Online failure prediction in cloud datacenters by real-time message pattern learning

摘要: Once failures occur in a cloud datacenter accommodating large number of virtual resources, they tend to spread rapidly and widely, impacting on many users (tenant owners). One the best ways prevent failure from spreading system is identifying signs before its occurrence deal with it proactively causes serious problems. Although several approaches have been proposed predict by analyzing past message logs relationship between messages failures, still difficult automatically for reasons such as various types log formats or time gaps pattern learning application identified patterns real systems. Based this understanding, we propose new prediction method paper which learns classifying their similarity without depending format re-Iearning frequently-changed configurations. We implemented our evaluated using data recorded an actual datacenter. The experimental result shows that approach predicted 80% precision covered 90% occurrences.