FedQueue protocol optimizes federated learning for HPC facilities

Researchers have developed FedQueue, a novel protocol designed to optimize federated learning across multiple High-Performance Computing (HPC) facilities. This system addresses the significant delays caused by batch scheduler queues, which can dominate training time. FedQueue incorporates queue delay predictions and cutoff-based admission to manage local work and buffer late arrivals, thereby bounding update staleness. The protocol also employs staleness-aware aggregation to stabilize heterogeneous workloads, leading to improved convergence and reduced training time. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Optimizes federated learning efficiency in distributed HPC environments, potentially reducing training times for large-scale AI models.

RANK_REASON This is a research paper detailing a new protocol for federated learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• FedQueue
• HPC
• arXiv

• paper
• infra