ETD

• edge computing
• function-as-a-service
• workload prediction

Accurate workload prediction is a key enabler for efficient and reliable resource alloca-
tion in Function-as-a-Service (FaaS) platforms deployed at the edge. In this thesis, the
effectiveness of foundation model–based approaches for workload forecasting in edge FaaS
environments is assessed. Specifically, two state-of-the-art time-series foundation mod-
els are evaluated: TimesFM, a large-scale pretrained model designed for general-purpose
time-series forecasting, and Lag-Llama, a transformer-based model that leverages lagged
representations to capture long-term temporal dependencies. The workload is defined
as the number of function invocations arriving at an edge FaaS platform over time, a
critical input for proactive resource management. The ability of the selected models to
accurately predict workload dynamics is assessed under different operating conditions by
also analyzing how their predictions can be exploited by resource allocation policies.