Experimenting with low-overhead OpenMP runtime on IBM Blue Gene/Q
IBM Journal of Research and Development
| Hi-index | 0.00 |
Barrier construct in OpenMP program is a directive used to remove race code before continuing. Each thread waits until all of the other threads of a team have reached the barrier region. And data dependence is a technology to find whether two statements can be run in parallel. This paper presents two ways to optimize the OpenMP programa barrier. The first one is to remove redundant barrier with data dependence information, that is, if all the statements across a barrier have no data dependence, then the barrier can be safely removed. And the second one is to reduce the cost of barrier. The paper gives an implementation of another form of parallelism- DOACROSS and a new form of OpenMP barrier-region barrier, in which both are synchronized with busy-waiting. Experimental results show that the performance of the optimized OpenMP program is improved.