Flame GPU 2

Flame GPU 2 implements a parallel agent-based modeling framework that uses MPI and OpenCL to accelerate large-scale simulations of complex biological and physical systems on distributed GPUs.

Key Features:

• Parallelization: Uses MPI for distributed-memory parallel execution and OpenCL to leverage distributed GPUs for large-scale simulations.
• Agent representation: Represents agents as finite-state automata with memory (Coakley et al., 2006) and employs a state-machine-based representation to facilitate execution ordering and communication.
• Parallel communication and scheduling: Implements parallel communication routines and mechanisms for optimal ordering of agent execution to support efficient parallel runs.
• High performance: Demonstrated parallel efficiency above 80% in experiments, for example executing models with half a million agents on 432 processors.
• Data dependency analysis (planned): Includes planned enhancements for analyzing data flow between agents to inform optimization of parallel execution.
• Vector operations (planned): Plans to implement operations over groups of agents to improve computational efficiency.
• Dynamic task scheduling (planned): Plans to support dynamic task scheduling for more flexible and efficient distribution of tasks during simulation runs.

Scientific Applications:

• Bioreactor studies: Applied to simulate and analyze bioreactor processes.
• Immunogenic studies: Used in research related to immune responses.
• Epidermis modeling: Employed for simulating skin layers and their interactions.

Methodology:

Implements finite-state automata with memory and state-machine-based agent representations, uses optimal agent execution ordering and parallel communication routines, and runs parallel simulations via MPI and OpenCL on distributed GPUs.