Parallel and fault-tolerant k-means clustering based on the actor model

K-means algorithm is a well-known unsupervised machine learning tool that aims at splitting a given dataset into a fixed number of clusters via iterative refinement approach. Running such an algorithm on today's datasets that are characterized by its high multidimensionality and huge size requires using fault-tolerance mechanisms to mitigate the impact of possible failures. In this paper, we propose an actor-based implementation of k-means algorithm. The algorithm was made fault-tolerant by periodically saving the centroids into a stable storage during the failure-free execution, and restarting from the last saved centroids upon a failure. This was implemented in two different ways: optimistic checkpointing (blocking) and pessimistic checkpointing (non-blocking). The actor-based k-means algorithm was evaluated on a machine with eight cores. The experiments showed that the proposed algorithm scales very well as the number of workers increases, and can be up to similar to 2x faster than a Java-thread-based implementation of k-means algorithm. The results also showed that the optimistic algorithm outperformed the pessimistic one, specifically, in the presence of competing I/O operations. Several failures were forced to occur during the execution to evaluate the performance of the fault-tolerant implementations. The experiments showed that the average amount of lost work ranged from 3-6%.