Towards a Multi-Armed Bandit Approach for Adaptive Load Balancing in Function-as-a-Service Systems

The Function-as-a-Service (FaaS) computing paradigm enjoys growing popularity, promising easier operations, increased scalability and more flexible pricing. The adoption of FaaS has started exceeding the borders of Cloud data centers, with interest of researchers and practitioners in exploiting serverless functions at the edge of the network as well. However, the constrained resource availability of Edge environments require careful resource allocation. In particular, a typical problem in FaaS system is how to distribute incoming invocation requests to pools of heterogeneous worker nodes, through suitable load balancing policies. In this paper, we argue that different workloads and scenarios might require different load balancing policies. As such, we study an approach to adaptive load balancing based on a Multi-Armed Bandit (MAB) problem formulation. Our approach is able to identify the best performing policy at run-time, according to user-specified performance objectives. Simulated experiments demonstrate that our approach is effective, even in non-stationary settings, where the objective to optimize is modified during execution.