EN FR
EN FR


Section: New Results

Load Balancing Management in a Distributed Task-Based Programming Model

Distributed task-based programming models such as StarPU optimize the execution of applications based on an initial distribution of data. The resulting computational load on each node may however evolve over the course of the application, to the point where this initial distribution of data leads becomes suboptimal. It becomes necessary to correct the distribution the distribution of data to rebalance the load among nodes. Tools such as Zoltan or ParMetis do exist to perform this rebalancing job. However, they cannot be employed without breaking the application execution flow, and force synchronizing steps in fundamentally asynchronous task parallelism paradigms. Within the context of the internship of Loïc Jouans, we proposed a mechanism to enable the detection of load imbalance as well as the application of corrective measures to rebalance it while preserving the execution asynchrony.