Double-gyre ocean circulation is a typical phenomenon in the northern mid-latitude ocean basins. Its low-frequency variability significantly influences on both ocean and climate. To enhance its predictability, the finding of optimal initial perturbation which can trigger the double-gyre variation is important. CNOP method is adopted to calculate the optimal initial perturbation and this method has already been widely applied. Previous studies show that intelligent algorithms are effective methods to solve CNOP on ENSO event in ZC model. In this paper, it is the first time that intelligent algorithm is used to solve CNOP in ROMS for double-gyre variation. The dimension of the objective function is much higher than that of the ZC model and the calculation time increases dramatically, therefore effective result may not be able to be acquired in acceptable time. We propose an improved intelligent algorithm in accompany with the principal component analysis to solve CNOP, which is called principal component analysis based flower pollination (PCAFP) algorithm. To speed up furtherly, we complete both the parallelization of the nonlinear process of ROMS and the PCAFP algorithm. Different frameworks like OpenMP, MPI and CUDA on GPU are utilized and their performances are compared. Using a 20 nodes' cluster (12 cores in each node) on TH-1A supercomputer system, the computation time is reduced from 324.4 hours (about 13.5 days) to 3.71 hours, and the 87.44x speedup demonstrates that our parallelization can remarkably reduce the time cost. Hence, the proposed PCAFP with its parallelization on clusters can not only obtain optimal initial perturbation that can trigger the double-gyre variation, but also solve CNOP effectively and efficiently.