Keywords: Multi Agent, Local Voting Protocol, Scheduling, Directed Acyclic Graph
Abstract: Scheduling computational workflows represented by directed acyclic graphs (DAGs) is crucial in many areas of computer science, such as cloud/edge tasks and data mining. The complexity of online DAG scheduling is compounded by the large number of computational nodes, data transfer delays, and the non-uniform arrival of tasks. This paper introduces the Multi-Agent Local Voting Protocol (MLVP), a novel approach focused on dynamic load balancing for DAG scheduling in heterogeneous computing environments, where executors represented as an agents. The MLVP employs a local voting protocol to achieve effective load distribution by formulating the problem as a differentiated consensus achievement. The algorithm calculates an aggregated DAG metric for each executor-node pair based on node dependencies, node availability, and executor performance. These metrics are optimized using a genetic algorithm to assign tasks probabilistically, achieving efficient workload distribution across the system and thus improving makespan. The effectiveness of the MLVP is demonstrated through comparisons with state-of-the-art DAG scheduling algorithm and popular heuristics such as DONF, FIFO, Min-Min, and Max-Min. Simulations show that MLVP achieves makepsan improvements of up to 70\% on specific graph topologies and an average makespan reduction of 23.99\% over DONF across various random DAGs. Notably, the algorithm's scalability is evidenced by enhanced performance with increasing numbers of executors and graph nodes.
Submission Number: 64
Loading