Stimulus-based Sandbox for Learning Resource Dependencies in Virtualized Distributed Applications

Armen Aghasaryan

Abstract: In this paper, we present an approach for the automated profiling of cloud-based distributed applications. The failure dependencies within or between application nodes can be methodically elucidated by dynamically applying a series of unitary perturbations on the underlying computing resources. Each such perturbation in a node acts as a stimulus that propagates to performance meters of dependent nodes and reveals correlations and causal relations between the respective entities. We have developed an instrumented framework for methodical elucidation of these dependencies which covers an extensive set of failure situations. The prime application of our approach is the behavior learning of a distributed application under various resource insufficiency conditions for Quality of Service management and Root Cause Analysis.