A scheduling algorithm based on reinforcement learning for heterogeneous environments

Efficient application scheduling is critical for achieving high performance in heterogeneous computing environments. For heterogeneous static scheduling problems, applications with a set of distributed tasks are modeled as directed acyclic graphs (DAG). The goal of scheduling is to properly allocate...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Applied soft computing Ročník 130; s. 109707
Hlavní autoři: Lin, Ziniu, Li, Chen, Tian, Lihua, Zhang, Bin
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier B.V 01.11.2022
Témata:
DAG scheduling
Heterogeneous systems
Hyperparameter optimization
Reinforcement learning
Heterogeneous systems
Hyperparameter optimization
Reinforcement learning
DAG scheduling
ISSN:1568-4946, 1872-9681
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Efficient application scheduling is critical for achieving high performance in heterogeneous computing environments. For heterogeneous static scheduling problems, applications with a set of distributed tasks are modeled as directed acyclic graphs (DAG). The goal of scheduling is to properly allocate processors to DAG nodes and minimize the average completion time of all DAG applications. Most existing scheduling algorithms cannot fully utilize scheduling information, and are designed to schedule one DAG application at a time. Therefore, this study proposes a scheduling algorithm based on reinforcement learning for heterogeneous computing environments. A convolutional neural network with graph attention is first used to fully process and encode the scheduling information. Then, a fully connected neural network selects an appropriate node from the DAG applications for execution. Finally, a heuristic method is employed to allocate a processor to the selected node based on the earliest finish time and node duplication. This process is then repeated for every DAG node. The proximal policy optimization algorithm is used to tune all network hyperparameters. The training phase employs a reward function suitable for multi-DAG scheduling to ensure optimal performance when multiple DAG applications are scheduled concurrently. Experiments of various scheduling scenarios and types of applications were conducted, and the results show that the proposed algorithm reduces the average termination time of applications by 13.03% over that of existing algorithms. •A graph neural network is adopted to process application structure.•The proximal policy optimization is applied to train neural networks.•A heuristic method is used to allocate processors and reduce the decision space.•The learning environment suitable for multi-application scheduling is applied.
ISSN:1568-4946
1872-9681
DOI:10.1016/j.asoc.2022.109707