Serverless filters and pipes : advanced FaaS invocation and routing mechanism based on in-network computation paradigm

Abstract

Serverless computing is a cloud service model that abstracts away infrastructure management, allowing developers to deploy individual functions and pay only for the resources consumed during execution. While this model offers significant advantages in scalability and simplicity of deployments, Function-as-a-Service (FaaS) architectures come with new challenges. On the one hand, the pay-per-use model makes executing inefficient functions more expensive. And on the other hand, the stateless and event-driven nature of functions require carefully crafted workflows and coordination. In-network computation, a paradigm that offloads computation to network components, presents a promising but largely unexplored solution for optimizing serverless architectures. This thesis investigates the application of in-network computation to FaaS platforms. A novel middleware layer that intercepts network communication between the API gateway and serverless functions is proposed and implemented. This middleware enables the dynamic application of operators that can perform pre-processing, post-processing, and function flow control, such as parallelization and chaining. To demonstrate the viability of this approach, two key use cases were implemented and evaluated: a parallelized audio processing workflow and an optimized distributed all-reduce algorithm. The evaluation shows that the middleware can achieve a significant performance speedup by parallelizing audio processing tasks. Moreover, it can mitigate the issue of double-billing by reducing function execution times for coordination-intensive workloads like the all-reduce algorithm. These results demonstrate that integrating in-network computation into serverless platforms via a flexible middleware provides a powerful mechanism for optimizing performance, reducing costs, and enabling more complex application architectures.