them from several threads and a race condition occurs, you may experience various exceptions thrown (such as, ConcurrentModificationException and ArrayIndexOutOfBoundsException), there may be silent data loss or your program may even stuck in an endless loop.

The Java concurrency API includes a lot of data structures that can be used in concurrent applications without risk. We can classify them in two groups:

These are some of the data structures:

This design pattern explains how to implement the situation where a task has to notify an event to another task. The easiest way to implement this pattern is with a semaphore or a mutex, using the ReentrantLock or Semaphore classes of the Java language or even the wait() and notify() methods included in the Object class.

See the following example:

public void task1() {
  section1();
  commonObject.notify();
}

public void task2() {
  commonObject.wait();
  section2();
}

Under these circumstances, the section2() method will always be executed after the section1() method.

This design pattern is a generalization of the Signaling pattern. In this case, the first task waits for an event of the second task and the second task waits for an event of the first task. The solution is similar to that of Signaling, but in this case you must use two objects instead of one.

See the following example:

public void task1() {
  section1_1();
  commonObject1.notify();
  commonObject2.wait();
  section1_2();
}
public void task2() {
  section2_1();
  commonObject2.notify();
  commonObject1.wait();
  section2_2();
}

Under these circumstances, section2_2() always will be executed after section1_1() and section1_2() after section2_1(), take into account that, if you put the call to the wait() method before the call to the notify() method, you will have a deadlock.

A mutex is a mechanism that you can use to implement a critical section ensuring mutual exclusion. That is to say, only one task can execute the portion of code protected by the mutex at one time. In Java, you can implement a critical section using the synchronized keyword (that allows you to protect a portion of code or a full method), the ReentrantLock class, or the Semaphore class.

Look at the following example:

public void task() {
  preCriticalSection();
  lockObject.lock() // The critical section begins
  criticalSection();
  lockObject.unlock(); // The critical section ends
  postCriticalSection();
}

The Multiplex design pattern is a generalization of the mutex. In this case, a determined number of tasks can execute the critical section at once. It is useful, for example, when you have multiple copies of a resource. The easiest way to implement this design pattern in Java is using the Semaphore class initialized to the number of tasks that can execute the critical section at once.

Look at the following example:

public void task() {
  preCriticalSection();
  semaphoreObject.acquire();
  criticalSection();
  semaphoreObject.release();
  postCriticalSection();
}

This design pattern explains how to implement the situation where you need to synchronize some tasks at a common point. None of the tasks can continue with their execution until all the tasks have arrived at the synchronization point. The Java concurrency API provides the CyclicBarrier class, which is an implementation of this design pattern.

Look at the following example:

public void task() {
  preSyncPoint();
  barrierObject.await();
  postSyncPoint();
}

This design pattern provides a solution to the problem that occurs when you acquire a lock and then check for a condition. If the condition is false, you have had the overhead of acquiring the lock ideally. An example of this situation is the lazy initialization of objects. If you have a class implementing the Singleton design pattern, you may have some code like this:

public class Singleton{
  private static Singleton reference;
  private static final Lock lock=new ReentrantLock();
  public static Singleton getReference() {
    lock.lock();
    try {
        if (reference==null) {
          reference=new Object();
        }
    } finally {
        lock.unlock();
    }
    return reference;
  }
}

A possible solution can be to include the lock inside the conditions:

public class Singleton{
  private Object reference;
  private Lock lock=new ReentrantLock();
  public Object getReference() {
    if (reference==null) {
      lock.lock();
      try {
          if (reference == null) {
            reference=new Object();
          }
      } finally {
          lock.unlock();
      }
    }
    return reference;
  }
}

