The major performance bottleneck of the parallel Fourier method on distributed memory systems is the network communication cost. In this study, we investigate the potential of using non-blocking all-to-all communications to solve this problem by overlapping computation and communication. We present the runtime comparison of a 3D seismic modeling problem with the Fourier method using non-blocking and blocking calls, respectively, on a Linux cluster. The data demonstrate that a performance improvement of up to 40% can be achieved by simply changing blocking all-to-all communication calls to non-blocking ones to introduce the overlapping capability. A 3D reverse-time migration result is also presented as an extension to the modeling work based on non-blocking collective communications.
Bibliographical noteKAUST Repository Item: Exported on 2020-10-01
Acknowledgements: We thank the Texas Advanced Computing Center forproviding us the computational resources. Chunlei Chuwould like to thank Phuong Vu for many insightfuldiscussions on parallel computing. This work was madepossible in part with funding from ConocoPhillips and theKing Abdullah University of Science and Technology(KAUST).
This publication acknowledges KAUST support, but has no KAUST affiliated authors.