Reproduce “SLA-MORL: SLA-Aware Multi-Objective Reinforcement Learning for HPC Resource Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Seraj Al Mahmud Mostafa, Aravind Mohan, Jianwu Wang
SLA-MORL, an adaptive multi-objective reinforcement learning framework, optimizes GPU and CPU allocation for machine learning training in High-Performance Computing and cloud environments. It achieves 67.2% reduction in training time, 68.8% in operational costs, and a 73.4% improvement in SLA compliance by dynamically adjusting optimization priorities and leveraging historical learning.
Key insights
- Dynamically adjusting optimization weights based on real-time SLA violation severity is crucial for achieving high compliance and efficiency in multi-objective resource management.
- Intelligent initialization, whether through historical learning or efficient baseline experiments, significantly reduces initial exploration overhead, making RL practical for diverse and evolving workloads.
- A comprehensive state representation, including current resource allocation, utilization, training progress, and SLA compliance flags, is essential for informed and adaptive resource allocation by an RL agent.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.