Supercomputers in the near future will be much more than just larger and more powerful machines than the ones used today. We are approaching exascale system – High Performance Computers that will be able to do 1018 floating point operations per second. With that, some paradigms of HPC will be-come obsolete. And energy consumption will become one of the key factors for the systems to come.
These systems will be very heterogeneous and complex on node level. They will leverage very different technologies, such as general-purpose CPUs as well as GPU and FPGA accelerators. For energy efficient computing with-out significant trade-off on the performance side, application will have to leverage these technologies dynamically. To achieve this, a new software stack is needed.
One of the new challenges for exascale is the total cost of ownership (TCO), and more specifically, the operating costs of exascale systems, the environmental footprint associated with the energy and power consumed, not only by the machine itself but also by the infrastructure supporting it, and power and performance inhomogeneity, stemming from manufacturing variability. The current state of practice in system wide resource manage-ment is rather passive in the sense that it merely serves user resource requests and applies brute-force enforcement of administrator-driven policies. What is needed is a holistic coordinator that will be able to sup-port power, energy, performance, and throughput optimizations across the whole system in a scalable manner.
In the race to greater performance using less energy, the focus has often been on hard-ware. However, system software can now play a crucial role towards a controlled balance between application performance, power constraints and energy consumption, taking advantage of software interfaces for hardware configuration, without neglecting application needs. REGALE intends to realize these benefits under real operational environments by equipping the HPC system soft-ware stack with the mechanisms and policies for effective performance, power and energy control and optimization, bridging the gap between research and practice.
The REGALE project started in April 2021. It aims to build a software stack that is not only capable of bringing efficiency to tomorrow’s HPC systems. It also will be open and scalable for massive supercomputers. REGALE brings together leading supercomputing stakeholders, prestigious academics, top European supercomputing centers and end users from critical target sectors, covering the entire value chain in system software and applications for extreme scale technologies.
