Performance Analysis: An introduction to the tools and performance analysis methodology used in POP CoE
Date + Time:
Tutorial at Supercomputing Frontiers Europe 2018
Thursday 15th March, 2018, 1:30 pm – 5:00 pm
Brief agenda:
- Introduction to performance analysis and POP CoE methodology.
- Performance analysis with Paraver and Scalasca for parallel codes.
- Parallel I/O Profiling and Darshan profiling tool.
- Hands-on session with profiling tools.
Presenters:
Jon Gibson, Numerical Algorithms Group Ltd., Great Britain
Wadud Miah, Numerical Algorithms Group Ltd., Great Britain
Abstract:
POP CoE (Performance Optimisation and Productivity Centre of Excellence) is funded by the Horizon 2020 programme to help people write more efficient parallel. We will present the systematic approach used to analyse and optimise parallel codes and an introduction to some open source profiling tools used within the project. Measuring application performance often results in large amounts of data that is difficult to handle or interpret beyond some simple first observations. The methodology used within POP CoE provides a quantitative way of measuring the relative impact of the different factors inherent in parallelisation, each metric reflecting a common cause of inefficiency. This provides the knowledge necessary to decide the best course of action to get performance via reproducible and comparable measurements of the performance. POP CoE is a collaboration between Barcelona Supercomputing Center, High Performance Computing Center Stuttgart, Jülich Supercomputing Centre, Numerical Algorithms Group Ltd, RWTH Aachen and TERATEC.