Scalable Visualization of Nsight Systems Traces with Paraver

Описание: The number of scientific applications using GPUs for accelerated compute is steadily growing, which adds new programming models and practices into play. Moreover, ML applications present new parallelization strategies and communication libraries new to HPC environments. These factors require a revision of the tools and methodologies that we use for performance analysis, while still applying our perspective for scalability and interest for detail.

In this webinar, we showed you how you can gain additional performance insight from Nvidia nsys traces with Paraver, by translating them with our tool nsys2prv. By enabling the visualization of these traces on one of POP's performance analysis tools, we can navigate through scales more easily and display new and interesting performance metrics, apply efficiency models, and compare and quantify differences between different executions, objects, or timeline regions.

About the Presenter

Marc Clascà is a master student in Computer Engineering at Universitat Politècnica de Catalunya, and a research engineer at the Barcelona Supercomputing Center since 2020. His research interests include programming models, performance tools, performance analysis and specialized hardware and accelerators. His current research focus is on analyzing the performance of scientific applications and AI use cases that use accelerators, with the aim of deriving new efficiency metrics and analysis methodologies. This includes exploring the potentials of GPU specific tracing and visualization tools, and understanding new programming models and communication patterns used in LLM training and inference.

Slides: https://pop-coe.eu/sites/default/file...
POP CoE: https://pop-coe.eu