Optimizing NEURON Simulation Environment Using Remote Memory Access with Recursive Doubling on Distributed Memory...

Area of Science:

• Computational Neuroscience
• High-Performance Computing

Background:

• Complex neuronal network models require efficient simulation environments like NEURON.
• Interprocessor communication, particularly spike exchange, is a bottleneck in parallel simulations.
• Current Message Passing Interface (MPI) usage, specifically MPI_Allgather, limits scalability with increased processors.

Purpose of the Study:

• To improve the efficiency of the NEURON simulation environment for large-scale neuronal networks.
• To reduce the communication overhead associated with spike exchange in distributed memory systems.
• To enhance the performance of NEURON on parallel computing architectures.

Main Methods:

• Modified the MPI_Allgather collective for spike exchange.
• Implemented Remote Memory Access (RMA) to transition from two-sided to one-sided communication.
• Utilized a recursive doubling mechanism for efficient, step-wise processor communication.

Main Results:

• Reduced communication time for spike exchange over distributed memory systems.
• Enhanced communication concurrency among processors.
• Improved the overall runtime performance of the NEURON simulation environment.

Conclusions:

• The optimized communication methodology significantly boosts NEURON's efficiency for large neuronal network simulations.
• The use of RMA and recursive doubling presents a scalable solution for parallel neuroscience simulations.
• This work contributes to advancing computational neuroscience by enabling more efficient modeling of complex neural systems.