Main Content

Accelerating FaaS Platforms by In-Network Computing

Author: Mohammed M. Sillah

Supervisors: Prof. Dr. Boris Koldehofe, M.Sc. Mehran Salmani

Format: Master's thesis, TU Ilmenau

Kubernetes-based serverless platforms rely on software-mediated mechanisms such as DNS resolution and node-local service proxies to perform service discovery and request routing. While these abstractions provide flexibility and correctness under dynamic workloads, they introduce additional latency, increase tail variability, and exhibit slow convergence when backend replicas scale or terminate. These effects are particularly pronounced in serverless environments, where functions are short-lived and backend membership changes frequently. This thesis investigates whether programmable network dataplanes can improve request routing efficiency and scaling responsiveness by offloading service selection into the network fabric. Using P4-programmable switches, service-aware forwarding logic is implemented directly in the dataplane, enabling backend selection and packet steering to be performed at line rate while preserving Kubernetes control-plane semantics. A hybrid architecture is proposed in which Kubernetes remains responsible for orchestration and lifecycle management, while latency-critical routing decisions are executed inside a programmable switch synchronised with Kubernetes endpoint updates. The system is implemented using the BMv2 software switch and evaluated against the default Kubernetes networking stack using OpenFaaS workloads. The evaluation examines end-to-end latency, delivered throughput (goodput), packet loss, load distribution fairness, and routing behaviour under steady-state load, saturation, and dynamic scaling events. Results show that dataplane-assisted routing consistently reduces average and tail latency, sustains higher goodput under load, and reacts deterministically to backend membership changes. However, the overall convergence time following pod deletion remains dominated by transport-layer and host-level effects outside the dataplane’s control. These findings demonstrate that programmable dataplanes can significantly improve per-request efficiency and scaling responsiveness in Kubernetes-based serverless platforms, while also highlighting that complete elimination of stale traffic requires coordinated handling across the control plane, dataplane, and transport layers.

References: https://bibliographie.tu-ilmenau.de/servlets/DozBibEntryServlet?mode=show&id=ilm_mods_00079552