Section: New Results

Data Aware Batch Scheduling

We obtained in 2018 two important results on on-line scheduling using resource augmentation. The main idea is that the algorithm is applied to a more powerful environment than that of the adversary. We focused more specifically on the mechanism of rejection based on the concept of duality for mathematical programming applied for the analysis of the algorithm’s performance. More precisely, we proposed a primal-dual algorithm for the online scheduling problem of minimizing the total weighted flow time of jobs on unrelated machines when the preemption of jobs is not allowed. This analysis concerned usual sequential jobs. These results have been distinguished among the most significant ones on the annual ACM review of on-line algorithms. We extended this work on a practical side by applying the analysis to actual batch schedulers with parallel jobs, rejection was interpreted as redirecting jobs to some predefined machines.

Machine Learning is a hot topic which received recently a great attention for dealing with the huge amount of data produced by the explosion of the digital applications and for dealing with uncertainties. The members of DataMove promoted a methodology based on simulation and machine learning to obtain efficient dynamic scheduling policies. The main idea is to focus the learning scheme targeting the policies them-selves, and not the specific parameters of the problem. Today, this methodology is mature and it is applied in several project like ANR Energumen (performances and replaced by energy saving). We also launched a new project at MIAI on edge Intelligence. The idea is to propose an alternative to the high-consuming classical IA by doing most of the computations close the the place where the data are produced. We are developing both an efficient task orchestration framework and distributed learning algorithms.

We wrote a survey [20] on scheduling on heterogeneous machines where we provided a complete benchmark suite and we recoded all existing algorithms and compared them.