Summary

Auto-Parallelizing Compilers can do a lot, but they are limited to loop level parallelism and in analysis of data dependencies
Manual parallelization with directives is easy to implement, however care must be given to the treatment of data
It is important to understand the exact semantics of the parallelization directives
To obtain scalable performance it is often necessary to make program parallel at coarse-grain level, e.g. across subroutine invocations
To obtain scalable performance it is often necessary to take care of proper data distribution on the machine
It is necessary to profile parallel programs to diagnose cases of load imbalance and parallelization overhead.