Reproduce “Deep Learning Workload Scheduling in GPU Datacenters: Taxonomy, Challenges and Vision” 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
Wei Gao, Qinghao Hu, Zhisheng Ye, Peng Sun, Xiaolin Wang, Yingwei Luo, Tianwei Zhang, Yonggang Wen
A taxonomy of deep-learning workload schedulers for GPU datacenters connects training, inference, hyperparameter optimization, and hybrid workloads within a common systems framework. The survey compares objectives, resource-management techniques, and deployment challenges, emphasizing the growing role of profiling, heterogeneous-resource modeling, GPU sharing, elastic execution, adaptive batching, and lifecycle-wide coordination.
Key insights
- GPU availability alone is an incomplete measure of capacity: CPU, memory, storage, network bandwidth, accelerator type, and interconnect locality can determine actual deep-learning performance.
- Sharing and elasticity improve utilization, but their benefits depend on interference, workload combinations, checkpointing cost, gang constraints, and the ability to adapt quickly to changing allocations.
- Training and inference should increasingly be coordinated across the application lifecycle, allowing idle serving capacity to support elastic training while preserving latency and accuracy objectives.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.