Timemory: Modular Performance Analysis for HPC

Abstract

HPC has undergone a significant transition toward heterogeneous architectures. This transition has introduced several issues in code migration to support multiple frameworks for targeting the various architectures. In order to cope with these challenges, projects such as Kokkos and LLVM create abstractions which map a generic front-end API to the backend that supports the targeted architecture. This paper presents a complementary framework for performance measurement and analysis. Several performance measurement and analysis tools in existence provide their capabilities through various methods but the common theme among these tools are prohibitive limitations in terms of user-level extensions. For this reason, software developers commonly have to learn multiple tools and valuable analysis methods, such as the roofline model, are frequently required to be generated manually. The timemory framework provides complete modularity for performance measurement and analysis and eliminates all restrictions on user-level extensions. The timemory framework also provides a highly-efficient and intuitive method for handling multiple tools/measurements (i.e., “components”) concurrently. The intersection of these characteristics provide ample evidence that timemory can serve as the common interface for existing performance measurement and analysis tools. Timemory components are developed in C++ but includes multi-language support for C, Fortran, and Python codes. Numerous components are provided by the library itself – including, but not limited to, timers, memory usage, hardware counters, and FLOP and instruction roofline models. Additionally, analysis of the intrinsic overhead demonstrates superior performance in comparison with popular tools.