Reinforcement Learning for Load-Balanced Parallel Particle Tracing.

Area of Science:

• High-performance computing
• Computational science
• Parallel algorithms

Background:

• Particle tracing is crucial for simulations in fluid dynamics, oceanography, and meteorology.
• Optimizing parallel performance in distributed-memory systems remains a significant challenge.
• Dynamic load balancing and communication cost reduction are key to efficient parallel processing.

Purpose of the Study:

• To develop an online reinforcement learning (RL) paradigm for dynamic optimization of parallel particle tracing.
• To improve performance metrics such as parallel efficiency, load balance, and I/O and communication costs.
• To create adaptive algorithms that handle diverse simulation data and scales.

Main Methods:

• Implemented an RL-based work donation algorithm to balance process workloads.
• Developed a high-order workload estimation model for predictive process load assessment.
• Designed a communication cost model accounting for block and particle data transfer.
• Integrated these components into a novel online RL framework.

Main Results:

• The RL algorithm dynamically optimized parallel particle tracing performance.
• Significant improvements were observed in parallel efficiency and load balance.
• Reductions in I/O and communication costs were achieved.
• The method demonstrated adaptability across various large-scale simulation datasets.

Conclusions:

• The proposed online RL paradigm effectively optimizes parallel particle tracing in distributed-memory systems.
• The combination of work donation, workload estimation, and communication cost modeling leads to superior performance.
• This approach offers a scalable and adaptive solution for large-scale scientific simulations.