Describir: Towards reducing communications in sparse matrix kernels