Prediction-based Coflow Scheduling

We explore the problem of minimizing the weighted average coflow completion time of coflows when flow sizes are unknown but unreliable predictions on them are available. We propose to use Sincronia, a 4-approximation algorithm when actual flow sizes are known, directly on the predictions. For this variant of Sincronia, we obtain a finite upper bound on the approximation ratio that combines unreliability on predictions and the weights of the coflows. On several numerical experiments, it is shown that this bound is too conservative, and that in practice Sincronia with predictions has a much better performance on average as well as in the worst-case.