Optimize Your Code: Intro to Profiling for HPC & AI Applications 

Date: Wednesday, December 11, 2024, 10:00-14:00, Cyprus time

Venue: This training event is held as a hybrid event. You are welcome to join us at the Andreas Mouskos Auditorium, José Mariano Gago Hall, The Cyprus Institute. Otherwise please, connect to our live stream of the discussion, available on Zoom (Password: VsSCz1) 

Language: English 

Registration: Registration for this event is open until Monday, December 09, 2024. Registration form here. 

Pre-requisites: Familiarity with using a terminal or command line interface is beneficial. In particular, you should be happy with connecting using SSH. Also, some previous experience with Pytorch will be useful.

Requirements: On-site attendees should bring their own laptop with a Mac, Linux, or Windows operating system that they have administrative privileges on.

Agenda

10:00 - 10:30

Dr Christodoulos Stylianou

Title: Profiling Basics: Understanding Metrics and Tools for HPC

This informative session will introduce participants to the foundational concepts of profiling in high-performance computing (HPC). The presentation will cover what profiling is and why it is crucial for optimizing code performance. Attendees will learn about key performance metrics and how to interpret them, as well as explore popular profiling tools used in HPC environments. This session aims to equip participants with the knowledge needed to understand the performance characteristics of their code and identify potential areas for optimization.

10:30 - 11:00

Mr Giorgos Kosta

Title: Profiling AI Applications: Techniques and Best Practices

This session transitions into the specialized area of profiling AI applications. Participants will learn about the unique performance considerations when profiling deep learning models and AI workloads. We will introduce tools specifically designed for AI profiling, such as those that measure GPU utilization and memory bandwidth. The session will provide insights into interpreting AI-specific metrics, understanding training and inference bottlenecks, and optimizing models for better performance. Attendees will gain the knowledge needed to fine-tune their AI applications for increased efficiency.

11:00 - 11:30

Mr Spiros Millas

Title: Scaling Up: Multi-GPU Parallelism for AI Models

This session transitions into the specialized area of profiling AI applications. Participants will learn about the unique performance considerations when profiling deep learning models and AI workloads. We will introduce tools specifically designed for AI profiling, such as those that measure GPU utilization and memory bandwidth. The session will provide insights into interpreting AI-specific metrics, understanding training and inference bottlenecks, and optimizing models for better performance. Attendees will gain the knowledge needed to fine-tune their AI applications for increased efficiency.

Coffee Break and Networking

12:00 - 14:00

Hands-On Session: Practical Profiling for AI Applications

In this interactive hands-on session, participants will put their knowledge into practice by learning how to profile AI models effectively. The session will guide attendees through setting up profiling tools, executing profiling tasks, and analyzing performance data to optimize their AI applications. Special emphasis will be placed on demonstrating how to apply these profiling techniques in a multi-GPU environment, allowing users to explore strategies for monitoring and improving the performance of parallel processing models. By the end of this session, participants will gain practical experience and confidence in profiling and optimizing their AI workflows, both for single and multi-GPU setups.