Load balancing for multi-threaded PDES of stochastic reaction-diffusion in neurons.

Area of Science:

• Computational Neuroscience
• Biophysics
• Scientific Computing

Background:

• Stochastic simulation of neuronal processes is crucial for understanding molecular dynamics.
• Existing simulation methods may face challenges with computational efficiency for complex neuronal models.

Purpose of the Study:

• To develop and evaluate a multi-threaded simulator, Neuron Time Warp-Multi Thread, for stochastic reaction-diffusion processes in neurons.
• To apply Q-Learning and Simulated Annealing for optimizing dynamic load balancing and window control within the simulator.

Main Methods:

• Development of a multi-threaded parallel-distributed event simulator (PDES) for neuronal simulations.
• Implementation of Q-Learning and Simulated Annealing algorithms to tune parameters for dynamic load balancing.
• Collection and analysis of runtime statistics from simulation threads to enable workload migration.

Main Results:

• Both Q-Learning and Simulated Annealing improved execution times for small simulations, by up to 31% and 19% respectively.
• Simulated Annealing demonstrated superior performance for larger populations, reducing execution time by 41%.
• The Neuron Time Warp-Multi Thread simulator effectively utilized multi-threading for enhanced computational performance.

Conclusions:

• Q-Learning and Simulated Annealing are effective AI techniques for optimizing parallel simulation parameters in computational neuroscience.
• The developed simulator and optimization methods offer significant speedups for stochastic neuronal simulations.
• This approach enhances the feasibility of large-scale, realistic molecular dynamics simulations within neurons.