In this paper, we study the reliable resource allocation (RRA) problem of allocating virtual machines (VMs) from multiple optically interconnected data centers (DCs) with the objective of minimizing the total failure probability based on the information obtained from the optical network virtulization. We first describe the framework of resource allocation, formulate the RRA problem, and prove that RRA is NP-complete. We provide an algorithm, named Minimum Failure Cover (MFC), to obtain optimal solutions for small scale problems. We then provide a greedy algorithm, named VM-over-Reliability (VOR), to solve large scale problems. Numerical results show that VOR achieves results close to optimal solutions gained by MFC for small scale problems. Numerical results also show that VOR outperforms the resource allocation through random DC selection (RDS).