On the use of computational geometry to detect software faults at runtime

Abstract

Despite advances in software engineering, software faults continue to cause system downtime. Software faults are difficult to detect before the system fails, especially since the first symptom of a fault is often system failure itself.This paper presents a computational geometry technique and a supporting tool to tackle the problem of timely fault detection during the execution of a software application. The approach involves collecting a variety of runtime measurements and building a geometric enclosure, such as a convex hull, which represents the normal (i.e., non-failing) operating space of the application being monitored. When collected runtime measurements are classified as being outside of the enclosure, the application is considered to be in an anomalous (i.e., failing) state. This paper presents experimental results that illustrate the advantages of using a computational geometry approach over the distance based approaches of Chi-Squared and Mahalanobis distance. Additionally, we present results illustrating the advantages of using the convex-hull enclosure for fault detection in favor of a simpler enclosure such as a hyperrectangle

Self-Healing of Operational Workflow Incidents on Distributed Computing Infrastructures

“…Zhang et al [34] Inria assume that incidents lead to non stationarity of the workload statistics and use the Page-Hinkely test to detect them. Stehle et al [32] present a method where the convex hull is used instead of hyper-rectangles to classify system states. As described in section 5, we use multiple threshold values for a given metric to use more than two levels to characterize incidents.…”

Entropy-Based Detection of Incipient Faults in Software Systems

“…The extent of this method is limited to transactionbased anomalous behavior in the sense that it can only detect anomalies that manifest themselves via changes in CPU utilization patterns. Stehle et al present an approach to detect software faults using computational geometry [10]. A convex hull is established as the geometric enclosure of an n-dimensional data set gathered over the full spectrum of nominal system operating conditions.…”

Diagnosis of software failures using computational geometry

“…The detection approach uses information collected by the runtime sensors to construct a geometric enclosure whose enclosing space represents the normal execution of the monitored application. [3] V. EXPERIMENTAL RESULTS This section contains the results of the case study on the diagnosis of software failures. The failure diagnosis results include the results of experiments designed to test Aniketos's ability to diagnose known fault types.…”

“…Our previous work [3] proposed a technique for detecting software failures using computational geometry. This paper proposes an technique for diagnosing software failures that extends the geometric approach from our previous work.…”