Solving large permutation flow-shop scheduling problems on GPU-accelerated supercomputers