Communication Sparsity in Distributed Spiking Neural Network Simulations to Improve Scalability.

Area of Science:

• Computational Neuroscience
• Large-Scale Simulations
• Brain Function Research

Background:

• Big science projects increasingly use Spiking Neuronal Network (SNN) simulations for brain research.
• Distributed computing is essential for brain-scale SNN simulations, but communication overhead limits scalability.
• Computational efficiency decreases as more nodes are added, hindering simulation performance.

Purpose of the Study:

• To improve computational efficiency in distributed SNN simulations.
• To address communication bottlenecks in implicit synchronization, process handshake, and data exchange.
• To enhance the scalability of large-scale brain simulations.

Main Methods:

• Modeled Spiking Neuronal Networks (SNNs) as hypergraphs for connectivity-aware neuron allocation.
• Utilized hypergraph partitioning to minimize interprocess communication and increase communication graph sparsity.
• Implemented dynamic sparse exchange for efficient data transfer in sparse communication scenarios.

Main Results:

• Achieved significant gains in computational efficiency, up to 40.8 percentage points.
• Reduced overall simulation time by up to 73% through combined methods.
• Demonstrated the effectiveness of hypergraph-based allocation and dynamic sparse communication.

Conclusions:

• Hypergraph modeling and dynamic sparse communication effectively improve distributed SNN simulation performance.
• Optimized communication strategies are crucial for achieving scalable and efficient large-scale brain simulations.
• Findings are applicable to other distributed complex system simulations using graph-based communication models.