Leveraging a Cluster-Booster Architecture for Brain-Scale Simulations

The European Dynamical Exascale Entry Platform (DEEP) is an example of a new type of heterogeneous supercomputing architecture that include both a standard multicore-based "Cluster" used to run less scalable parts of an application, and an Intel MIC-based "Booster" used to run highly scalable compute kernels. In this paper we describe how the compute engine of the widely used NEURON scientific application has been ported on both the DEEP and the Intel MIC platform. We discuss the design and implementation of the core simulator with an emphasis on the development workflow and implementation details that enable the efficient use of the new "Cluster-Booster" type of architectures. We describe optimizations of the data structures and algorithms tailored to the Intel Xeon Phi coprocessor which contributed to improve the overall performance of NEURON by a factor 5. Validation of our implementation has first been done on STAMPEDE supercomputer in order to emulate the DEEP architecture performance. Building on these results, we then explored opportunities offered by the DEEP platform to efficiently support complex scientific workflow.