In this dissertation, we propose a metaheuristic Genetic Algorithm (GA) adopted in a distributed and parallel implementation, providing a fast resource allocation approach while maintaining a good performance efficiency. In the first place, we study the role of virtual link mapping stage in online virtual network embedding (VNE) problem where virtual node mapping (VNoM) and virtual link mapping (VLiM) are conducted separately. We propose several efficient GA-based algorithms driven by nonlinear fitness functions for VLiM stage. Extensive simulation results reveal that our proposed algorithms are not only significantly faster than many existing VNE algorithms, but also better than those in performance. Although uncoordinated VNoM and VLiM stages can simplified an algorithmic implementation, a lack of coordination between them might result in low embedding outcomes. To fill this gap, in the second part, we propose a new approach that jointly coordinates virtual node and link mappings in a single stage, where the virtual link mapping is based on several efficient path searching methods. Moreover, a novel heuristic conciliation mechanism is presented to handle a possible set of infeasible link mappings. Our proposed algorithms outperform state-of-the-art VNE approaches in all performance metrics we adopted. In the last part of this dissertation, we propose a collaborative edge computing framework empowered by parked vehicles (PVs) to efficiently handle online computational tasks during peak hours. We formulate the online task offloading as a binary integer programming (BIP) problem aimed at minimizing offloading cost while maximizing accumulative rewards of PVs by sharing their idle resources. To solve the problems of time complexity and scalability of BIP, a heuristic algorithm, namely M&M, and a distributed and parallel GA-based algorithm are introduced. They are then compared with several heuristic algorithms to demonstrate their efficiency on different sizes of generic parking lots.