Extremely Scalable Spiking Neuronal Network Simulation Code: From Laptops to Exascale Computers.

Area of Science:

• Computational Neuroscience
• High-Performance Computing
• Neuroscience Software Development

Background:

• Current neuronal network simulation tools can model large networks, up to 10% of the human cortex.
• At this scale, network connectivity becomes sparse due to connection limits per neuron.
• Efficiently managing computational costs requires exploiting this sparsity in simulation software.

Purpose of the Study:

• To develop and implement a computational framework that exploits sparsity in brain-scale neuronal networks.
• To enhance simulation software for efficient execution on future high-performance computing systems.
• To ensure performance is maintained across different system scales.

Main Methods:

• Introduction of a two-tier connection infrastructure.
• Development of a directed communication framework for compute nodes.
• Implementation of the framework within the NEST simulation code.
• Performance evaluation across various network simulation scaling scenarios.

Main Results:

• Demonstrated feasibility of the proposed two-tier connection and communication framework.
• The implemented technology effectively accounts for network sparsity in large-scale simulations.
• Performance analysis confirmed the approach's suitability for post-petascale computing without compromising smaller systems.

Conclusions:

• The new data structures and communication scheme are essential for brain-scale neural network simulations.
• The developed framework prepares the NEST simulation kernel for future high-performance computing environments.
• This approach ensures efficient and scalable simulations of complex neural networks.