Chapter 44. NSCL DAQ Thread Library

The NSCL DAQ thread library supplies object oriented threading support. This chapter describes:

44.1. The thread and synchronization model

The NSCL Daq software models each thread as an object from a class that is derived from the Thread base class. This class provides functions that allow a thread to be created, started, exited, joined to and detached.

The body of a thread is provided by you when you create your concrete Thread subclass. The thread body is just your implementation of the virtual run method. Threads exit by returning from the run method.

Given a running thread, the join thread allows the caller to block until the thread represented by the object exits. joining is a necessary part of thread cleanup, unless the thread invokes its own detach member.

In the following example, a thread is designed that will block itself for a second, print a message and exit. Code is shown that creates the thread, starts it, joins it an deletes it. Note that deleting a running thread object is a bad idea and has undefined consequences.

Example 44-1. The life of a thread


        ...
class MyThread : public Thread {      (1)
public:
    virtual void run();
};
void
MyThread::run()                      (2)
{
    sleep(1);
    std::cerr << "My thread " << getId() << " is exiting\n";
    return;
};
   ...
   
MyThread* aThread = new MyThread;   (3)
aThread->start();                     (4)

   ...
   
aThread->join();                    (5)
delete aTrhead;                     (6)
 
    ...

            
(1)
By making MyThread a subclass of Thread, objects from class MyThread can be started as independent threads of execution
(2)
The run member of a thread is an abstract method. Your thread classes must supply the behavior for this member. When the thread is started, the run method gains control in the context of the new thread.
(3)
Creating a thread is simple. Just create an object that is of the thread class type.
(4)
Starting a thread is equally simple, Just call the thread's start method. That starts the thread with an entry point that eventually calls the run method.
(5)
It's not safe to destroy a thread that is executing. Calling a thread's join method blocks the caller's thread until the thread exits. Note that it is not safe for a thread to call its own join method since that will block the thread forever.
(6)
Once a thread has exited, the object that ran it can be destroyed. That effectively destroys the thread. If the state of the exiting thread allows, it is possible to start the thread again after it has exited.

Non trivial threaded software will almost always need some means for threads to synchronize against one another. Consider the following trivial, but wrong, example:

Example 44-2. Why synchonization is needed


int someCounter = 0;

class MyThread : public Thread
{
    virtual void run() {
        for (int i=0; i < 10000; i++) {
            someCounter++;
        }
};

...
MyThread* th1 = new MyThread;
MyThread* th2 = new MyThread;
th1->start();
th2->start();
th1->join();
th2->join();
delete th1;
delete th2;
std::cerr << "someCounter = " << someCounter << std::endl;
...


            

In this program, two threads increment the variable someCounter in parallel 10,000 times each. You might expect the output of this program to always be 20000. Most of the time, it probably will be. Sometimes it will be something less than 20000.

Consider what


someCounter++;
            
actually does. The value of someCounter is fetched to a processor register, the register is incremented and finally, the register is stored back into someCounter.

Suppose thread th1 fetches someCounter, and increments the register but before it has a chance to store the incremented value back into someCounter th2 executes, fetches someCounter (the old value), increments the register and stores the value back. Now th1 gets scheduled, and stores its value back.

This sequence of steps results in a lost increment. It is possible to construct sequences of execution, that result in a final value of someCounter holding any value from 10000 through 20000 depending on how access to someCounter is interleaved.

One way to fix this is to ensure that the increment of someCounter is atomic with respect to the increment. The NSCLDAQ threading library provides a synchonization primitive called a SyncGuard that can be used to implement the Monitor construct first developed by Per Brinch Hansen (see Wikipedia's Monitor (synchronization) page.

Let's rewrite the previous example so that the increment is atomic with respect to the scheduler. To do this we will isolate the counter in a class/object of its own so that it is not possible to use it incorrectly

Example 44-3. Using SyncGuard to implement a monitor


class ThreadedCounter {
private:
    Synchronizeable  m_guard;      (1)
    int              m_counter;
public:
    void increment();
    int  get() const;
};
void
ThreadedCounter::increment()
{
    sync_begin(m_guard);          (2)
    m_counter++;
    sync_end();                  (3)
}
int
ThreadedCounter::get() const
{
    return m_counter;
}

ThreadedCounter someCounter;
class MyThread : public Thread
{
    virtual void run() {
        for (int i=0; i < 10000; i++) {
            someCounter.increment();     (4)
        }
};

...
MyThread* th1 = new MyThread;
MyThread* th2 = new MyThread;
th1->start();
th2->start();
th1->join();
th2->join();
delete th1;
delete th2;
std::cerr << "someCounter = " << someCounter.get() << std::endl;
...

            
(1)
This member is the synchronization element. We will see it used in the increment member.
(2)
The sync_begin() enters the monitor. Only one thread at a time is allowed to execute the code between a sync_begin and a sync_end for the same synchronizing object.
(3)
This call leaves the monitor. The effect of the monitor is to make the increment atomic with respect to the scheduler.
(4)
By using the increment function, the counter is incremented atomically.