InvarNet-X: A Comprehensive Invariant Based Approach for Performance Diagnosis in Big Data Platform

To provide a high performance and reliable big data platform, this paper proposes a comprehensive invariant-based performance diagnosis approach named InvarNetX. InvarNet-X not only covers performance anomaly detection but also root cause inference, both of which are conducted under the consideration of operation context of big data applications. The performance anomaly detection procedure is adopted to trigger the cause inference procedure and accomplished by checking the ARIMA model drift on Cycle Per Instruction (CPI) data of big data applications. The oracle of cause inference is the unobservable root causes of performance problems always expose themselves via the violations of the associations amongst directly observable performance metrics. In InvarNet-X, such observable associations as the likely invariants are established by the Maximal Information Criteria (MIC) and each performance problem is signified by a set of violations of those likely invariants. Finally,the root cause is uncovered by searching a similar signature in the signature database. With such a comprehensive analysis, InvarNet-X can provide much detailed clues for performance problems and even pinpoint the root causes if the signature database is given. Through experimental evaluations in a small prototype, we find out InvarNet-X can achieve an average 91% precision and 87% recall in diagnosing some real faults reported in software bug repositories, which is superior to several state-ofthe-art approaches. Meanwhile, the local modeling methodology makes InvarNet-X easily facilitated in real-time and large scale big data platforms.