On optimization of finite-difference time-domain (FDTD) computation on heterogeneous and GPU clusters