A fault tolerant K-means algorithm based on storage-class memory

Checkpointing is a pervasive method to provide fault-tolerance in high performance computing systems. However, with the rapid growth of system scale, more frequent checkpointing are needed, thus making the overhead brought by checkpointing intolerable. In this paper we present a checkpoint-free fault tolerance method. It takes advantage of non-volatile property of storage-class memory (SCM) to store execution-relevant data. Our method introduces negligible overhead to the algorithm when failures strikes and substantially reduces the recovery overhead. We add fault-tolerant capacity to K-means clustering algorithm using our method. Experimental results indicate that our approach introduces much less overhead than checkpointing does.