Hire a web Developer and Designer to upgrade and boost your online presence with cutting edge Technologies

Thursday, February 3, 2011

Java Synchronization

The Java programming language supports multiple threads, using a monitor construction for synchronization. In the Java language, a method can be declared as synchronized.Threads communicate primarily by sharing access to fields and the objects reference fields refer to. This form of communication is extremely efficient, but makes two kinds of errors possible: thread interference and memory consistency errors. The tool needed to prevent these errors is synchronization.
A built-in mechanism ensures that only one Java thread can execute an object's synchronized methods at a time. The mechanism also allows threads to wait for resources to become available, and allows a thread that makes a resource available to notify other threads that are waiting for the resource. This web page describes this mechanism and its use. But first, we need to look at an example that illustrates why synchronization is needed.
The Bounded Buffer Problem

In the Bounded Buffers problem, there are two kinds of processes or threads: producers and consumers. Producers take buffers from a queue of empty buffers, put content into the buffers, and place the buffers into a queue of full buffers. Consumers take buffers from a queue of full buffers, do something with the buffers' contents, and place the buffers back onto the queue of empty buffers. Each of the buffer queues has a bounded capacity for buffers. Processes communicating through a pipe or communicating through mailboxes are examples of the Bounded Buffers problem.

There is generally only one queue of empty buffers, maintained by the operating system. It provides all buffers for processes or threads. There can be multiple producers, consumers, and queues of full buffers. The queues of full buffers usually have a small capacity, on the order of 10 buffers. We generally only need to consider a single queue of full buffers at a time. The following picture illustrates the flow of buffers.

Figure 1. Buffer flow in the Bounded Buffers Problem.
A First Solution

When we have a good solution to the Bounded Buffers problem, the coordination between producers and consumers will be controlled by two BufferQueue objects representing the queue of empty buffers and the queue of full buffers. The following code is an implementation of a BufferQueue class that works well in a single-threaded environment.
Testing the First Solution

We are limited in the kind of testing we can do on this first version of the BufferQueue class. There is no good way of making producers wait for empty buffers or making consumers wait for full buffers. So we will only use producers. Initially the queue of empty buffers has 10 buffers and the queue of full buffers has none. We first run a single producer that moves buffers from the queue of empty buffers to the queue of full buffers as long as the queue of empty buffers is not empty. The producer kills time for 20 milliseconds on the average after removing a buffer from the queue of empty buffers and kills time for 20 milliseconds on the average after putting the buffer on the queue of full buffers.

Here is the log for this kind of test.

0001: Producer 1 got Buffer 0 from BufferQueue empty.
0042: Producer 1 removed Buffer 0 from BufferQueue empty.
0068: Producer 1 got Buffer 1 from BufferQueue empty.
0087: Producer 1 removed Buffer 1 from BufferQueue empty.
0134: Producer 1 got Buffer 2 from BufferQueue empty.
0140: Producer 1 removed Buffer 2 from BufferQueue empty.
0173: Producer 1 got Buffer 3 from BufferQueue empty.
0222: Producer 1 removed Buffer 3 from BufferQueue empty.
0251: Producer 1 got Buffer 4 from BufferQueue empty.
0276: Producer 1 removed Buffer 4 from BufferQueue empty.
0294: Producer 1 got Buffer 5 from BufferQueue empty.
0298: Producer 1 removed Buffer 5 from BufferQueue empty.
0316: Producer 1 got Buffer 6 from BufferQueue empty.
0335: Producer 1 removed Buffer 6 from BufferQueue empty.
0380: Producer 1 got Buffer 7 from BufferQueue empty.
0390: Producer 1 removed Buffer 7 from BufferQueue empty.
0423: Producer 1 got Buffer 8 from BufferQueue empty.
0438: Producer 1 removed Buffer 8 from BufferQueue empty.
0464: Producer 1 got Buffer 9 from BufferQueue empty.
0532: Producer 1 removed Buffer 9 from BufferQueue empty.

As you can see, the queue of empty buffers appears to be working correctly.

Now let's try the same test with two producers. Here is the log.

0001: Producer 1 got Buffer 0 from BufferQueue empty.
0004: Producer 2 got Buffer 0 from BufferQueue empty.
0018: Producer 2 removed Buffer 0 from BufferQueue empty.
0050: Producer 2 got Buffer 1 from BufferQueue empty.
0059: Producer 1 removed Buffer 1 from BufferQueue empty.
0063: Producer 2 removed Buffer 2 from BufferQueue empty.
0064: Producer 1 got Buffer 3 from BufferQueue empty.
0086: Producer 1 removed Buffer 3 from BufferQueue empty.
0098: Producer 1 got Buffer 4 from BufferQueue empty.
0110: Producer 2 got Buffer 4 from BufferQueue empty.
0114: Producer 2 removed Buffer 4 from BufferQueue empty.
0120: Producer 1 removed Buffer 5 from BufferQueue empty.
0146: Producer 1 got Buffer 6 from BufferQueue empty.
0161: Producer 1 removed Buffer 6 from BufferQueue empty.
0167: Producer 1 got Buffer 7 from BufferQueue empty.
0174: Producer 1 removed Buffer 7 from BufferQueue empty.
0177: Producer 2 got Buffer 8 from BufferQueue empty.
0184: Producer 1 got Buffer 8 from BufferQueue empty.
0200: Producer 1 removed Buffer 8 from BufferQueue empty.
0202: Producer 2 removed Buffer 9 from BufferQueue empty.

If you look at the "got Buffer" reports, you can see that some of the buffers were gotten twice (Buffer 0, for example) and others weren't gotten at all (Buffer 2, for example).
A Second Solution

The first solution has some obvious problems in a multithreaded environment. First of all, it makes no provision for making a thread wait. To get the BufferQueue class tested, we needed to get rid of the consumers, who need to wait for a full buffer. Second, multiple threads that both read and modify shared data can make a mess of the data.