This solution still has problems. If two tasks check the condition at once, you will create two objects. The best solution to this problem doesn't use any explicit synchronization mechanism:

public class Singleton {

  private static class LazySingleton {
    private static final Singleton INSTANCE = new Singleton();
  }

  public static Singleton getSingleton() {
    return LazySingleton.INSTANCE;
  }

}

When you protect access to a shared variable with a lock, only one task can access that variable, independently of the operation you are going to perform on it. Sometimes, you will have variables that you modify a few times but read many times. In this circumstance, a lock provides poor performance because all the read operations can be made concurrently without any problem. To solve this problem, there exists the read-write lock design pattern. This pattern defines a special kind of lock with two internal locks: one for read operations and the other for write operations. The behavior of this lock is as follows:

The Java concurrency API includes the class ReentrantReadWriteLock that implements this design pattern. If you want to implement this pattern from scratch, you have to be very careful with the priority between read-tasks and write-tasks. If too many read-tasks exist, write-tasks can be waiting too long.

This design pattern tries to remove the overhead introduced by creating a thread for the task you want to execute. It's formed by a set of threads and a queue of tasks you want to execute. The set of threads usually has a fixed size. When a thread approaches the execution of a task, it doesn't finish its execution; it looks for another task in the queue. If there is another task, it executes it. If not, the thread waits until a task is inserted in the queue, but it's not destroyed.

The Java concurrency API includes some classes that implement the ExecutorService interface, which internally uses a pool of threads.

This design pattern defines how to use global or static variables locally to tasks. When you have a static attribute in a class, all the objects of a class access the same occurrences of the attribute. If you use thread local storage, each thread accesses a different instance of the variable.

The Java concurrency API includes the ThreadLocal class to implement this design pattern.

You can only execute concurrent tasks that are independent of each other. If you have two or more tasks with an order dependency between them, maybe you have no interest in trying to execute them concurrently and including a synchronization mechanism to guarantee the execution order. The tasks will execute in a sequential way, and you will have to overcome the synchronization mechanism. A different situation is when you have a task with some prerequisites, but these prerequisites are independent of each other. In this case, you can execute the prerequisites concurrently and then use a synchronization class to control the execution of the task after completion of all the prerequisites.

Another situation where you can't use concurrency is when you have a loop, and all the steps use data generated in the step before or there is some status information that goes from one step to the next step.

Rich threading APIs, as the Java concurrency API, offer you different classes to implement concurrency in your applications. In the case of Java, you can control the creation and synchronization of threads using the Thread or Lock classes, but it also offers you high-level concurrency objects, such as executors or the Fork/Join framework that allows you to execute concurrent tasks. This high-level mechanism offers you the following benefits:

In summary, for performance and development time reasons, analyze the high-level mechanisms your thread API offers you before implementing your concurrent algorithm.

One of the main objectives when you implement a concurrent algorithm is to take advantage of all the resources of your computer, especially the number of processors or cores. But this number may change over time. Hardware is constantly evolving and its cost becomes lower each year.

When you design a concurrent algorithm using data decomposition, don't presuppose the number of cores or processors that your application will execute on. Get the information about the system dynamically (for example, in Java you can get it with the method Runtime.getRuntime().availableProcessors()) and make your algorithm use that information to calculate the number of tasks it's going to execute. This process will have an overhead over the execution time of your algorithm, but your algorithm will be more scalable.

If you design a concurrent algorithm using task decomposition, the situation can be more difficult. You depend on the number of independent tasks you have in the algorithm and forcing a bigger number of tasks will increment the overhead introduced by synchronization mechanisms, and the global performance of the application can be even worse. Analyze in detail the algorithm to determine whether you can have a dynamic number of tasks or not.

If you need to use a Java library in a concurrent application, read its documentation first to know if it's thread-safe or not. If it's thread-safe, you can use it in your application without any problem. If it's not, you have the following two options:

For example, if you need a List in a concurrent application, you should not use the ArrayList class if you are going to update it from several threads because it's not thread-safe. In this case, you can use a thread-safe class such as ConcurrentLinkedDeque,CopyOnWriteArrayList, or LinkedBlockingDeque. If the class you want to use is not thread-safe, first you must look for a thread-safe alternative. Probably, it will be more optimized to work with concurrency that any alternative that you can implement.

