We study the generation of random fields applied to problems where the domain is much larger than the characteristic distance of the properties fluctuation. Spectral Representation is used and a variation for isotropic fields is proposed. To improve scalability, a method to divide the domain in independent overlapping subsets is developed. The effectiveness of these modifications will be shown on tests in a 3D seismic waves propagation code running over hundreds of cores. Comparison with other schemes for large scale generation of random fields will also be performed.