I know it's not the more rigorous way to make speedtest and some specifics tools exist in order to achieve it (never try them, i should !) but results difference is important enough to let thing that there are some underlying reasons to prefer functional approach compare to imperativ one. I don't know which ones, probably some funny compilator optimizations but i'm sure it is way to explore. I will be happy to do more testing with better tools if i have the occasion.
You can run it once inside sbt with run numberIterations s1,s2,s3 where numberIterations, s1, s2, s3 are Int representing number of run of the test and s1, s2, s3 are vector sizes.
I took two version of sqdist, imparativ and functional ones. For input vector size i run 60% of warming up sqdist and measure time on the left 40% for different number of iterations from 100k to 10M. At each iterations 2 random DenseVector are generated. Code spoke better by himself i think.
Things like computation of the mean and standard deviations should be done !
It can be found on theses StackOverFlow, this one which highlight a question of JVM version, or this other one which propose to prevent tailrecursion optimization which can be usefull to see at which degree theses optimizations are important.