Serverless computing has gained popularity as a novel cloud execution model for applications in recent times. Businesses constantly try to leverage this new paradigm to add value to their revenue streams. The serverless eco-system accommodates many application domains successfully. However, its inherent properties such as cold start delays and relatively high per unit charges appear as a shortcoming for certain application workloads, when compared to a traditional Virtual Machine (VM) based execution scenario. A few research works exist, that study how serverless computing could be used to mitigate the challenges in a VM based cluster environment, for certain applications. In contrast, this work proposes a generalized framework for determining which workloads are best able to reap benefits of a serverless computing environment. In essence, we present a potential hybrid scheduling solution for exploiting the benefits of both a serverless and a VM based serverful computing environment. Our proposed framework leverages the actor-critic based deep reinforcement learning architecture coupled with the proximal policy optimization technique, in determining the best scheduling decision for workload executions. Extensive experiments conducted demonstrate the effectiveness of such a solution, in terms of user cost and application performance, with improvements of up to 44% and 11% respectively.

Serverless computing enables rapid application development and deployment by composing loosely coupled microservices at a scale. This emerging paradigm greatly unburdens the users of cloud environments, from the need to provision and manage the underlying cloud resources. With this shift in responsibility, the cloud provider faces the challenge of providing acceptable performance to the user without compromising on reliability, while having minimal knowledge of the application requirements. Sub-optimal resource allocations, specifically the CPU resources, could result in the violation of performance requirements of applications. Further, the fine-grained serverless billing model only charges for resource usage in terms of function execution time. At the same time, the provider has to maintain the underlying infrastructure in always-on mode to facilitate asynchronous function calls. Thus, achieving optimum utilization of cloud resources without compromising on application requirements is of high importance to the provider. Most of the current works only focus on minimizing function execution times caused by delays in infrastructure set up and reducing resource costs for the end-user. However, in this paper, we focus on both the provider and user's perspective and propose a function placement policy and a dynamic resource management policy for applications deployed in serverless computing environments. The policies minimize the resource consumption cost for the service provider while meeting the user's application requirement, i.e., deadline. The proposed solutions are sensitive to deadline and efficiently increase the resource utilization for the provider, while dynamically managing resources to improve function response times. We implement and evaluate our approach through simulation using ContainerCloudSim toolkit. The proposed function placement policy when compared with baseline scheduling techniques can reduce resource consumption by up to three times. The dynamic resource allocation policy when evaluated with a fixed resource allocation policy and a proportional CPU-shares policy shows improvements of up to 25% in meeting the required function deadlines.