Score-P and OMPT: Bridging the Gap for Offloading in OpenMP Performance Analysis

Reuter, Jan André; Mohr, Bernd

doi:10.1007/978-3-032-41029-0_14

Contribution to a conference proceedings/Contribution to a book

FZJ-2026-04928

Score-P and OMPT: Bridging the Gap for Offloading in OpenMP Performance Analysis

Reuter, J. A. (Corresponding author)FZJ* ; Mohr, B.FZJ*

2027
Springer Nature Switzerland Cham
ISBN: 978-3-032-41028-3 (print), 978-3-032-41029-0 (electronic)

OpenMP: Adaptability for Heterogeneous Multi‑Device Systems
22nd International Workshop on OpenMP, IWOMP 2026, Vienna, Austria, 7 Oct 2026 - 9 Oct 2026 Cham : Springer Nature Switzerland, Lecture Notes in Computer Science 17099, 215 - 230 (2027) [10.1007/978-3-032-41029-0_14]

This record in other databases:

Please use a persistent id in citations: doi:10.1007/978-3-032-41029-0_14

Abstract: With the diverse space of accelerators in HPC, OpenMP target offloading has become a prevalent approach in accelerator programming. However, performance analysis has historically been challenging due to the distribution of relevant events across distinct tooling interfaces. OMPT enables the observation of these events through a standardized runtime interface, thereby facilitating the execution of OpenMP-level offload analysis. This paper extends the OMPT-based measurement support of Score-P to include support for target offloading. We outline the implementation approach and portability workarounds necessary to accommodate the discrepancies between current runtime and compiler implementations. For assessment, we investigate the overhead and compare OMPT-reported durations with vendor low-level interfaces utilizing microbenchmarks and SPEChpc 2021 benchmark workloads for the OpenMP runtimes provided by AMD ROCm and the HPE Cray Compiler Environment. The findings of the study suggest that OMPT-based offload analysis is a valuable tool for the evaluation of specific configurations. However, it is important to note that runtime-dependent limitations persist, particularly with regard to timestamp behavior and robustness. Still, the portability of the interface allows analysis across different runtime implementations, without the need for supporting multiple vendor-specific interfaces.

Contributing Institute(s):