Exploring Concurrency: Basics in Java

4th April 2018

This is the first part of my 'Exploring Concurrency' series.

As a web developer, I spend most of my time working in a single-threaded environment. I would say this is a fairly safe assumption for most other web developers as well. Ultimately, it's much easier to write code when you don't have to worry about your code producing weird results and locking up indefinitely.

Recently I've decided to peer into Pandora's Box and learn more about concurrent programming and how I can incorporate it into my toolkit. As part of that learning I am going to write up my findings in a series of posts to consolidate some of that knowledge in a useful way.

For most of this series, I will mostly be working with the JVM languages (perhaps with some other novel languages thrown in here and there). Let's kick things off by exploring the basics using the most popular option - Java.

The Thread primitive

As you might expect from Java, the out-of-the-box concurrency primitives are simply OOP abstractions. To work with a thread, you simply create a new Thread and pass it a Runnable method. With Java 8 Lambda syntax this looks like:

new Thread(() -> System.out.println("Thread finished")).start();

Of course, if you need object-like behaviours, then you just need to implement Runnable or extend Thread and use that implementation instead:

new Thread(new Runnable {
  private int counter = 0;
  
  @Override
  public void run() { 
    counter++;
    System.out.println("Thread finished"); 
  }
})
  .start();

Threads can be interrupted to signal that they should do some other work, or simply stop. For example:

Thread childThread = new Thread(() -> {
  try {
    Thread.sleep(4000L);
    System.out.println("Child thread finished");
  } catch (InterruptedException e) {
    System.out.println("Child thread interrupted");
    // Propagate interrupt back up to the parent thread
    Thread.currentThread().interrupt();
  }
});

childThread.start();

// Sleep main thread then interrupt child prematurely
Thread.sleep(2000L);
childThread.interrupt();

Outputs after a 2 second delay:

Child thread interrupted

Methods such as sleep have a checked InterruptedException. The reasoning for this is that the language designers wanted to force users to handle such interruptions all the way back up the call stack, which seems reasonable. Unfortunately this leads to a lot of noise in our code (similarly for most checked exceptions). I think it would be good if we could attach some kind of exception handler to Thread instead, like the following pseudo-code:

new Thread(
  () -> Thread.sleep(4000L),
  e -> Thread.currentThread().interrupt()
);

Unfortunately, I doubt this kind of API change will happen any time soon! For brevity, I will not be including any further unnecessary checked exception handling for the rest of this post.

Continuing on, threads can be made to wait until other threads are completed before continuing execution by using join.

Thread childThread = new Thread(() -> {
  Thread.sleep(2000L);
  System.out.println("Child thread finished");
});

childThread.start();
childThread.join();

System.out.println("Main thread finished");

Outputs:

Child thread finished
Main thread finished

You might expect the opposite ordering as the childThread has to sleep for 2 seconds first. However, due to the childThread.join call, the main thread waits for childThread to complete before continuing with its own execution. Fairly straightforward so far.

Ready, set... go!

Whilst the basic usage of threads is pretty simple, the reality of using multiple threads can quickly become quite messy and unintuitive. We'll discuss the major problems that most concurrent code will have to face.

Thread interleaving

The first major issue is that threads do not have any guarantees on the order that they will run. This means we can get weird and wonderful results from innocuous looking code:

int x = 0;

new Thread(() -> x++).start(); // Thread A
new Thread(() -> x++).start(); // Thread B

assertThat(x).isEqualTo(2); // Could be true OR false

Whoa what? How can x be 1? It's like one operation was completely disregarded!

We would logically expect that:

  1. Thread A increments x before thread B.
  2. Thread B picks up the most recent increment of x.
  3. Thread B then increments x again using this up-to-date value.

As it turns out, threads can be scheduled (internally) to execute at any time. This means that you can end up in a scenario where they execute simultaneously (or overlap) and both see the value of x as 0. When this happens, they both set x to 1 instead of the expected 2.

Life gets even harder if we increase the number of threads. We increase the number of permutations and make the results even more unpredictable. Using three threads, x could equal 1, 2 or 3 depending on the order they run. Clearly not great news if we want to have bug-free code!

This issue is due to thread interleaving and is at odds with our normal, sequential thinking. A popular expression for this issue is the term - race condition.

To remedy this we need our code to be atomic. We need to ensure that when we perform compound operations they are actually performed in a single operation that is synchronized between threads.

More on compound operations

Typically compound operations consist of a read operation followed by some action i.e. 'check-then-act' or 'read-modify-write' operations. The differences between these types of compound operation are subtle, but worth thinking about. The main difference is that 'read-modify-write' expects that an update operation is performed using the most recent read operation's value.

