Predictive function placement for distributed serverless environments

Serverless computing has gained popularity in recent years, and it is being used in an increasing number of use cases. It is still limited by some of its challenges, such as high latency, statelessness, vendor lock-in or difficulty to test. However, it presents advantages that could open new opportunities for more applications, such as event-driven IoT or mobile applications that could benefit from the serverless elasticity and resilience and reduce operating costs. Networks are becoming more geographically distributed and adopting a hybrid infrastructure of cloud and edge nodes, for example on 5G base stations. If the latency is high, a wide range of mobile applications cannot make use of serverless computing. In this paper, we propose an approach to reduce the end-to-end latency perceived by an application using serverless computing by predicting the resource utilization of the available nodes to decide the location to deploy the serverless function instances. In this first approach, we apply a Kalman filter to predict the CPU load of each of the nodes. The results of the experiments with two serverless nodes show a latency reduction that increases as the requested computation becomes more complex, reaching a 17% reduction compared to resource-based load balancer with direct measurement.