The execution of tasks in a concurrent application when you don't use any synchronization mechanism is nondeterministic. The order in which the tasks are executed and the time each task is in execution before the processor moves on to another task is determined by the scheduler of the operating system. It doesn't care if you observe that the execution order is the same in a number of executions. The next one could be different.

The result of this assumption used to be a data race problem. The final result of your algorithm depends on the execution order of the tasks. Sometimes, the result can be right, but at other times it can be wrong. It can be very difficult to detect the cause of data race conditions, so you must be careful not to forget all the necessary synchronization elements.

Thread local variables are a special kind of variable. Every task will have an independent value for this variable, so you don't need any synchronization mechanism to protect access to this variable.

This can sound a little strange. Every object has its own copy of the attributes of the class, so why do we need the thread local variables? Consider this situation. You create a Runnable task, and you want to execute multiple instances of that task. You can create a Runnable object for each thread you want to execute, but another option is to create a Runnable object and use that object to create all the threads. In the last case, all the threads will have access to the same copy of the attributes of the class except if you use the ThreadLocal class. The ThreadLocal class guarantees you that every thread will access its own instance of the variable without the use of a Lock, a semaphore, or a similar class.

Another situation when you can take advantage of Thread local variables is with static attributes. All instances of a class share the static attributes, but you declare them with the ThreadLocal class. In this case, every thread will have access to its own copy.

Another option you have is to use something like ConcurrentHashMap<Thread, MyType> and use it like var.get(Thread.currentThread()) or var.put(Thread.currentThread(), newValue). Usually, this approach is significantly slower than ThreadLocal because of possible contention (ThreadLocal has no contention at all). It has an advantage though: you can clear the map completely and the value will disappear for every thread; thus, sometimes it's useful to use such an approach.

We can define an algorithm as a sequence of steps to solve a problem. There are different ways to solve the same problem. Some are faster, some use fewer resources, and others fit better with special characteristics of the input data. For example, if you want to order a set of numbers, you can use one of the multiple sorting algorithms that have been implemented.

In a previous section of this chapter, we recommended you use a sequential algorithm as the starting point to implement a concurrent algorithm. There are two main advantages to this approach:

But not every algorithm can be parallelized, at least not so easily. You might think that the best starting point could be the sequential algorithm with the best performance solving the problem you want to parallelize, but this can be a wrong assumption. You should look for an algorithm than can be easily parallelized. Then, you can compare the concurrent algorithm with the sequential one with the best performance to see which offers the best throughput.

One of the main problems you can have in a concurrent application is a data race condition. As we explained before, this happens when two or more tasks modify the data stored in a shared variable and access to that variable is not implemented inside a critical section.

For example, when you work with an object-oriented language such as Java, you implement your application as a collection of objects. Each object has a number of attributes and some methods to read and change the values of the attributes. If some tasks share an object and call to a method to change a value of an attribute of that object and that method is not protected by a synchronization mechanism, you probably will have data inconsistency problems.

There are special kinds of object named immutable objects. Their main characteristic is that you can't modify any attributes after initialization. If you want to modify the value of an attribute, you must create another object. The String class in Java is the best example of immutable objects. When you use an operator (for example, = or +=) that we might think changes the value of a String, you are really creating a new object.

The use of immutable objects in a concurrent application has two very important advantages:

There is a drawback with immutable objects. If you create too many objects, this may affect the throughput and memory use of the application. If you have a simple object without internal data structures, it's usually not a problem to make it immutable. However, making immutable complex objects that incorporate collections of other objects usually leads to serious performance problems.

One of the best mechanisms to avoid a deadlock situation in a concurrent application is to force tasks to get shared resources always in the same order. An easy way to do this is to assign a number to every resource. When a task needs more than one resource, it has to request them in order.

