Parallel Programming Guide for HP-UX Systems

Parallel synchronization

Synchronizing code

Chapter 8154

C$DIR ORDERED_SECTION(ORDGATE)

A(I,J) = A(I+1,J)

C$DIR END_ORDERED_SECTION

ENDDO

Recall that once a given thread has passed through an ordered section, it

cannot reenter it until all other threads have passed through in order.

This is only possible in the given example if the number of available

threads integrally divides 99 (the I loop limit). If not, deadlock results.

To better understand this:

• Assume 6 threads, numbered 0 through 5, are running the parallel I

loop.

•ForI = 1, J = 1, thread 0 passes through the ordered section and

loops back through J, stopping when it reaches the ordered section

again for I = 1, J = 2. It cannot enter until threads 1 through 5

(which are executing I = 2 through 6, J = 1 respectively) pass

through in sequence. This is not a problem, and the loop proceeds

through I = 96 in this fashion in parallel.

•ForI > 96, all 6 threads are no longer needed. In a single loop nest

this would not pose a problem as the leftover 3 iterations would be

handled by threads 0 through 2. When thread 2 exited the ordered

section it would hit the ENDDO and the I loop would terminate

normally.

• But in this example, the J loop isolates the ordered section from the I

loop, so thread 0 executes J = 1 for I = 97, loops through J and waits

during J = 2 at the ordered section for thread 5, which has gone idle,

to complete. Threads 1 and 2 similarly execute J = 1 for I = 98 and

I = 99, and similarly wait after incrementing J to 2. The entire J loop

must terminate before the I loop can terminate, but the J loop can

never terminate because the idle threads 3, 4, and 5 never pass

through the ordered section. As a result, deadlock occurs.

To handle this problem, you can expand the ordered section to include

the entire j loop, as shown in the following C example: