I have opened a couple of threads asking about k-means performance problem in Spark. I think I made a little progress.

Previous I use the simplest way of KMeans.train(rdd, k, maxIterations). It uses the "kmeans||" initialization algorithm which supposedly to be a faster version of kmeans++ and give better results in general.

But I observed that if the k is very large, the initialization step takes a long time. From the CPU utilization chart, it looks like only one thread is working. Please see https://stackoverflow.com/questions/29326433/cpu-gap-when-doing-k-means-with-spark.

I read the paper, http://theory.stanford.edu/~sergei/papers/vldb12-kmpar.pdf, and it points out kmeans++ initialization algorithm will suffer if k is large. That's why the paper contributed the kmeans|| algorithm.

If I invoke KMeans.train by using the random initialization algorithm, I do not observe this problem, even with very large k, like k=5000. This makes me suspect that the kmeans|| in Spark is not properly implemented and do not utilize parallel implementation.

I have also tested my code and data set with Spark 1.3.0, and I still observe this problem. I quickly checked the PR regarding the KMeans algorithm change from 1.2.0 to 1.3.0. It seems to be only code improvement and polish, not changing/improving the algorithm.

I originally worked on Windows 64bit environment, and I also tested on Linux 64bit environment. I could provide the code and data set if anyone want to reproduce this problem.

I hope a Spark developer could comment on this problem and help identifying if it is a bug.