Cassandra token range not distributing evenly across all the nodes

Our Cassandra cluster has 21 nodes and each node is set **num_token=16**. From my understanding, the token range is between -263 to +263-1, each partition key is hashed (MurMur3 hash function) to a token within this range. We thought the token range for each vnode would be **(-263 to +263-1)/num** of vnode (in this case 21x16, totally 336 vnodes). so the sum of token range for each node would be 4.7%. When we check on nodetool ring and calculated the sum as follow: As shown above the percentage varies from 3 to 6%. Is there a way to make the token range distributed more evenly across all the nodes? We are seeing uneven disk usage in this cluster likely cause by this. Any suggestion would help.