(4) Modular Supercomputing
[PIs: Amunts, Kirchner, Lippert, Luu, Memmesheimer, Schweitzer, Suarez, Urbach]
One promising paradigm towards building an exascale computer architecture is called “Modular Supercomputing”. Driven by the diverse needs of the various fields of science involved in CASCADE, we will investigate and implement this concept. Instead of hosting one or two monolithic HPC systems, we aim for a variety of compute modules deeply integrated into a single HPC infrastructure. While at least one module will be based on general-purpose processors, there will be modules e.g. based on accelerators, optimized for high floating point operation throughput, or based on hardware with highly performant access to large data arrays, optimized for data analytics. Facilitated by a high performance interconnect, applications can map their algorithmic content to the architecture using the optimized modules. Visualization is a prime example for a class of problems which needs to use several of these modules. This infrastructure must be fully transparent for the user and it must support workflows where applications can flexibly be executed on different modules. This development will play an important role for the envisaged exa-scale installation at JSC, from which research in CASCADE will benefit.
