JURECA Booster overview

General

The Booster module attached to JURECA provides a theoretical peak performance of 5 PetaFLOPS on 1640 Intel KNL nodes. A specification for the JURECA Booster module can be found on the Configuration page. The hardware design of the Booster means user codes need to be adapted to take advantage of the compute capabilities. To judge whether this is the case for your application code, as a guideline at least one of the following criteria should be met:

  • the compute kernels can profit from AVX512 vectorisation/SIMD
  • the code scales to 68 cores or more
  • the code relies on fast memory access and profits from MCDRAM

Access to the booster resources is handled by a separate JURECA partition. Please follow the guideline in the Quick introduction guide to choose the correct partition in your job script.

Resources

Numerous resources on the Intel KNL architecture and porting to it exist. In particular, we would like to refer to the following:

Preinstalled Software

A list of all software that is currently pre-installed on the Booster architecture can be found here. This does not imply that all software packages scale on the Booster for all inputs. We are building up experience in this area and will share it here as soon as possible.

To access the software installed for the KNL architecture you have to excecute the following command before loading any other module:

module load Architecture/KNL

If you want to cross-compile for the booster partition on the JURECA login nodes, make sure that you have loaded the Architecture/KNL as well. In addition you have to add the module load command in your job script.

Tools

We strongly advise you to make use of standard HPC performance tools when adapting and optimising code for the Booster. Examples for performance tools are Scalasca (Score-P), Vampir, Intel Inspector, Intel Advisor, and Intel VTune.