The incrementation example earlier is a 'read-modify-write' operation. We get the most recent value, increment it then try to update the corresponding variable. The intention was to update x sequentially in two increments, however one of the updates was invalidated by the second thread's update.

On the other hand, 'check-then-act' usually revolves around the potential for a condition to pass when it shouldn't due to using a stale value. For example:

boolean doAction = true;

Runnable action = () -> {
  if (doAction) {
    System.out.println("Action was ran.");
    doAction = false;
  }
}

new Thread(action).start();
new Thread(action).start();

Can print the following:

Action was ran.
Action was ran.

Clearly our intention was only for 'Action was ran.' to be printed once, but here we can see it twice. You can imagine how this could be very unwarranted in a production application!

Memory consistency

Our problems don't stop there. We also have another problem which is even weirder.

Can we guarantee that the threads see the correct state of the resource at the correct moment in time during execution?

Unfortunately not! Changes to shared state are not guaranteed to ever be made visible to other threads that also have access to the shared state. This means that you can have peculiar situations like the following:

boolean looping = true;

new Thread(() -> {
  Thread.sleep(100L);
  looping = false;
})
  .start();

new Thread(() -> {
  while (looping) {
    // Do some work
  }
  
  System.out.println("Thread 2 finished");
})
  .start();

Again this looks innocent enough, but you might be surprised to find out that thread 2 may never finish looping. The update to looping by thread 1 is never made available to thread 2, so we get stuck in a continuous loop.

This time the problem is one of memory consistency. There are no guarantees that a thread's local memory will be flushed back to the main memory (shared by other threads) in a timely fashion. Another way of expressing this would be to say there is insufficient visibility between threads.

Teaching threads to work together

As we can see there are some tricky problems that we face to take advantage of having those extra threads.

Thankfully there are a few immediate features we can use to get threads to coordinate their efforts correctly.

The sychronized keyword

The easiest way of controlling the activity multiple threads is to use the synchronized keyword. It can be attached to class methods allowing only one thread to access the method at a time. For example:

class SyncedCounter {
  int counter = 0;
  
  synchronized inc() { counter++; }
}

SyncedCounter counter = new SyncedCounter();

new Thread(() -> counter.inc()).start();
new Thread(() -> counter.inc()).start();

assertThat(counter.counter).isEqualTo(2); 

It can also be applied in statements like the following for more 'fine-grained' control over specific blocks of code:

int counter = 0;
Object lock = new Object();

new Thread(() -> {
  synchronized (lock) {
    counter++;
  }
})
  .start();

new Thread(() -> {
  synchronized (lock) {
    counter++;
  }
})
  .start();

assertThat(counter).isEqualTo(2);

Note that in this instance, we also have to manually provide a lock object. If the thread has access to that lock object, then only it can execute the code block - it is said to have acquired the lock. Typically we could just pass in this (instead of some other object) as the lock for convenience. Using the synchronized keyword on class methods directly does the same thing, but automatically uses this as the lock.

We can see that synchronization is able to solve our thread interleaving and memory consistency problems. Unfortunately, we cannot apply it wholesale as it works against the performance benefits of using multiple threads in the first place. Synchronization is expensive and can cause thread starvation - a situation where threads sit idle when they could be better utilised elsewhere. Consequently, the ultimate aim should be to limit the scope of every synchronized block to be as small as possible.

One thing to note - whenever you use synchronized around some shared state, you need to ensure that ALL access (read or write) to the shared state is synchronized to ensure thread safety.

The volatile keyword

Another mechanism we can employ is the use of the volatile keyword on class fields. This allows us to get around the problem of visibility that we saw earlier. With this keyword, updates to the class field will be made immediately visible to other threads (rather than potentially never).

Augmenting our previous example to avoid visibility problems:

class LoopChecker {
  // Volatile can only be added to instance members
  volatile boolean looping = true;
}

LoopChecker loopChecker = new LoopChecker();

new Thread(() -> {
  Thread.sleep(100L);
  loopChecker.looping = false;
})
  .start();

new Thread(() -> {
  while (loopChecker.looping) {
    // Do some work
  }
  
  System.out.println("Thread 2 finished");
})
  .start();

This time it actually prints the following:

Thread 2 finished

volatile can be considered a 'lightweight' form of synchronization. However, it is not a complete replacement for synchronized as we can still run into problems where we require atomicity e.g. increments/decrements on numbers.

