Performance Evaluation, Profiling and Modeling Tools for ARM-based HPC Systems

The Arm microprocessor architecture has emerged as a legitimate alternative and competitor to predominant x86-based supercomputing architectures. An easy-to-use and robust tool chain for understanding, porting and refactoring existing HPC codes (e.g., commercial, academic, and DOE Office of Science codes) is essential to unlocking the full potential of Arm in the HPC realm. EP Analytics is developing and commercializing PerfPal, a set of performance analysis, modeling and prediction tools that can be deployed to design future Arm-based architectures for High-Performance Computing (HPC) applications and identify potential optimizations on application codes that can improve performance on existing and upcoming Arm-based systems. PerfPal relies on raw performance data extracted through direct analysis of the production binary, through a technique known as Binary Instrumentation (BI), to provide an in- depth view of application performance. Performance modeling/co-design component of PerfPal can utilize performance data collected on existing systems to project performance on future hardware. During Phase I, we prototyped foundational tools to enable PerfPal to present high-level performance analysis to the end-users. In addition, a new tool was designed that facilitates bringing SST (Structural Simulation Toolkit), a popular co-design framework, to the Arm architecture. EP Analytics will complete the development and commercialization PerfPal. The base binary instrumentation toolkit, EPAX, will be extended by adding more robust performance analysis-driven instrumentation modes. The toolkit will also be optimized to minimize the overhead of running instrumented binaries. In addition, a rich set of proprietary performance analysis tools will be developed and released as part of the PerfPal-Arm performance framework. Finally, a new feature that will inform future hardware/software co-design by utilizing raw performance data collected on existing supercomputing systems will also be added to the PerfPal framework. The Arm software development tools being created by EP Analytics have commercial applications across high performance computing, “hyper-scale” (Internet) data centers and cloud computing. This research also has the potential to contribute to and advance the national leadership in “Exascale” supercomputing technology.