|
Welcome
|
We should all understand and follow the ARCHER2 Code of Conduct to ensure this course is conducted in the best teaching environment.
The course will be flexible to best meet the learning needs of the attendees.
Feedback is an essential part of our training to allow us to continue to improve and make sure the course is as useful as possible to attendees.
|
|
Understanding your HPC workflow
|
Your HPC research workflow consists of many components.
The potential performance improvement depends on the whole workflow, not just the individual component.
|
|
Overview of ARCHER2 and connecting via SSH
|
ARCHER2 consists of high performance login nodes, compute nodes, storage systems and interconnect.
The system is based on standard Linux with command line access.
ARCHER2’s login address is login.archer2.ac.uk.
You connect to ARCHER2 using an SSH client.
|
|
HPC performance and benchmarking
|
Different timing and performance metrics are used for different applications.
Use the lowest node/core count that is feasible for your baseline.
Plan your benchmarking before you start, make sure you understand which parameters you want to vary and why.
|
|
Analysing performance
|
You can use benchmarking data to make decisions on how to best use your HPC resources.
Parallel efficiency is often the key decision metric for limits of parallel scaling.
Parallel performance is ultimately limited by the serial parts of the calculation.
|
|
Automating performance data collection and analysis
|
|
|
Introduction to profiling
|
Profiling allows to see what routines are consuming runtime in our application.
Profiling can also provide information on memory use and IO by our application.
The differences between profiles at different node/core counts can be more revealing than a single profile.
|
|
Where next?
|
|