The aim of this paper is to propose a muscle control method for a humanoid musculoskeletal arm that mimics human muscle coordination from an engineering viewpoint. As muscle control is posed as a redundancy problem, the resulting muscle force corresponds to different desired control criteria (e.g., minimizing the total metabolic energy, minimizing muscle activity). In this research, the criterion we choose is "load balance", which is essential for reducing actuator demands on systems subject to repetitive and demanding tasks. In order to achieve a balanced force distribution throughout the muscles, we obtain a minimum-load muscle force by considering the acceleration contribution in both muscle and joint space. The orthogonal space of the minimum-load muscle force solution is designed to place the muscle force close to the midpoint of the muscle force limits. The proposed method is tested by tracking a table tennis movement of a real human subject. The results not only provide an explanation on the concepts of "minimum-load" and "balance-load", but also show that the proposed method has advantageous properties such as computational efficiency and stability.