A multi-agent deep reinforcement learning approach for optimal resource management in serverless computing

The flexibility of serverless resources enables cloud service providers to scale up dynamically and down the resource requirements over time. However, the varying incoming workload and prefixed capacity of worker nodes can lead to resource under-utilization. To tackle this issue and to improve the performance of data centers, a Multi-Agent deep reinforcement learning-based Resource Management model (MARM) is presented, that reduces the overhead of resource under-utilization. It utilizes the resources efficiently by choosing the best worker node for the incoming workload resource requirements. It allows multiple agents to operate cooperatively while sharing communication. The proposed model keeps track of minimal resources requested by incoming function instances and schedules it to the most suited worker node to encourage efficient resource utilization. Extensive experiments are conducted using the proposed model in a simulated environment using a synthetic dataset and scaled-down Alibaba Cluster Trace. The experimental results show the efficacy of the proposed model in terms of improved CPU and memory utilization up to 39 and 30% over state-of-the-art schedulers, respectively.