As one of the indicators of similarity, you can create a data set nthat is somewhat intelligently designed, and then run each of your algorithms on all of these data sets. Then you get a n-dimensional runtime vector associated with each algorithm, which can then be dropped to any old distance. I would suggest that something like cosine distance would be a good guess, because if your datasets are of different sizes, you would like to classify your algorithms as they scale. In addition to runtime, you can keep track of maximum memory usage or anything you can think of for measurement.