Seer: a lightweight online failure prediction approach

Özçelik, Burcu and Yılmaz, Cemal (2015) Seer: a lightweight online failure prediction approach. (Accepted/In Press)

WarningThere is a more recent version of this item available.

PDF (This is a RoMEO green journal -- author can archive pre-print (ie pre-refereeing) and post-print (ie final draft post-refereeing)) - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader


Online failure prediction aims to predict the manifestation of failures at runtime before the failures actually occur. Existing online failure prediction approaches typically operate on data which is either directly reported by the system under test or directly observable from outside system executions. These approaches generally refrain themselves from collecting internal execution data that can further improve the prediction quality. One reason behind this general trend is due to the runtime overhead cost incurred by the measurement instruments that are required to collect the data. In this work we conjecture that large cost reductions in collecting internal execution data for online failure prediction can derive from reducing the cost of the measurement instruments, while still supporting acceptable levels of prediction quality. To evaluate this conjecture, we present a lightweight online failure prediction approach, called Seer. Seer uses fast hardware performance counters to perform most of the data collection work. The data is augmented with further data collected by a minimal amount of software instrumentation that is added to the systems software. We refer to the data collected in this manner as hybrid spectra. We applied the proposed approach to three widely used open source subject applications and evaluated it by comparing and contrasting three types of hybrid spectra and two types of traditional software spectra. At the lowest level of runtime overheads attained in the experiments, the hybrid spectra predicted the failures about half way through the executions with an F-measure of 0.77 and a runtime overhead of 1.98%, on average. Comparing hybrid spectra to software spectra, we observed that, for comparable runtime overhead levels, the hybrid spectra provided significantly better prediction accuracies and earlier warnings for failures than the software spectra. Alternatively, for comparable accuracy levels, the hybrid spectra incurred significantly less runtime overheads and provided earlier warnings.

Item Type:Article
Uncontrolled Keywords:Software reliability; software quality assurance; online failure prediction
Subjects:Q Science > Q Science (General)
ID Code:27106
Deposited By:Cemal Yılmaz
Deposited On:27 Aug 2015 11:02
Last Modified:12 Nov 2016 14:23

Available Versions of this Item

Repository Staff Only: item control page