Meeting the memory challenges of brain-scale network simulation.

Area of Science:

• Computational Neuroscience
• Neuroinformatics
• High-Performance Computing

Background:

• Studying the brain connectome requires simulation software capable of handling diverse scales, from local circuits to whole-brain interactions.
• Existing simulation technologies for large-scale brain networks lack detailed descriptions, hindering reproducibility and accessibility.
• Memory consumption becomes a critical bottleneck on modern supercomputers as network models approach meso- and macroscale simulations.

Purpose of the Study:

• To analyze memory consumption in neuronal simulators as a function of network size and computational resources.
• To identify key software components contributing to memory saturation in large-scale neural network simulations.
• To develop and implement strategies for reducing memory footprint to enable scalable brain simulations.

Main Methods:

• Development of a linear model to analyze memory consumption of neuronal simulator components.
• Application of the model to the Neural Simulation Tool (NEST) to identify dominant memory-consuming components at different scales.
• Implementation of data structures that exploit network sparseness to reduce memory usage.

Main Results:

• The linear model effectively predicts memory consumption and identifies critical components for optimization.
• Strategies exploiting network sparseness significantly reduce memory requirements for large-scale simulations.
• Optimized NEST software demonstrates scalability to 10,000 processors and beyond.

Conclusions:

• Memory consumption is a primary challenge for large-scale connectome simulations on modern supercomputing architectures.
• The developed analytical model and optimization strategies provide a framework for efficient neuroinformatics software development.
• These findings are crucial for advancing the capabilities of the Human Connectome Project and related research.