HPC Tunathon 2020

Postponed (25 - 29 May 2020)

Forschungszentrum Jülich
Jülich Supercomputing Centre

Currently installed systems at JSC and the ones planned for the future are characterised by ever increasing levels of complexity. This makes it necessary to master different programming paradigms and tools for analysis. At the same time, the size of computer systems and problem sizes tackled by scientists increases. This increased potential and demand in net compute power asks for better scalability of user codes. Consequently, the HPC Tunathon aims to tune and optimise application execution on large scale. Tools and techniques to achieve scalability will be briefly introduced and extensively used during the workshop.

The HPC Tunathon is focused on hands-on practical training in a cooperative atmosphere. Experts will be available to assist with a variety of tools and support performance analysis, tuning, and scaling of the users' HPC applications. The outcome we aim for is a more efficient use of the resources allocated to the users on JUWELS and the future JUWELS Booster system.

Topics covered include:

• Structured parallel performance engineering
• Execution characterisation and MPI correctness validation
• Multi-threading and node-level/roof-line analysis
• Multi-node performance/scalability analysis
• Optimising CUDA-aware MPI
• Single-core performance analysis including vectorisation

Our experts will help attendees to investigate the scalability of their codes, aiming to run on the full JUWELS CPU or GPU systems, aiming for 1000+ nodes in the CPU case.
For an idea of more details, please refer to the HPC Tunathon Guide or the agenda of the workshop.

Prerequisites

In order to gain maximum benefit from this workshop, participants are required to have an in-depth (hands-on) knowledge of the algorithms and codes used in their project. We also encourage teams of two or more developers. Furthermore, if you do not already have your code running across multiple compute nodes, then the workshop is not for you. Similarly, if you want to investigate GPU scalability your code needs to already be GPU-capable.

In addition, we also ask participants to prepare test cases for different scales, e.g. to evaluate the initial performance of their code and to monitor progress when trying out modifications.

If you want to familiarise yourself with the available tools or prepare for the use of GPUs, take a look at the related workshops.

Application

The HPC Tunathon is open to all users but is limited to a small number of teams. The first step for participation is to apply as a code team, we thus expect one team member to register the team and provide details about the team and code that will be used. Individual participants will then be asked in a second step to register via the provided link after the acceptance by the organisers to provide personal details necessary to enter Forschungszentrum Jülich.

When you register, you will need to indicate the computer system(s) which have been used, with the maximum number of compute nodes and the execution configuration of processes and threads, and provide a graph or table of (strong or weak) scaling performance. If any tools have been used to investigate execution performance, brief details should also be included.

Contact

For any questions concerning the workshop please send e-mail to
d.broemmel@fz-juelich.de (Dr. Dirk Brömmel),
a.herten@fz-juelich.de (Dr. Andreas Herten),
and / or
b.wylie@fz-juelich.de (Dr. Brian J.N. Wylie)

Jülich Supercomputing Centre (JSC)
Institute for Advanced Simulation (IAS)
Forschungszentrum Jülich GmbH
52425 Jülich
Germany

Area Menu

Additional Information

Sponsors

Nvidia

Supporters

arm

Intel