Simulation of reaction diffusion processes over biologically relevant size and time scales using multi-GPU...

Area of Science:

• Computational Biology
• Biophysics
• Biochemistry

Background:

• Simulating in vivo cellular processes using the reaction-diffusion master equation (RDME) is computationally intensive.
• Previous single-GPU implementations were limited to small systems or short timescales.
• Simulating larger eukaryotic systems or long timescales on single GPUs is impractical due to memory and processing constraints.

Purpose of the Study:

• To develop a scalable multi-GPU parallel implementation of the MPD-RDME method.
• To overcome the memory and computational limitations of single-GPU simulations for complex cellular processes.
• To enable efficient simulation of larger biological systems and longer time scales.

Main Methods:

• Developed a multi-GPU parallel implementation of the MPD-RDME method using spatial decomposition.
• Incorporated dynamic load balancing to accommodate GPUs with varying performance and memory.
• Utilized CUDA for high-performance peer-to-peer GPU memory transfers.

Main Results:

• Demonstrated parallel efficiency and performance gains with increasing system size, particle counts, and reaction numbers.
• Successfully simulated the Min protein system in E. coli using the multi-GPU approach.
• Evaluated algorithm performance on state-of-the-art GPU devices.

Conclusions:

• The new multi-GPU implementation significantly enhances the computational feasibility of RDME simulations.
• The spatial decomposition and dynamic load balancing approach is effective for heterogeneous GPU environments.
• This method can be generalized to other lattice-based computational problems in biology and beyond.