3.4.1 Parallel Execution Model

The parallel execution of loops is performed by Solaris threads. The thread starting the initial execution of the program is called the master thread. At program start-up the master thread creates multiple slave threads as shown in the following figure. At the end of the program all the slave threads are terminated. Slave thread creation is performed exactly once to minimize the overhead.

Figure 3–1 Master and Slave Threads

After start-up, the master thread starts the execution of the program while slave threads wait idly. When the master thread encounters a parallel loop, different iterations of the loop are distributed among the slave and master threads which start the execution of the loop. After each thread finishes execution of its chunk it synchronizes with the remaining threads. This synchronization point is called a barrier. The master thread cannot continue executing the remainder of the program until all the threads have finished their work and reached the barrier. The slave threads go into a wait state after the barrier waiting for more parallel work, and the master thread continues to execute the program.

During this process, various overheads can occur:

• The overhead of synchronization and work distribution

• The overhead of barrier synchronization

In general, there may be some parallel loops for which the amount of useful work performed is not enough to justify the overhead. For such loops, there may be appreciable slowdown. In the following figure, a loop is parallelized. However the barriers, represented by horizontal bars, introduce significant overhead. The work between the barriers is performed serially or in parallel as indicated. The amount of time required to execute the loop in parallel is considerably less than the amount of time required to synchronize the master and slave threads at the barriers.

Figure 3–2 Parallel Execution of a Loop