Asynchronous Branch-Parallel Simulation of Detailed Neuron Models.

Area of Science:

• Computational Neuroscience
• Neuroscience
• Computer Science

Background:

• Simulating detailed neuron models is crucial for understanding brain function.
• Current simulation methods face challenges in strong scaling and utilizing multi-core architectures.
• Load imbalance and inefficient resource utilization hinder progress in large-scale neural simulations.

Purpose of the Study:

• To develop an advanced computational strategy for accelerating detailed neuron network simulations.
• To overcome limitations in load balancing and resource utilization on distributed multi-core systems.
• To enable faster and more scalable simulations for studying synaptic plasticity and long-term neural dynamics.

Main Methods:

• Extracted flow-dependencies between ODE parameters and algebraic solvers for individual neurons.
• Developed techniques for memory, communication, and computation reorganization for load-balanced, asynchronous execution.
• Implemented a new computation model leveraging tree-based parallelism and SIMD acceleration within the NEURON simulation environment using the HPX runtime system.

Main Results:

• Achieved asynchronous distributed and parallel simulation of single neurons to medium-sized networks.
• Demonstrated improved strong scaling properties and finer-grained parallelism compared to state-of-the-art methods.
• Reduced time to solution across various distributed multi-core architectures.

Conclusions:

• The proposed strategy significantly accelerates neural network simulations.
• The new computation model effectively balances workload and utilizes computing resources.
• This advancement facilitates more extensive and efficient research in computational neuroscience.