Federated Learning Baselines

FedCompass profiles stable compute-throughput heterogeneity rather than stochastic scheduler admission delay. Buffered and semi-asynchronous methods generally rely on bounded staleness assumptions or stale-update filtering. FedAsync reached early loss thresholds quickly but plateaued at a worse final loss in the production experiment. Synchronous FedAvg is vulnerable in cross-facility HPC because the slowest queued facility determines round progress. Fully asynchronous federated learning avoids blocking but can aggregate very stale updates when queue times spike.