Using Hessian Locally Linear Embedding for autonomic failure prediction

The increasing complexity of modern distributed systems makes conventional fault tolerance and recovery prohibitively expensive. One of the promising approaches is online failure prediction. However, the process of feature extraction depends on the experienced administrators and their domain knowledge to filtering and compressing error events into a form that is easy for failure prediction. In this paper, we present a novel performance-centric approach to automate failure prediction with Manifold Learning techniques. More specifically, we focus on methods that use Supervised Hessian Locally Embedding algorithm to achieve autonomic failure prediction. In our experimental work we found that our method can automatically predict more than 60% of the CPU and memory failures, and around 70% of the network failure based on the runtime monitoring of the performance metrics.