Modeling operating system crash behavior through multifractal analysis, long range dependence and mining of memory usage patterns.

摘要

Software Aging is a phenomenon where the state of the operating systems degrades over a period of time due to transient errors. These transient errors can result in resource exhaustion and operating system hangups or crashes.; Three different techniques from fractal geometry are studied using the same datasets for operating system crash modeling and prediction. Holder Exponent is an indicator of how chaotic a signal is. M5 Prime is a nominal classification algorithm that allows prediction of a numerical quantity such as time to crash based on current and previous data. Hurst exponent measures the self similarity and long range dependence or memory of a process or data set and has been used to predict river flows and network usage.; For each of these techniques, a thorough investigation was conducted using crash, hangup and nominal operating system monitoring data. All three approaches demonstrated a promising ability to identify software aging and predict upcoming operating system crashes. This thesis describes the experiments, reports the best candidate techniques and identifies the topics for further investigation.