ENERGY AND LATENCY AWARE RESOURCE MANAGEMENT IN EDGE FAAS PLATFORMS

Edge computing is reshaping the way computation is performed by moving it closer to the data source, thereby enabling faster response times and reducing the dependency on centralized data centers. This architectural shift is increasingly relevant for applications that demand low-latency processing and high responsiveness, such as autonomous systems, smart manufacturing, and real-time monitoring. The primary goal of edge computing is to support these use cases through localized computation, which minimizes delays and enables decision-making at the point of data generation. However, deploying applications in edge environments introduces a new set of challenges, particularly w.r.t. resource management. Unlike traditional cloud data centers, edge nodes typically operate under constrained resources, with limited processing power, memory capacity, and energy availability. This makes efficient utilization of these resources essential to sustaining performance while maintaining system viability. Furthermore, edge workloads are inherently dynamic and often unpredictable, leading to scenarios where over-provisioning wastes energy and under-provisioning degrades performance. Managing this trade-off requires a system that can rapidly adapt to changing workloads while minimizing unnecessary energy consumption. Function-as-a-Service (FaaS) platforms offer a promising solution to these challenges. As a model of serverless computing, FaaS allows developers to deploy fine-grained computational units, or functions, that are triggered by events and automatically scaled according to demand. This abstraction decouples the logic of functions from resource provisioning, which is especially appealing in dynamic and distributed environments, such as the edge. By invoking functions only when they are needed, FaaS enables more granular control over resource usage and offers the potential for higher energy efficiency compared to monolithic or continuously running services. Despite these advantages, current FaaS implementations are primarily designed for cloud environments and are not well-suited to the constraints of edge computing. The problem becomes more pronounced when considering energy efficiency. Suboptimal function placement and static resource allocation policies can lead to significant energy waste when applied to edge scenarios. Moreover, centralized scheduling mechanisms often fail to account for the heterogeneity and distribution of edge nodes, resulting in the inefficient utilization of available resources. This thesis investigates how FaaS platforms can be adapted for energy-efficient and low-latency operations in edge environments. It focuses on designing intelligent resource management strategies that dynamically allocate computational resources based on real-time workload demands while minimizing the energy footprint of the system. At the same time, it addresses one of the most critical performance challenges of FaaS, i.e., cold start delays and long completion times, by exploring techniques that proactively mitigate these latencies through predictive scheduling and efficient container reuse. The goal is to leverage the elasticity of FaaS in a way that aligns with both the energy constraints and stringent latency requirements of edge computing. Building upon these motivations, this thesis sets out to design and evaluate resource management techniques that simultaneously reduce energy consumption and improve service responsiveness. Therefore, this work explores how the core principles of FaaS elasticity, event-driven execution, and fine-grained scaling can be adapted for energy-efficient and low-latency edge computing scenarios. The first contribution introduces an energy-aware function scheduling mechanism tailored for edge-based FaaS environments. By leveraging adaptive workload placement strategies, execution is consolidated onto fewer active nodes during low-demand periods, while inactive nodes are temporarily shut down to reduce energy waste. As demand increases, additional nodes are reactivated to maintain service responsiveness. Evaluation in the context of a video stream processing scenario demonstrates the capability of this approach to significantly lower energy consumption without compromising the timeliness or reliability of function execution using a formulation of a MixedInteger Programming (MIP) approach. The second contribution explores the integration of time series forecasting into energy-aware function scheduling in edge-based FaaS environments. Various neural prediction models are assessed, including those based on recurrent structures, Gaussian processes, and transformer architectures, to estimate future function invocation patterns. These predictions are used to guide a scheduling algorithm that minimizes energy use by proactively adjusting the number of active edge nodes. The MIP approach governs resource allocation decisions, balancing function demand with minimal node usage. The method remains robust even under fluctuating prediction accuracies, highlighting its adaptability in practical and dynamic workloads. The third contribution addresses a key latency challenge in Edge FaaS environments, i.e., cold start delays and long function completion times. To mitigate these issues, a predictive hybrid scheduling mechanism is introduced, combining time series forecasting of future load with a hybrid memory-retention policy. By anticipating function invocations and proactively preparing warm execution containers, the approach reduces both the frequency and impact of cold starts. The hybrid policy considers both the recency of function usage and container load to make informed retention and eviction decisions, thereby improving overall scheduling efficiency. Evaluation with realistic edge workloads demonstrates significant reductions in cold start latency and completion times while enhancing overall responsiveness. In summary, this thesis presents a comprehensive examination of how FaaS platforms can be enhanced for improved energy efficiency and low latency in edge computing environments. Through a sequence of complementary contributions, it demonstrates that combining intelligent scheduling, predictive modeling, and adaptive resource management can effectively address both energy and performance challenges inherent to Edge FaaS. The proposed mechanisms, ranging from energy-aware scheduling and predictive resource provisioning to hybrid latency mitigation, collectively demonstrate how FaaS elasticity can be leveraged to strike a balance between energy consumption and service responsiveness. Overall, the work advances the state of the art in edge-oriented FaaS systems by offering practical, data-driven strategies that enhance sustainability and performance in distributed, resource-constrained environments.