FLINT: Performance-Aware In-Network Computing for Dynamic 6G Sub-Networks

To address the demand for near real-time applications in 6G networks, we propose FLINT, a context-aware task scheduling system for in-network computing. Unlike hardwaredependent approaches, FLINT utilizes dynamic compute nodes connected to routers, enabling flexible hardware utilization. By employing static code embeddings from a large transformer model and small metric estimation models on each compute blade, FLINT efficiently assigns unseen tasks based on their binaries and arguments. This cooperative scheduling approach between blades and the central router allows for dynamic deployments and blade exchangeability. Evaluation of diverse WebAssembly tasks and hardware demonstrates FLINT's effectiveness, achieving an F1 score and accuracy of over 0.9 on our taskset while maintaining low response time overhead.