In this paper, the multi-user multiple-input multiple-output (MU-MIMO) relay channel is investigated, where the source node provides multi-beams to multi-users via a multi-antenna relay node. In this scenario, linear processing matrix at the relay node is designed around block diagonal (BD) scheme to improve the system sum-rate. Compared with the traditional linear processing matrix with zero-forcing (ZF) scheme at the relay node, the proposed matrix based on BD scheme can not only eliminate the multi-user interference to the same extent as the ZF scheme, but also realize optimal power allocation at the relay node. Numerical simulations demonstrate the BD scheme outperforms the ZF scheme and can significantly improve the sum-rate performance.