By Dr. Dirk Pleiter, Research Group Leader, Jülich Supercomputing Centre
The Human Brain Project (HBP), a flagship project funded by the European Commission, has set itself an ambitious goal: Unifying our understanding of the human brain. To achieve it, researchers need a High-Performance Analytics and Compute Platform comprised of supercomputers with features that are currently not available, but OpenPOWER is working to make them a reality.
Through a Pre-Commercial Procurement (PCP) the HBP initiated the necessary R&D, and turned to the OpenPOWER Foundation for help. During three consecutive phases, a consortium of IBM and NVIDIA has successfully been awarded with R&D contracts. As part of this effort, a pilot system called JURON (a combination of Jülich and neuron) has been installed at Jülich Supercomputing Centre (JSC). It is based on the new IBM S822LC for HPC servers, each equipped with two POWER8 processors and four NVIDIA P100 GPUs.
Marcel Huysegoms, a scientist from the Institute for Neuroscience and Medicine, with support from the JSC could demonstrate soon after deployment the usability of the system for his brain image registration application. Exploiting the processing capabilities of the GPUs without further tuning, could achieve a significant speed-up compared to the currently used production system based on Haswell x86 processors and K80 GPUs.
Not only do the improved compute capabilities matter for brain research, but also by designing and implementing the Global Sharing Layer (GSL), the non-volatile memory cards mounted on all nodes became a byte addressable, globally accessible memory resource. Using JURON it could be shown that data can be read at a rate that is only limited by network performance. These new technologies will open new opportunities for enabling data-intensive workflows in brain research, including data visualization.
The pilot system will be the first system based on POWER processors where graphics support is being brought to the HPC node. In combination with the GSL it will be possible to visualize large data volumes that are, as an example, generated by brain model simulations. Flexible allocation of resources to compute applications, data analytics and visualization pipelines will be facilitated through another new component, namely the dynamic resource management. It allows for suspension of execution of parallel jobs for a later restart with a different number of processes.
JURON clearly demonstrates the potential of a technology ecosystem settled around a processor architecture with interfaces that facilitate efficient integration of various devices for efficient processing, moving and storing of data. In other words, it demonstrates the collaborative potential of OpenPOWER.