Part of the problem is the interface for the BufferQueue class. The fact that the functionality of the getBuffer() and removeBuffer() methods are separated causes trouble whenever one process does a getBuffer() in between another process's getBuffer() and removeBuffer() - both processes get the same buffer. If the first process did a removeBuffer() between another process's getBuffer() and removeBuffer() then the second buffer to be removed will never have been gotten.

When a data structure is modified, you usually need to get local information about the structure, then modify it. You cannot do this correctly if the structure is modified in between these two steps. When you deal with multiple threads, there are usually operations that have to be executed atomically. That is, the operations have to be performed completely without the possibility of another thread trying to make modifications at the same time.

To fix our BufferQueue, we must start by modifying its methods. We need to combine the getBuffer() and removeBuffer() functionality. So we combine them into a takeBuffer() method that removes and returns a buffer. While we're at it, we may as well get rid of the isEmpty() and isFull() methods. The boolean values that they return can easily change before they are used. We thus end up with the second version of BufferQueue.java.

This version of the BufferQueue class cannot properly be tested since there is no way of properly stopping a producer when the queue of empty buffers is empty. If we did come up with a way of doing that would it work? Yes, for a while, but eventually it would fail. Since a thread swith can occur at any time, we do not have any guarantee that takeBuffer() and putBuffer() are executed atomically. As a matter of fact, these methods execute so fast relative to the time between thread switches that failure would probably occur only once every few hundred thousand operations. Imagine examining logs that big to try to figure out what is going wrong. Problems that show up rarely are much harder to troubleshoot than problems that occur frequently.
Correctly Solving Synchronization Problems

A correct solution to a synchronization problem requires the following.

* atomic method execution that gives threads mutually exclusive access to bookkeeping data,
* a mechanism for waiting until an operation can be performed, and
* a mechanism for notifying other threads of a change in state.

The first two needs have been discussed earlier. The third is a consequence of the second. In the Bounded Buffers problem, for example, consumers must wait until a producer has put a buffer onto the queue of full buffers. That means that there must be something in the putBuffer() method that wakes up threads that are waiting.

A monitor is a convenient mechanism that tackles addresses these needs. In order to understand the monitor concept, let's take a look at how monitors are implemented in Java.
Java's Implementation of Monitors

Any object in Java is a monitor if some of its methods are declared with the synchronized keyword. As a monitor, there are three mechanisms for addressing the needs described above.

* Atomic method execution that gives threads mutually exclusive access to bookkeeping data - The synchronized keyword:
When this keyword is placed before a method's return type, threads that invoke the method must acquire a lock on the receiving object. Only one thread can hold this lock at a time. This thread is then the object's lock holder. Any other thread that sends a message to the object invoking a synchronized method becomes a lock waiter.

* A mechanism for waiting until an operation can be performed - The wait() method:
An object with synchronized methods has a wait set. The wait() method can only be invoked by the object's lock holder. When the lock holder invokes the wait() method it is placed into the object's wait set. If a thread other than the lock holder invokes the wait() method it will cause an IllegalMonitorStateException to be thrown.

* A mechanism for notifying other threads of a change in state - The notifyAll() method:
This method wakes up all threads in the wait set. Since more than one thread may be woken up, waiting threads usually need to check again to see if they can accomplish their objective. The notifyAll() method also throws an IllegalMonitorStateException if invoked by a thread other than the lock holder.

A Properly Synchronized BufferQueue

In order to properly synchronize a BufferQueue for the Bounded Buffers problem, the following changes need to be made.

* Declare takeBuffer() and putBuffer() synchronized. Put the keyword synchronized before the return types for the methods.

* Use a wait loop at the beginning of takeBuffer() and putBuffer(). This wait loop has the following form.

while (unable to proceed) {
try {
wait();
} catch(InterruptedException ex) {
}
}

The takeBuffer() method is unable to proceed when the count variable is 0. The putBuffer() method is unable to proceed when the count variable is equal to buffers.length.

* Invoke notifyAll() to wake up the wait set. Normally, this method is invoked at the end of a synchronized method. It should only be invoked when the change in state can make some thread in the wait set able to proceed. When the count variable is buffers.length - 1 at the end of the putBuffer() method then a thread that is waiting because the queue is full is able to proceed. When the count variable is equal to 1 at the end of the takeBuffer() method then a thread that is waiting because the queue is empty is able to proceed.

Execution of the Properly Synchronized Solution

The following table shows the monitor state for the queue of full buffers. For this example, each producer fills one buffer, puts it onto the queue of full buffers, then terminates. Each consumer gets one buffer from the queue of full buffers, consumes it, then terminates. At the start, there are two consumers and two producers, and the queue of full buffers is empty. All threads are attempting to access the queue sumultaneously.

In this table, count is the number of buffers in the queue, lh is its lock holder, lw is its lock waiters, and ws is its wait set.

Consumer Producer full BufferQueue runable
A B C D count lh lw ws threads
takeBuffer takeBuffer putBuffer putBuffer 0 - - - A, B, C, D
runs blocks blocks blocks | A B, C, D A
wait | - - A B, C, D
resumes blocks blocks | B C, D | B
wait | - - A, B C, D
resumes blocks | C D | C
add Buffer 1 | | | |
notifyAll | | A, B, D - |
returns | - - A, B, D
resumes blocks blocks | A B, D A
take buffer 0 | | |
returns | - - B, D
resumes blocks | B D B
wait | - - B D
resumes | D | |
add buffer 1 | | |
notifyAll | | B - |
returns | - - B
resumes | B |
take buffer 0 | |
returns | - |

No comments:

Post a Comment