Accelerating the Gillespie Exact Stochastic Simulation Algorithm using hybrid parallel execution on graphics...

Area of Science:

• Computational Biology
• Biophysics
• Chemical Kinetics

Background:

• The Gillespie Stochastic Simulation Algorithm (GSSA) is crucial for simulating biochemical reactions with low reactant concentrations, where deterministic methods fail.
• Serial GSSA implementations are computationally expensive, limiting their use for complex models and parameter sweeps.
• Stochastic simulation is essential for understanding inherent randomness in biological systems.

Purpose of the Study:

• To develop a novel variant of the exact GSSA that leverages graphics processing units (GPUs) for significant acceleration.
• To overcome the computational bottlenecks of traditional GSSA methods, enabling faster and more extensive simulations.
• To improve the efficiency of parameter sweep exercises in stochastic reaction kinetics.

Main Methods:

• Parallelization of a single GSSA realization across GPU threads (fine-grained parallelism).
• Coarse-grained parallelism by executing multiple warps on different multi-processors to generate multiple trajectories simultaneously.
• Development of novel data structures and algorithms to minimize memory traffic, a key bottleneck in GSSA.

Main Results:

• Achieved performance gains of 8×-120× compared to state-of-the-art serial algorithms.
• Demonstrated significant speedups across various model types.
• Successfully reduced the computational cost of simulating stochastic reaction kinetics.

Conclusions:

• The novel GPU-accelerated GSSA offers a substantial performance improvement over serial methods.
• This approach makes complex stochastic simulations and large-scale parameter sweeps computationally feasible.
• The optimized GSSA variant is a valuable tool for computational biology and related fields.