Optimizing Resource Allocation in Cloud-Based Deep Learning Systems Using Multi-Agent Reinforcement Learning

Resource management in cloud-based deep learning system is a difficult task because of fluctuations in deep learning workloads and the high computation requirements.The resource allocation approaches which have been put in place include the greedy allocation and the methods that are under the control of a single central authority, which have the disadvantages of not being swift to adapt to the changes in demand patterns, are costly and ineffective.This paper presents MARL based approach for proactive resource management in cloud based deep learning as a problem solution.MARL covers several agents, where each of them oversees a particular cloud resource, such as dispatch, CPU, GPU or memory.These real-life agents work collaboratively and further learn from the environment to help achieve optimal utilization, energy, and cost optimization.The results of the tests show that the proposed MARL-based approach outperforms conventional strategies, in terms of higher throughput, highest utilization of resources, and low energy consumption.This also results in great savings more than if allocation is done in a greedy manner and controlled by a central authority.Thus, MARL can be advantageous for cloud resource management as a feasible and robust approach to deep learning.These results can be beneficial to guide further research and to propose the future developments and implementations of MARL in cloud computing systems.