CVaR-Based Adaptive Scheduling of Fork–Join Workflows with Selective Task Replication in Edge–Cloud Environments under Computation and Network Uncertainty
CVaR-Based Adaptive Scheduling of Fork–Join Workflows with Selective Task Replication in Edge–Cloud Environments under Computation and Network Uncertainty
Ali Azimi1 Amin Babazadeh Sangar2
1) Department of Computer Engineering, Ur.C., Islamic Azad University, Urmia, Iran
2) Department of Computer Engineering, Ur.C., Islamic Azad University, Urmia, Iran
IOI Code
:
AAHD-ABKBKG
Publication :
National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
The increasing adoption of heterogeneous edge–cloud infrastructures for executing scientific workflows has highlighted the need for scheduling approaches that are robust against fluctuations in computational and communication resources. In workflows characterized by fork–join structures, a delay in a single branch can significantly postpone the completion of the corresponding join node and, consequently, the entire workflow.
This study introduces TRAFJS, an adaptive, multi-objective, and tail-risk-aware scheduling algorithm designed for heterogeneous edge–cloud environments. The proposed algorithm combines the HEFT rank with the structural positions of tasks within fork–join patterns, critical-path information, deadline urgency, and branch-delay risk quantified using the Conditional Value-at-Risk (CVaR) metric. In addition, selective task replication is applied only to high-risk tasks that have a substantial impact on join nodes, while event-driven rescheduling is triggered in response to changes in processing times or network bandwidth.
The proposed algorithm was evaluated using the Montage, SIPHT, CyberShake, and Epigenomics scientific workflows, with workflow sizes of 50, 100, and 200 tasks, under uncertainty levels of 0.1, 0.3, and 0.5. Each experimental configuration was executed independently 30 times, and the performance of TRAFJS was compared with that of HEFT, MFJS, and CPOP.
The results show that TRAFJS achieved an average makespan of 11.403, representing reductions of 2.21%, 2.54%, and 15.52% compared with HEFT, MFJS, and CPOP, respectively. The delay CVaR was reduced to 1.514, which was 31.74%, 34.67%, and 71.77% lower than the corresponding values obtained by the three baseline algorithms. TRAFJS also achieved a deadline violation rate of 12.5%, corresponding to reductions of 33.17%, 35.41%, and 79.70%, respectively.
This improvement in reliability was accompanied by limited increases of 4.75% in energy consumption and 2.52% in execution cost compared with HEFT. Overall, by establishing a controlled trade-off among makespan, tail risk, energy consumption, and cost, TRAFJS improves the robustness and stability of fork–join workflow execution under uncertain edge–cloud conditions.