Atomic classes

The java.util.concurrent package provides atomic versions of primitives such as AtomicBoolean, AtomicInteger, AtomicReference, etc. These classes are designed to provide both atomicity and visibility when working with their equivalent primitive types. For example:

AtomicInteger counter = new AtomicInteger(0);

new Thread(() -> {
  System.out.println(counter.getAndIncrement());
})
  .start();

new Thread(() -> {
  System.out.println(counter.getAndIncrement());
})
  .start();

Prints the following:

1
2

Normally without any synchronisation, it could be possible that both printed lines are equal to 1. Fortunately getAndIncrement is an atomic operation ensuring we get the desired output. This atomicity is actually achieved via a low-level native 'compare-and-swap' mechanism that is not synchronization-based. The update operation will only be performed if the current value is exactly the same as an expected value. In atomic classes, this is exposed as the compareAndSet method, which is used under-the-hood for most (if not all) atomic operations in the atomic classes.

For example, getAndIncrement can be simplified to the following pseudo-code:

int current;
do {
  current = get();
} while (!compareAndSet(current, current + 1));

This loop essentially calls the atomic compareAndSet operation continuously with the 'most recent' value (obtained from get) until it succeeds. I think it's quite interesting that some programming problems are 'best' solved simply by using brute-force.

In summary, atomic classes should be considered as wrappers around volatile variables, and they are certainly a great option if we require atomicity. If not, then just using volatile may be sufficient, however, it seems that atomic classes are usually more versatile and useful.

Collections

As working with collections is commonplace in nearly any application, there are two packages we can immediately reach for when attempting to make our collections thread-safe.

Synchronized collections

These are actually just wrapper classes that add synchronized to every public method of the internal collection. They are created by calling their respective static methods:

Collections.synchronizedList(new ArrayList<Integer>());
Collections.synchronizedMap(new HashMap<String, Integer>());
Collections.synchronizedSet(new HashSet<Integer>());

As these collections just add synchronized to every public method, they do not exhibit good concurrent performance as only one thread will be able to access the collection at a time. As we've discussed earlier, this can cause thread starvation and is not scalable in a multi-threaded environment. These synchronized collections may be useful in certain scenarios, but there are other types of collections that should be considered first.

Concurrent collections

These are available via the java.util.concurrent package and are designed to solve the performance deficiencies in the synchronized collections. Examples of equivalent concurrent collections include ConcurrentHashMap, CopyOnWriteArrayList and CopyOnWriteArraySet.

So how do these concurrent collections differ?

Let's first look at ConcurrentHashMap. This implementation uses a more fine-grained locking mechanism called lock-striping to allow greater shared access between threads (instead of a single lock on all public methods). This allows many threads to read from the map, and a limited number of threads to write to it concurrently. Consequently, the get, put, containsKey and remove methods are well-optimized for concurrency.

Unfortunately, these optimizations are not without trade-offs. Operations such as size and isEmpty are purposefully not guaranteed to be accurate. The collection size is considered to be a 'moving target' - definitely something to be aware of!

As ConcurrentHashMap is not designed for exclusive locking (like the synchronized version), it exposes helpful atomic operations such as putIfAbsent, remove and replace.

Looking at CopyOnWriteArrayList/Set, these collections provide thread safety by copying the entire collection into a new collection every time there is an update. Iterators on this collection are constructed only with a snapshot of the collection at that point in time. However, this also means that later updates will not be propagated down to any of the iterators that already exist.

These collections are effectively immutable avoiding the need for heavyweight synchronization methods. Again, this is not without trade-off. Copying large collections is not fast so we should avoid using these collections when updates are more common than read-only iterations.

Conclusion

Concurrency is an attractive method for us to squeeze more performance out of our code. Unfortunately it is not for free as there are caveats that are not obvious and we have to be careful with.

  • Threads can interleave meaning that that operations are executed in unpredictable orders. When using some shared state, this can result in unexpected results.
  • Threads are not guaranteed to make their local changes to shared state available to other threads. This causes memory consistency issues.

These issues can be very problematic as their effects can be very subtle and dangerous. To battle these issues, we can employ several language features:

  • Use the synchronized keyword. This should be made as fine-grained as possible to avoid introducing performance problems (which defeats the point of multi-threading our code).
  • Use the volatile keyword. This allows us to flush updates to variables back to main memory immediately and make them visible to other threads.
  • Use the Atomic classes. These are like volatile variables, except that they also provide atomicity via methods such as compareAndSet.
  • Use concurrent collections. These collections maximize concurrent performance whilst also providing thread safety with relatively little overhead (and some caveats). They should be preferred over synchronized collections.

Hopefully this has been a relatively gentle introduction to concurrency in Java. I've tried my best to summarise some of the raft of information on the topic, but there is plenty more to discuss! For further reading I would recommend checking out Java Concurrency in Practice, which is a very comprehensive resource that is still very relevant to this topic.

In the next post of this series, I will be discussing some of the higher-level abstractions that are available for dealing with concurrency in the Java standard library. I look forward to seeing you there!