LANS Seminar

Seminar Title: Towards Scalable Federated Learning for Scientific Computing

Speaker: Yijiang Li, Postdoctoral Appointee, Mathematics and Computer Science Division, Argonne National Laboratory

Date: Thursday, March 26, 2026

Time: 3:00 PM-4:00 PM (In-Person)

Location: Hybrid, Bldg. 240, Conference Room 4301

Description: Federated learning (FL) enables collaborative model training without centralizing sensitive data, but making FL work in practice requires more than better algorithms. It requires algorithms co-designed with the infrastructure they run on. This principle becomes especially apparent in high-performance computing (HPC) environments, where heterogeneous architectures, strict firewall policies, and batch job schedulers introduce challenges qualitatively different from those studied in the existing FL literature. We present a systematic deployment of FL across multiple supercomputers, taking four DOE facilities as a case study representative of general HPC ecosystems. Extensive experiments expose key system-level bottlenecks and reveal that stochastic batch scheduler delays critically affect training performance,  a challenge for which existing FL algorithms, designed without infrastructure in mind, are poorly suited. This motivates FedQueue, a queue-aware FL protocol that co-designs training and aggregation directly with scheduler behavior, supported by convergence guarantees and demonstrated improvements over baselines in these deployments.