For example, if you have two tasks, T1 and T2, and both need two resources, R1 and R2, you can force both to request first the R1 resource and then the R2 resource. You will never have a deadlock.

On the other hand, if T1 first requests R1 and then R2 and T2 first requests R2 and then R1, you can have a deadlock.

For example, a bad use of this tip is as follows. You have two tasks that need to get two Lock objects. They try to get the locks in different order:

public void operation1() {
  lock1.lock();
  lock2.lock();
  ….
}
public void operation2() {
  lock2.lock();
  lock1.lock();
  …..
}

It's possible that operation1() executes its first sentence and operation2() its first sentence too, so they will be waiting for the other Lock and you will have a deadlock.

You can avoid this simply by getting the locks in the same order. If you change operation2(), you will never have a deadlock as follows:

public void operation2() {
  lock1.lock();
  lock2.lock();
  …..
}

When you have to share data between two or more tasks, you have to use a synchronization mechanism to protect the access to that data and avoid any data inconsistency problems.

Under some circumstances, you can use the volatile keyword and not use a synchronization mechanism. If only one of the tasks modifies the data and the rest of the tasks read it, you can use the volatile keyword without any synchronization or data inconsistency problem. In other scenarios, you need to use a lock, the synchronized keyword, or any other synchronization method.

In Java 5, the concurrency API included a new kind of variable called atomic variables. These variables are classes that support atomic operations on single variables. They include a method, denominated by compareAndSet(oldValue, newValue), that includes a mechanism to detect if assigning to the new value to the variable is done in one step. If the value of the variable is equal to oldValue, it changes it to newValue and returns true. Otherwise, it returns false. There are more methods that work in a similar way, such as getAndIncrement() or getAndDecrement(). These methods are also atomic.

This solution is lock-free; that is to say, it doesn't use locks or any synchronization mechanism, so its performance is better than any synchronized solution.

The most important atomic variables that you can use in Java are:

Locks, as with any other synchronization mechanism, allow you to define a critical section that only one task can execute at a time. While a task is executing the critical section, the other tasks that want to execute it are blocked and have to wait for the liberation of the critical section. The application is working in a sequential way.

You have to pay special attention to the instructions you include in your critical sections because you can degrade the performance of your application without realizing it. You must make your critical section as small as possible, and it must include only the instructions that work on shared data with other tasks, so the time that the application is executing in a sequential way will be minimal.

Avoid executing inside the critical section the code you don't control. For example, you are writing a library that accepts a user-defined Callable, which you need to launch sometimes. You don't know what exactly will be in that Callable. Maybe it blocks input/output, acquires some locks, calls other methods of your library, or just works for a very long time. Thus, whenever possible, try to execute it when your library does not hold any locks. If it's impossible for your algorithm, specify this behavior in your library documentation and possibly specify the limitations to the user-supplied code (for example, it should not take any locks). A good example of such documentation can be found in the compute() method of the ConcurrentHashMap class.

Lazy initialization is a mechanism that delays object creation until the object is used in the application for the first time. Its main advantage is it minimizes the use of memory because you only create the objects that are really needed, but it can be a problem in concurrent applications.

If you have a method that initializes an object and this method is called by two different tasks at once, you can initialize two different objects. This, for example, can be a problem with singleton classes because you only want to create one object of these classes.

A elegant solution to this problem has been implemented, as the Initialization-on-demand holder idiom (https://en.wikipedia.org/wiki/Initialization-on-demand_holder_idiom).

Blocking operations are those operations that block the task that calls them until an event occurs. For example, when you read data from a file or write data to the console, the task that calls these operations must wait until they finish.

If you include one of these operations into a critical section, you are degrading the performance of your application because none of the tasks that want to execute that critical section can execute it. The one that is inside the critical section is waiting for the finalization of an I/O operation, and the others are waiting for the critical section.

Unless it is imperative, don't include blocking operations inside a critical section.