TY - GEN
T1 - OpenMP Advisor
T2 - Proceedings of the 19th International Workshop on OpenMP, IWOMP 2023
AU - Mishra, Alok
AU - Malik, Abid M.
AU - Lin, Meifeng
AU - Chapman, Barbara
N1 - Publisher Copyright:
© 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2023
Y1 - 2023
N2 - With the increasing diversity of heterogeneous architecture in the HPC industry, porting a legacy application to run on different architectures is a tough challenge. In this paper, we present OpenMP Advisor, a novel compiler tool that enables code offloading to a GPU with OpenMP using Machine Learning. Although the tool is currently limited to GPUs, it can be extended to support other OpenMP-capable devices. The tool has two modes: Training and Prediction. It analyzes benchmark codes, generates every possible code variant on the target device, runs and gathers data to train an ML-based cost model in the training mode, which predicts the runtime of every code variant in the prediction mode. The main objective behind this tool is to maintain the portability aspect of OpenMP. Our Advisor produced code for several applications on seven architectures with four compilers, and accurately anticipated the top ten options for each application on every architecture. Initial results suggest that this tool can help compiler developers and HPC researchers migrate their legacy codes to the new heterogeneous computing environment.
AB - With the increasing diversity of heterogeneous architecture in the HPC industry, porting a legacy application to run on different architectures is a tough challenge. In this paper, we present OpenMP Advisor, a novel compiler tool that enables code offloading to a GPU with OpenMP using Machine Learning. Although the tool is currently limited to GPUs, it can be extended to support other OpenMP-capable devices. The tool has two modes: Training and Prediction. It analyzes benchmark codes, generates every possible code variant on the target device, runs and gathers data to train an ML-based cost model in the training mode, which predicts the runtime of every code variant in the prediction mode. The main objective behind this tool is to maintain the portability aspect of OpenMP. Our Advisor produced code for several applications on seven architectures with four compilers, and accurately anticipated the top ten options for each application on every architecture. Initial results suggest that this tool can help compiler developers and HPC researchers migrate their legacy codes to the new heterogeneous computing environment.
KW - compiler
KW - cost model
KW - gpu
KW - machine learning
KW - openmp
UR - https://www.scopus.com/pages/publications/85172077071
U2 - 10.1007/978-3-031-40744-4_3
DO - 10.1007/978-3-031-40744-4_3
M3 - Conference contribution
AN - SCOPUS:85172077071
SN - 9783031407437
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 34
EP - 48
BT - OpenMP
A2 - McIntosh-Smith, Simon
A2 - Deakin, Tom
A2 - Klemm, Michael
A2 - de Supinski, Bronis R.
A2 - Klinkenberg, Jannis
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 13 September 2023 through 15 September 2023
ER -