Normalize row or column while each row is an observation

Suppose I have a matrix compose of row as each observation, column as each property and I want to calculate the distance between each observation.

In this case I think I should normalize each column, instead of row, to a unit vector? I think that would make more sense since each column can have different unit. In contrast if I normalize it by rows to unit vectors, it will be dominated by the property that has larger average value.