FlowState: Scaling Serverless Execution via Fluid Worker Repurposing
Serverless computing, Edge computing, Smart Cities, Dynamic workers
The increasing demand for scalable, cost-efficient, and event-driven computing has
fueled the rise of serverless computing as a transformative paradigm within cloud ecosys-
tems. By abstracting infrastructure management and dynamically allocating resources,
Function-as-a-Service (FaaS) platforms enable high elasticity. However, applying this
model to resource-constrained Smart City edge infrastructures exposes severe structural
inefficiencies. Modern urban workloads exhibit deep temporal variability and non-overlapping,
cyclical demand patterns across disparate domains (e.g., daytime intelligent transporta-
tion vs. nighttime municipal waste collection). Traditional serverless architectures isolate
functions within rigid, dedicated container boundaries, forcing an inefficient binary choice
between high startup latency (cold starts) and aggressive resource waste (provisioned
concurrency). This architectural coupling leads to severe resource fragmentation, where
dedicated infrastructure sits idle in one domain while adjacent services experience arti-
ficial resource starvation and contention. To resolve these limitations, this dissertation
introduces FlowState, a novel serverless execution model that decouples application logic
from dedicated container lifecycles to enable rapid runtime worker repurposing. Instead
of maintaining function-specific execution sandboxes, FlowState employs a pool of per-
manently warm, generic workers capable of hot-swapping functionality at the application
tier. We realize this decoupled framework via a functional middleware prototype that
shifts the isolation boundary to a managed runtime layer. This approach encapsulates
function logic into lightweight, dynamically interchangeable application components and
executes deterministic hot-swapping policies to eliminate virtualization-level initialization
costs entirely. To validate the framework, we execute a rigorous empirical evaluation ex-
posing the proposed architecture to synthetic, multi-domain Smart City workload profiles
that emulate cyclical application shifts. Experimental results demonstrate that the de-
coupled approach maximizes hardware efficiency, minimizes function transition delays,
and drastically reduces infrastructure memory footprints under dynamic stress. Com-
pared with container-based baselines, the proposed model reduces average response times
for high-intensity workloads by more than 40% while maintaining a lightweight opera-
tional footprint. These results demonstrate the potential of dynamic worker repurposing
as a highly efficient, secure, and adaptive execution model for distributed cloud-to-edge
infrastructures.