Electricity Cost Minimization for Workflows Scheduling in Geo-Distributed Data Centers

Geo-distributed data centers (GDCs) located around the world serve massive workflow applications, incurring high electricity cost. Since electricity prices of GDCs vary according to both geographical locations and time periods, different scheduling schemes result in varied electricity costs. How to reduce electricity costs while satisfying deadline constraints of workflows is critical for efficient operation of GDCs due to geographical and temporal variations in electricity prices. To solve the problem, a Frequency and Electricity Cost aware Multiple Workflows Scheduling algorithm (FECMWS) is proposed based on the Dynamic Voltage and Frequency Scaling (DVFS) technique, aiming at minimizing the electricity costs of all workflows while satisfying the deadline constraints. This algorithm first constructs the scheduling sequence of workflow tasks through three stages, including workflow sequencing, deadline partition and task sequencing. To approach the global optimum of scheduling objective, the algorithm adopts two graph embedding models and a policy network to solve the Markov Decision Process (MDP) of task resource allocation, assigning VMs to each task in the sequence. Experiments are conducted based on a large number of randomly generated scientific workflow instances and the results are analyzed using multi-factor analysis of variance (ANOVA). The FECMWS is calibrated and then compared to existed methods. The results demonstrate the effectiveness of the proposal.