Data Center Diagnostics with Network Provenance

Diagnosing problems in data centers has always been a challenging problem due to their complexity and heterogeneity. Among recent proposals for addressing this challenge, one promising approach leverages provenance, which provides the fundamental functionality that is needed for performing fault diagnosis and debugging—a way to track direct and indirect causal relationships between system states and their changes. This information is valuable, since it permits system operators to tie observed symptoms of a faults to their potential root causes. However, capturing provenance in a data center is challenging because, at high data rates, it would impose a substantial cost. In this paper, we introduce techniques that can help with this: We show how to reduce the cost of maintaining provenance by leveraging structural similarities for compression, and by offloading expensive but highly parallel operations to hardware. We also discuss our progress towards transforming provenance into compact actionable diagnostic decisions to repair problems caused by misconfigurations and program bugs.