You're right, I should be running each test at least a few times and taking the average. I found that the results vary a few percent here and there, but the drastic difference between the different datastructures is always consistent and visible.