Profile-based, load-independent anomaly detection and analysis in performance regression testing of software systems

Abstract

Performance evaluation through regression testing is an important step in the software production process. It aims to make sure that the performance of new releases do not regress under a field-like load. The main outputs of regression tests are the metrics that represent the response time of various transactions as well as the resource utilization (CPU, disk I/O and Network). In this paper, we propose to use a concept known as Transaction Profile, which can provide a detailed representation for the transaction in a load independent manner, to detect anomalies through performance test runs. The approach uses data readily available in performance regression tests and a queueing network model of the system under test to infer the Transactions Profiles. Our initial results show that the Transactions Profiles calculated from load regression test data uncover the performance impact of any update to the software. Therefore we conclude that using Transactions Profiles is an effective approach to allow testing teams to easily assure each new software release does not suffer performance regression.