Robustness to Dynamic Worker-Speed Variation

Determine whether adaptive workload balancing remains effective when relative worker-speed ratios vary substantially during training, and establish whether profiling must be repeated during training under such conditions.

Background

Adaptive workload balancing profiles each worker once at initialization and assumes that relative worker-speed ratios remain stable throughout training. Although using speed ratios mitigates common-mode throughput changes caused by thermal state or background load, extreme speed variation could invalidate the fixed workload assignment. The paper identifies repeated profiling during training as a possible solution but does not establish its effectiveness or necessity.

References

Under extreme variations in worker speed during training, running the profiling again during training may be needed, which remains future work.

GeoMesh: Workload-Balanced and Sign-Compressed Geo-Distributed LLM Training  (2609.18388 - Shin et al., 16 Sep 2026) in Section 6, Limitations, subsection “AWB profiling assumptions”