StochShed is a research project that consists in investigating classic scheduling strategies for HPC (High Performance Computing) and in analyzing their performance on average on instances whose distribution is known for common scheduling problems. It fits into the theme of predicting the performance of complex systems represented by data centers.

The objective is to define the allocation of applications to machines in order to obtain the best possible execution performance. Many studies propose algorithms and validate them by measuring the gains obtained compared to existing solutions. To carry out these measurements, simulations are often used which require implementing the scheduling algorithm and having an application model. In the considered problem, the application is modeled by an execution cost for each task that depends on the machine that executes it. The way to generate these instances for the simulation can greatly influence the results. Beyond this general evaluation approach, it is relevant to focus on an analysis further upstream, which allows directly characterizing the performance of the studied algorithms.

A complementary axis will consist in applying artificial intelligence algorithms to generate difficult instances (a network generating instances coupled to a network discriminating the difficulty of the instances) or to generate realistic instances (the second network discriminating the realism of the instances).

This project is held by FEMTO-ST (LC Canon and is supported by the Bourgogne Franche-Comté region and the National Research Agency (ANR):

One Ph.D student has been recruited for this project.