Work

TANGO: Inter-Job and Intra-Job Orchestration for Distributed Learning on Heterogeneous Wireless IoT Clusters

Year: 2026
Type: preprint
Abstract: Distributed learning on wireless Internet of Things (IoT) clusters requires coordination of shared communication resources and heterogeneous device capacities. Concurrent synchronous training jobs can generate overlapping communication bursts, while memory limits and thermal throttling alter their execution rates. We present TANGO, a hierarchical closed-loop co-scheduler that couples inter-job synchronization control with intra-job runtime adaptation. The inter-job controller coordinates collective timing and adjusts gradient accumulation to regulate synchronization frequency using congestion feedback. The intra-job controller adapts batch size, learning rate, and input-pipeline settings within an admissible configuration region, while reporting telemetry that updates inter-job decisions. We evaluate TANGO on a 12-node heterogeneous Raspberry Pi 5 testbed with nine vision and text workloads. Under a 50-job Poisson trace, TANGO reduces average job completion time by 16.8\%, makespan by 18.0\%, and P99 queueing delay by 19.7\% relative to Sia. Differences in final accuracy from Sia remain below 1\% point for the three reported workloads at the evaluated training budgets. Trace-driven simulations extend the evaluation to 2{,}048 nodes. These results support the usefulness of coupled communication and execution control for the evaluated contention-heavy edge-learning setting.
Language: English
Open Access status: green
OpenAlex ID: W7235177485
Cites: 34
Cited by:
Related to: 10