For large simulation data, Parallel Marching Cubes algorithm is efficient and commonly used to extract isosurfaces in 3D scalar field. However, the isosurface meshes are sometimes too dense and it is difficult for scientists to specify the areas they are interested in. In this paper, we provide them a new way to define mesh importance for decimation using transfer functions and visualize large simulation data in case the normal visualization methods cannot handle due to memory limit. We also introduce a parallel isosurface simplification framework which uses pyramid peeling to extract the decimated meshes progressively without generating the original surface. Since the implementation uses OpenCL which is oriented to heterogeneous computing, our method can be applied to different parallel systems and scientists can see the visualization results while doing simulations. Finally, we evaluate the performances of our algorithm and use different scientific datasets to show the efficiency of our method.