Performance Analysis of OpenMP Target Offloading in Score-P
Monday, May 13, 2024 3:00 PM to Wednesday, May 15, 2024 4:00 PM · 2 days 1 hr. (Europe/Berlin)
Foyer D-G - 2nd floor
Research Poster
Optimizing for Energy and PerformancePerformance Measurement
Information
Poster is on display and will be presented at the poster pitch session.
With increasing demand in compute performance of HPC systems, accelerators are getting the main focus for application development. Many of the Top500 HPC systems now include accelerators, with the top 3 systems alone having accelerators of three different vendors. This diversity requires application developers to choose portable frameworks to support all at the same time, as developing applications via each native API is time consuming. One of the available frameworks is OpenMP with its offloading capability and availability for C, C++ and Fortran. With OpenMP offloading gaining more traction recently, performance analysis becomes important as well. With this poster, we present our first results in adding support for OpenMP offloading to our instrumentation and measurement infrastructure Score-P using the OpenMP Tools Interface. We demonstrate how we can use both host side callbacks and the device tracing interface to build a measurement adapter capable of analyzing OpenMP applications effectively. We show the current support landscape between different compilers and present first results for profiles and event traces based on the SPEC HPC 2021 618.tealeaf_s benchmark running on the LUMI HPC cluster at CSC in Finland.
Contributors:
With increasing demand in compute performance of HPC systems, accelerators are getting the main focus for application development. Many of the Top500 HPC systems now include accelerators, with the top 3 systems alone having accelerators of three different vendors. This diversity requires application developers to choose portable frameworks to support all at the same time, as developing applications via each native API is time consuming. One of the available frameworks is OpenMP with its offloading capability and availability for C, C++ and Fortran. With OpenMP offloading gaining more traction recently, performance analysis becomes important as well. With this poster, we present our first results in adding support for OpenMP offloading to our instrumentation and measurement infrastructure Score-P using the OpenMP Tools Interface. We demonstrate how we can use both host side callbacks and the device tracing interface to build a measurement adapter capable of analyzing OpenMP applications effectively. We show the current support landscape between different compilers and present first results for profiles and event traces based on the SPEC HPC 2021 618.tealeaf_s benchmark running on the LUMI HPC cluster at CSC in Finland.
Contributors:
Format
On-site
