Deep learning (DL) has found extensive application in supporting various mobile applications. The efficient execution of DL tasks is paramount for ensuring the effectiveness of AI-driven mobile applications. While previous research has predominantly focused on minimizing the completion time of DL tasks, the associated cost of execution has often been overlooked. Nonetheless, cost becomes a critical factor, particularly when utilizing DL infrastructure rented from third-party cloud service providers. In this paper, we propose a cost-aware resource scheduling framework named CRS for orchestrating DL task execution in mobile cloud systems. Our aim is to minimize server rental costs by strategically orchestrating DL jobs with diverse deadlines and workload scales across rented cloud servers. We formally define the problem and prove its NP-hardness by reducing it to a multiple knapsack problem (MKP). To solve this problem, we devise an approximation algorithm with a guaranteed upper bound performance ratio of$1+\frac{1}{e-1}$. We evaluate CRS against state-of-the-art baselines through simulations of various job arrival scenarios in a real elastic mobile cloud system. The results demonstrate that CRS, on average, reduces rental costs by 45.1% compared to other baselines, while simultaneously achieving a shorter average job completion time (JCT) and maximum job completion time (i.e., makespan).

With the increasing demand for artificial intelligence (AI), cluster jobs require high-performance GPU instances and often face stringent deadline constraints. Previous studies have proposed various scheduling strategies to minimize job completion time and instance costs, assuming accurate prediction of job execution time across different instances. However, such assumptions are often unrealistic due to the inherent unpredictability of job execution time. Additionally, jobs typically involve large data sets and incur substantial cold-start time, leading to increased job completion latency and degradation of user quality of service. To tackle these challenges, we introduce an algorithm named Exponential Weighting Scheduling (EWS) for GPU clusters, which employs an online learning approach based on the combinatorial multi-armed bandit (CMAB) framework. EWS dynamically updates the probability distribution of decision spaces using real-time information on job performance across different instances and employs randomized decision-making for scheduling. Furthermore, we provide theoretical proof that this online strategy guarantees sublinear regret in terms of performance. Extensive experiments validate that our algorithm significantly enhances user quality of service and reduces instance utilization costs compared to other state-of-the-art baselines.