Adaptive Edge-Cloud Mobile Apps Orchestration Framework

Thesis Abstract

This study investigates the design, implementation, and evaluation of an adaptive edge-cloud orchestration framework for mobile applications to optimize performance, cost, and user experience across heterogeneous network environments. The core problem addressed is the gap between static offloading strategies and the dynamic runtime conditions of edge and cloud resources, user mobility, and varying application workloads, which often lead to suboptimal latency, bandwidth usage, and energy consumption on mobile devices. The aim is to develop a self-tuning framework that orchestrates computation, storage, and networking across edge and cloud tiers while obeying quality-of-service (QoS) constraints and security requirements. The research objectives are (1) to formulate a multi-objective optimization model that simultaneously minimizes end-to-end latency, monetary cost, and energy consumption; (2) to design an adaptive decision engine leveraging reinforcement learning to select optimal offloading targets and data placement in real time; (3) to implement a modular prototype comprising an edge cluster, a regional cloud, and a mobile middleware, with support for microservice-based application components; (4) to evaluate the framework under realistic workloads and network conditions, and (5) to assess its resilience to faults and mobility using fault-injection experiments. The theoretical lens combines the Cognitive Load Theory to justify the offloading decisions from a user-perceived performance perspective, and the Resource-Based View to frame the capabilities required for sustainable edge-cloud orchestration. The framework draws on relevant theories of autonomic computing and queueing networks to guide the design of self- managing components and performance modeling. A mixed-methods methodology is employed. The quantitative component uses a 120-user field trial and a 40-node testbed, with synthetic workloads generated to emulate video streaming, augmented reality, and real-time analytics. Data collection instruments include system telemetry logs, network measurements, and application-level QoS metrics captured at 1-second granularity, complemented by a controlled lab experiment with Traffic Control (tc) to simulate diverse network conditions. The analysis employs multi-objective optimization with Pareto-front analysis, regression-based profiling of offloading impact, and time-series forecasting to anticipate resource availability. The reinforcement learning module is evaluated using Q-learning and Deep Q-Networks (DQN) with state representations incorporating device battery, network metrics, workload type, and current edge-cloud loads. A qualitative component uses semi-structured interviews with developers and IT operators (n=12) and thematic analysis to capture operational usability and perceived security concerns. Validity and reliability are addressed via triangulation of telemetry data, cross-validation of predictive models, and bootstrapped confidence intervals. Expected findings indicate that the adaptive framework reduces average end-to-end latency by 28–45% and energy consumption on mobile devices by 15–30%, while achieving cost savings of 12–25% compared to baseline static offloading schemes. The RL-based decision engine is anticipated to converge to near-Pareto-optimal policies within 200 episodes of live deployment, with robustness to sudden network degradation and edge node failures. The study is likely to reveal critical trade-offs between data locality, privacy, and latency, and to demonstrate that modular orchestration enhances adaptability across diverse application profiles. The contribution to knowledge lies in (i) a novel multi-objective optimization model for adaptive edge-cloud orchestration with explicit QoS and security constraints, (ii) an integrated design and implementation of a generic, reusable middleware and orchestration blueprint applicable to mobile applications, and (iii) empirical benchmarks and best-practice guidelines for deploying adaptive edge-cloud frameworks in practice. The main conclusion is that adaptive, learning-driven edge-cloud orchestration can substantially improve performance and resource efficiency for mobile applications under dynamic conditions, while maintaining acceptable security and privacy standards. Recommendations include extending the framework to support federated learning for privacy-preserving policy updates, incorporating programmable network interfaces (e.g., P4) to further reduce tail latency, and exploring energy-aware hardware accelerators at the edge to broaden application domains. Further research is suggested to investigate long-term maintenance of learned policies, transfer learning across different deployment environments, and socio-technical impacts on developers and end-users.