Energy-efficiency and applications
A structured software stack and advice on efficient supercomputing
Moore’s law continues to apply. According to this law, the computing power of HPC systems doubles in a period of approx. 18 months. This results in a dramatic increase in the degree of parallelisation on the application level, and accelerator technologies are being also used to increase computing power. That is why users are confronted with increasingly complex HPC systems; their applications and codes should not just be able to connect to and control different processor types and hierarchical storage technologies, such as processor registers, L1 and L3 caches and main and background storages, but should also be as energy-efficient as possible. In collaboration with partner organisations and the Gauss Centre for Supercomputing (GCS), the LRZ has therefore expanded its training and workshop programme in recent years by placing focus on programming schemes. Experienced mentors from the HPC application labs also supervise research groups and support them from the point at which their application is made to the point at which their code is ported and executed.
Active support for programming and porting
The complexity of new computing resources increases the amount of effort involved in programming and porting on LRZ’s HPC systems: At the LRZ, HPC users have access to a structured software stack with libraries that have been optimised for available HPC and AI systems, compliers, parallel runtime environments and open-source and commercial applications. In close collaboration with the Munich Centre of Advanced Computing, LRZ employees are also developing tools to automatically profile application performance and efficiency. The Competence Network for Scientific High-Performance Computing in Bavaria (KONWIHR) provides further support for programming, optimising and implementing scientific code. This network, to which the LRZ also belongs, develops and optimises scalable programmes for simulations, calculations or data analysis based on of tried and tested principles of software, algorithm and performance engineering.
The use of and work with the Bavarian Energy, Architecture and Software Testbed (BEAST) has also proven its worth in these tasks: This is where benchmarks for sustainable, efficient HPC can be created and codes tested in conjunction with hardware components.
