Monitoring of metrics to identify abnormalities in a large scale distributed computing environment

Abstract

Methods and apparatus, including computer program products, implementing and using techniques for finding application latency degradation causes in an application on a distributed computing system. Variables associated with the application and the computing system are identified, including variables that are candidates for being responsible for latency degradation. A total latency is divided into latency components, that each corresponds to an aspect of the total application latency. Each latency component is divided into study classes, that each includes a subset of the candidate variables. For each study class, combinations are generated for the subset of the variables; a latency distribution is determined for each variable combination; the determined latency distributions for the combinations are compared with corresponding latency benchmark values for the same combinations to determine whether a degradation in latency distributions has occurred for particular combinations among the one or more combinations, and a result is provided to a user.