Constructing Neuronal Network Models in Massively Parallel Environments.

Area of Science:

• Computational Neuroscience
• High-Performance Computing
• Artificial Neural Networks

Background:

• Advancements in data structures allow for large-scale spiking neuron network simulations on petascale computers.
• Efficiently utilizing supercomputer power for neuronal network creation is crucial for advancing brain-scale simulations.

Purpose of the Study:

• To investigate the scalability and performance of neuronal network construction on supercomputers.
• To identify bottlenecks in current network creation algorithms and propose optimizations for enhanced performance.

Main Methods:

• Divided simulation runtime into network construction and dynamical state advancement phases.
• Evaluated process-parallel and thread-parallel network creation on multi-core nodes.
• Analyzed scaling of neuron and connection instance creation for varying network sizes.
• Investigated memory allocation strategies and loop order for construction algorithms.

Main Results:

• Process-parallel creation scales well but has high memory consumption; thread-parallel creation has low memory overhead but limited speedup.
• Neuron/connection instance creation algorithms scale well up to 10,000 neurons but not for millions.
• Inadequate memory allocation hinders thread-parallel scaling; optimized allocators restore excellent scaling.
• Improved loop order analysis enhances locality, significantly reducing runtime and improving scaling.

Conclusions:

• Optimized memory allocation and loop order strategies dramatically improve neuronal network construction performance on supercomputers.
• These combined techniques increase performance by an order of magnitude, enabling better utilization of parallel computing power for brain-scale simulations.