Simulation models of different fidelity levels are often available for a complex system. High-fidelity simulations are accurate but time-consuming. Therefore, they can only be applied to a small number of solutions. Low-fidelity simulations are faster and can evaluate a large number of solutions. But their results may contain significant bias and variability. We propose an Multi-fidelity Optimization with Ordinal Transformation and Optimal Sampling (MO2TOS) framework to exploit the benefits of high- and low-fidelity simulations to efficiently identify a (near) optimal solution. MO2TOS uses low-fidelity simulations for all solutions and then assigns a fixed budget of high-fidelity simulations to solutions based on low-fidelity simulation results. We show the benefits of MO2TOS via theoretical analysis and numerical experiments with deterministic simulations and stochastic simulations where noise is negligible with sufficient replications. We compare MO2TOS to Equal Allocation (EA) and Optimal Computing Budget Allocation (OCBA). MO2TOS consistently outperforms both EA and OCBA.