.png){kind=icon}
.png){kind=icon}
Return
Machine learning and data analysis method for predicting an efficient algorithm for heterogeneous multicore scheduling
.png){kind=icon}
PRE
AI
DOI：10.1016/j.cor.2026.107478.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
We propose a machine learning method to predict an efficient algorithm for scheduling tasks on hybrid multiprocessor systems with precedence constraints. The prediction is also based on data analysis, with the aim of identifying the influencing parameters in the input. Each task has processor-dependent processing times, and the objective is to construct a feasible non-preemptive schedule that minimizes the makespan. We define 129 features and benchmark seven state-of-the-art scheduling heuristics on synthetic instances generated using controllable DAG and processing time generators. The collected performance data are then used to train multi-output regression models, which predict the makespan and scheduling runtime of each heuristic for new instances. An analysis of feature contributions shows that 57 features account for 94% of cumulative model importance. At runtime, the heuristic minimizing a weighted combination of these two predicted metrics is automatically selected. Experimental results demonstrate that this approach, combining efficient prediction models and systematic instance generation, substantially improves scheduling performance compared to relying on any single heuristic alone. Additional experiments indicate that, despite substantially higher computational efforts, metaheuristic refinements yield only marginal improvements over the proposed algorithm.
Keywords:
Machine learning
Data analysis
Hybrid multiprocessor scheduling
Makespan minimization
Journal
C
IF:
4.3
Papers:
197
Citations:
0