Graph partitioning is an important subproblem in many applications. To solve it efficiently, the multilevel strategy in combination with a matching algorithm and a local refinement heuristic has proven to be a powerful method, and several libraries exist providing such an implementation. Due to the large involvement of heuristics, the evaluation of these libraries is usually based on experiments. In this paper we show that single experiments are usually not sufficient to judge the quality of an algorithm, since even results obtained for graphs of and identical structure show high variations. This is still true, even if the applied algorithms do not contain any nondeterminism. We propose a scheme that considers these variations and therefore makes evaluations and comparisons of different implementations more meaningful. We have applied this technique to evaluate the improvements of the Helpful-Set 2-partitioning implementation and present the obtained results.