Memory Mapping with RL

Reinforcement-learning solution to the compiler problem of mapping tensors to memory layers, minimizing machine-learning execution time.

Specifications & Metrics

type

Learning Methods & Systems

originator

P. Wang et al.

introduced

2023

Core Principles & Analysis

• Targets the memory mapping problem that occurs during compilation of machine-learning programs: mapping tensors to different memory layers to optimize execution time.
• Resource scheduling problems of this kind range from congestion control to cloud computing, with significant impact on cost, device wear, and even carbon emissions.
• Reinforcement learning suits such sequential decision-making problems with combinatorial search spaces and high-dimensional inputs.

Sources

• Optimizing Memory Mapping Using Deep Reinforcement Learning (arXiv 2305.07440)