Predicting Enterprise Application Performance Measures through Time-series Forecasting Daniel Elsner, 21st August 2017, Scientific advisor: Pouya Aleatrati

Agenda Motivation and Approach Research Artifact Research Questions Data Architecture Project Plan © sebis

“Companies are sitting on a treasure trove Motivation Problem Domains in Application Performance Monitoring (APM) Performance Availability Maintainability “Companies are sitting on a treasure trove – if only they knew how to use it.” A. Samuel, Wall Street Journal, 2015 Evolution in Enterprise Architecture (EA) Growing complexity and high business relevance of Enterprise Architecture (EA) Detect root causes and reduce complexity of distributed, large-scale systems Detect root causes, reduce complexity and lead to a higher agility in EA management Harness potential of monitoring data © sebis

Initial Research Questions What consecutive patterns can be identified by analyzing performance flaws in APM data? 1 How can APM data be used to forecast availability insufficiencies in advance of occurrence? 2 What proactive actions can be derived to avoid the identified patterns and increase performance and availability? 3 © sebis

Approach Data ETL Data Interpretation Research Artifact ML APM Techniques APM Data Gain insights and create value from Application Performance Monitoring (APM) data by applying machine learning techniques. © sebis

10+ potential use cases identified Approach Literature Review APM Performance Analysis EA Software Engineering Machine Learning Reviewed work from 10+ potential use cases identified Use Case Conceptualization 11 Use Cases Data Sources 3 Problem Domains Discussing with Researchers Use Case Evaluation Evaluate use cases by Interviewing Experts  Define Use Case © sebis

Time-series Forecasting Research Artifact Time-series Forecasting e.g. Response Time [ms] e.g. Crash Occurence From historical sequential APM data create forecasts for performance measures and incident probabilities. Linear/Non-linear Regression Gaussian mixture model / Gaussian Processes (Recurrent) Neuronale Netze Hidden Markov Autoregressive integrated moving average Kalman/Partikelfilter 3 Optional Automatic ticket severity evaluation driven by ML 1 Forecasting of relevant performance measures 2 Incident prediction © sebis

Research Questions 1 2 3 How accurately can a time-series forecasting model predict APM measures? 1 How well can a time-series forecasting model predict availability lacks (i.e. application crashes) in enterprise applications/services? 2 To what extend can we evaluate automatic ticket severity classification by analyzing APM data? 3 © sebis

Data Architecture – Layers and Data Sources Application Client Web Server Web Server Middleware Ticketing (Incident) Data Application Server Application Server Application Server Application Server Database Layer Database Database © sebis

Data Architecture – Dimensions and Measures User Device Application Activity Location Time Application App. Server Webserver Database Event Time Method Endpoint Measures Measures Crash rate Response Time Hang Time Crash rate Response Time Request Load Resource Util. Network I/O Resource Util. Network I/O Method Count © sebis

Design + Implementation Project Plan Initial Meeting Prof. Matthes Final Presentation Kickoff Conceptualization Data Exploration Use Case Definition Initial Research Research Artifact Pipeline (ETL) Model Design + Implementation Thesis Evaluation Writing Thesis August September October November December January © sebis

Cand. M. Sc. Daniel Elsner 17135 Daniel.elsner@tum.de

Backup <Date> Short Title © sebis