Energy-efficiency and software
Effectively managing HPC systems and the virtualisation of IT technology
The continuous increases in processors performance means that the potential of server systems at computing centres is often underexploited. Managing energy in a targeted manner using virtualisation can help to compensate for this imbalance: Operating several virtual instances on one physical server hardware not only saves on space. The performance of the available devices can be flexible aligned to current and future capacities. The virtualisation of technical resources makes a computing centre not only more flexible, but more sustainable too. This represents an important strategic component for sustainable energy management in the area of ‘Virtual Infrastructure Services’, but also when it comes to storage and data network technology at the LRZ.
Software stack requirements
The LRZ’s multiple HPC systems have thousands of processors and computing cores working in tandem – a challenge for the software stack, especially for the operating software and scheduling tools used to manage HPC system capacities. The tools and programmes also have to ensure that the mechanisms built into the hardware to reduce power consumption work perfectly.
The LRZ supports these tasks by using frameworks such as the Energy Aware Runtime (EAR). Because many scientific applications hardly benefit from a higher clock rate, the HPC systems at the LRZ are mostly operated at a reduced clock rate, which allows them to work more efficiently Without this control, the computers would consume to 30 per cent more power on average per year.
Other possibilities for reducing the power consumed by high-performance computers include scheduling or planning supercomputing tasks in a targeted way. The LRZ relies on SLURM, as well as energy-aware scheduling, for which research and technology are developing more and more tools. Computing tasks are combined in a way that keeps storage devices and processors as busy as possible. Goal: To automatically divide individual work orders so that all components of a system are kept equally busy at low clock rates. Power consumption can also be reduced in this way.