The paper addresses an important question of parallel programming: the degree to which an application should be parallelized to keep parallelization overhead low. This is also known as the time improvement versus tradeoff problem.

This problem is studied in the context of real-time program scheduling on multicore architectures. An auto-tuning approach is suggested for optimizing four performance-critical parameters of every task to be scheduled: artificial period, artificial deadline, offset, and parallelization.

The authors suggest a novel three-step heuristic-based approach for determining the optimized values of the parameters, which is efficient also in cases of many tasks to schedule: in step 1, optimized parameter values are determined for each individual task; in step 2, the values are determined for groups of tasks with the same time period; and in step 3, values are determined for all tasks together.

The experimental evaluation shows that the suggested approach significantly improves the state of the art: up to 80 percent more tasks can be scheduled within the same time period.

The paper studies an important problem of parallel programming and can be interesting for researchers and practitioners in the field. Unfortunately, the paper is not always well written, for example, important definitions are often only vaguely formulated or come too late.