TY - GEN
T1 - Performance Modeling of Hybrid MPI/OpenMP Scientific Applications on Large-scale Multicore Cluster Systems
AU - Wu, Xingfu
AU - Taylor, Valerie
N1 - KAUST Repository Item: Exported on 2020-10-01
Acknowledged KAUST grant number(s): KUS-I1-010-01
Acknowledgements: This work is supported by NSF grant CNS-0911023 and theAward No. KUS-I1-010-01 made by King AbdullahUniversity of Science and Technology (KAUST). Theauthors would like to acknowledge Argonne LeadershipComputing Facility for the use of BlueGene/P under DOEINCITE project “Performance Evaluation and AnalysisConsortium End Station”, the SDSC for the use of DataStarP655 under TeraGrid project TG-ASC040031, and TAMUSupercomputing Facilities for the use of Hydra. We wouldalso like to thank Stephane Ethier from Princeton PlasmaPhysics Laboratory and Shirley Moore from University ofTennessee for providing the GTC code.
This publication acknowledges KAUST support, but has no KAUST affiliated authors.
PY - 2011/8
Y1 - 2011/8
N2 - In this paper, we present a performance modeling framework based on memory bandwidth contention time and a parameterized communication model to predict the performance of OpenMP, MPI and hybrid applications with weak scaling on three large-scale multicore clusters: IBM POWER4, POWER5+ and Blue Gene/P, and analyze the performance of these MPI, OpenMP and hybrid applications. We use STREAM memory benchmarks to provide initial performance analysis and model validation of MPI and OpenMP applications on these multicore clusters because the measured sustained memory bandwidth can provide insight into the memory bandwidth that a system should sustain on scientific applications with the same amount of workload per core. In addition to using these benchmarks, we also use a weak-scaling hybrid MPI/OpenMP large-scale scientific application: Gyro kinetic Toroidal Code in magnetic fusion to validate our performance model of the hybrid application on these multicore clusters. The validation results for our performance modeling method show less than 7.77% error rate in predicting the performance of hybrid MPI/OpenMP GTC on up to 512 cores on these multicore clusters. © 2011 IEEE.
AB - In this paper, we present a performance modeling framework based on memory bandwidth contention time and a parameterized communication model to predict the performance of OpenMP, MPI and hybrid applications with weak scaling on three large-scale multicore clusters: IBM POWER4, POWER5+ and Blue Gene/P, and analyze the performance of these MPI, OpenMP and hybrid applications. We use STREAM memory benchmarks to provide initial performance analysis and model validation of MPI and OpenMP applications on these multicore clusters because the measured sustained memory bandwidth can provide insight into the memory bandwidth that a system should sustain on scientific applications with the same amount of workload per core. In addition to using these benchmarks, we also use a weak-scaling hybrid MPI/OpenMP large-scale scientific application: Gyro kinetic Toroidal Code in magnetic fusion to validate our performance model of the hybrid application on these multicore clusters. The validation results for our performance modeling method show less than 7.77% error rate in predicting the performance of hybrid MPI/OpenMP GTC on up to 512 cores on these multicore clusters. © 2011 IEEE.
UR - http://hdl.handle.net/10754/599161
UR - http://ieeexplore.ieee.org/document/6062871/
UR - http://www.scopus.com/inward/record.url?scp=81455155118&partnerID=8YFLogxK
U2 - 10.1109/CSE.2011.42
DO - 10.1109/CSE.2011.42
M3 - Conference contribution
SN - 9781457709746
SP - 181
EP - 190
BT - 2011 14th IEEE International Conference on Computational Science and Engineering
PB - Institute of Electrical and Electronics Engineers (IEEE)
ER -