A DRL-based online VM scheduler for cost optimization in cloud brokers.

Area of Science:

• Computer Science
• Cloud Computing
• Artificial Intelligence

Background:

• Cloud bursting relies on Infrastructure as a Service (IaaS) virtual machines (VMs), introducing uncertainty in scheduling due to unpredictable request arrivals and lifecycles.
• Existing deep reinforcement learning (DRL) approaches for VM scheduling often neglect Quality of Service (QoS) guarantees.

Purpose of the Study:

• To address the cost optimization challenge in online VM scheduling for cloud brokers supporting cloud bursting.
• To develop a DRL-based scheduler that minimizes public cloud expenses while adhering to QoS constraints.

Main Methods:

• Proposing DeepBS, a DRL-based online VM scheduler designed for cloud brokers.
• Implementing adaptive learning strategies to manage non-smooth and uncertain user requests.
• Evaluating DeepBS using cluster traces from Google and Alibaba for request arrival patterns.

Main Results:

• DeepBS demonstrates significant cost optimization advantages over benchmark algorithms.
• The scheduler effectively adapts to uncertain and dynamic VM request environments.
• QoS restrictions are satisfied while minimizing operational costs.

Conclusions:

• DeepBS offers a robust solution for cost-effective VM scheduling in cloud bursting environments.
• The DRL approach provides adaptive and efficient management of uncertain IaaS demands.
• This method enhances the economic viability of cloud bursting strategies.