On Learning Parametric Dependencies from Monitoring Data Johannes - PowerPoint PPT Presentation

On Learning Parametric Dependencies from Monitoring Data Johannes Grohmann, Simon Eismann, Samuel Kounev Symposium on Software Performance (SSP) 2019 05.11.2019 https://se.informatik.uni-wuerzburg.de/

Software Performance Models Introduction Related Work Approach Evaluation Conclusion  Performance models are a common approach to predict software performance Server system 2 Johannes Grohmann – On Learning Parametric Dependencies

Software Performance Models Introduction Related Work Approach Evaluation Conclusion  Performance models are a common approach to predict software performance maximize Server system Efficiency 3 Johannes Grohmann – On Learning Parametric Dependencies

Software Performance Models Introduction Related Work Approach Evaluation Conclusion  Performance models are a common approach to predict software performance maximize Server system Efficiency e.g. Auto-scaler Performance model 4 Johannes Grohmann – On Learning Parametric Dependencies

Software Performance Models Introduction Related Work Approach Evaluation Conclusion  Performance models are a common approach to predict software performance  However, correctly modeling a software system is difficult maximize Server system Efficiency e.g. Auto-scaler Performance model 5 Johannes Grohmann – On Learning Parametric Dependencies

Parametric Dependencies Introduction Related Work Approach Evaluation Conclusion Image Provider Database Recommender UI 6 Johannes Grohmann – On Learning Parametric Dependencies

Parametric Dependencies Introduction Related Work Approach Evaluation Conclusion Image Provider Database Recommender Workload QoS-metrics UI 20 Users (Quality of Service) 7 Johannes Grohmann – On Learning Parametric Dependencies

Parametric Dependencies Introduction Related Work Approach Evaluation Conclusion Image Provider Database Recommender ? Workload Workload QoS-metrics UI 20 Users 50 Users (Quality of Service) 8 Johannes Grohmann – On Learning Parametric Dependencies

Parametric Dependencies Introduction Related Work Approach Evaluation Conclusion  One parameter of performance models are parametric dependencies Image Provider Database Recommender ? Workload Workload QoS-metrics UI 20 Users 50 Users (Quality of Service) ResourceDemand(Recommender) [milliseconds] 9 Johannes Grohmann – On Learning Parametric Dependencies

Parametric Dependencies Introduction Related Work Approach Evaluation Conclusion  One parameter of performance models are parametric dependencies Image Provider Database Recommender ? Workload Workload QoS-metrics UI 20 Users 50 Users (Quality of Service) ResourceDemand(Recommender) = 17 * currentItems.size() [milliseconds] 10 Johannes Grohmann – On Learning Parametric Dependencies

Parametric Dependencies Introduction Related Work Approach Evaluation Conclusion  One parameter of performance models are parametric dependencies Image Provider Database Recommender ? Workload Workload QoS-metrics UI 20 Users 50 Users (Quality of Service) ResourceDemand(Recommender) = 17 * currentItems.size() + 0 * user ID [milliseconds] 11 Johannes Grohmann – On Learning Parametric Dependencies

Parametric Dependencies Introduction Related Work Approach Evaluation Conclusion  One parameter of performance models are parametric dependencies Image Provider Database Recommender ? Workload Workload QoS-metrics UI 20 Users 50 Users (Quality of Service) Goal: Autonomically detect such parametric dependencies 12 Johannes Grohmann – On Learning Parametric Dependencies

Example Introduction Related Work Approach Evaluation Conclusion 13 Johannes Grohmann – On Learning Parametric Dependencies

Related Work Introduction Related Work Approach Evaluation Conclusion  Krogmann et al. [KKR10] or Mazkatli and Koziolek [MK18] require source code for detection of dependencies. In contrast, our approch is solely based on monitoring data. 15 Johannes Grohmann – On Learning Parametric Dependencies

Related Work Introduction Related Work Approach Evaluation Conclusion  Krogmann et al. [KKR10] or Mazkatli and Koziolek [MK18] require source code for detection of dependencies. In contrast, our approch is solely based on monitoring data. Monitoring data 16 Johannes Grohmann – On Learning Parametric Dependencies

Related Work Introduction Related Work Approach Evaluation Conclusion  Krogmann et al. [KKR10] or Mazkatli and Koziolek [MK18] require source code for detection of dependencies. In contrast, our approch is solely based on monitoring data. Parameterized Parameterization Monitoring Performance Model Extraction Performance [SC+15, BHK11, [BHK11, WS+17, data Model SG+19, RV95, KP+09] HW+99, IL+05, MF11] Model 17 Johannes Grohmann – On Learning Parametric Dependencies

Related Work Introduction Related Work Approach Evaluation Conclusion  Krogmann et al. [KKR10] or Mazkatli and Koziolek [MK18] require source code for detection of dependencies. In contrast, our approch is solely based on monitoring data. Parameterized Parameterization Monitoring Performance Model Extraction Performance [SC+15, BHK11, [BHK11, WS+17, data Model SG+19, RV95, KP+09] HW+99, IL+05, MF11] Model Identified Dependencies 18 Johannes Grohmann – On Learning Parametric Dependencies

Related Work Introduction Related Work Approach Evaluation Conclusion  Krogmann et al. [KKR10] or Mazkatli and Koziolek [MK18] require source code for detection of dependencies. In contrast, our approch is solely based on monitoring data. Parameterized Parameterization Monitoring Performance Model Extraction Performance [SC+15, BHK11, [BHK11, WS+17, data Model SG+19, RV95, KP+09] HW+99, IL+05, MF11] Model Dependency Parameterized Identified Characterization Dependencies Dependencies [BH11, CW00, AG+18] 19 Johannes Grohmann – On Learning Parametric Dependencies

In a nutshell Introduction Related Work Approach Evaluation Conclusion 20 Johannes Grohmann – On Learning Parametric Dependencies

In a nutshell Introduction Related Work Approach Evaluation Conclusion Manual identification of parametric dependencies is Problem not always possible, time-intensive and error-prone 21 Johannes Grohmann – On Learning Parametric Dependencies

In a nutshell Introduction Related Work Approach Evaluation Conclusion Manual identification of parametric dependencies is Problem not always possible, time-intensive and error-prone Learning of dependencies using standard monitoring Idea data collected during production 22 Johannes Grohmann – On Learning Parametric Dependencies

In a nutshell Introduction Related Work Approach Evaluation Conclusion Manual identification of parametric dependencies is Problem not always possible, time-intensive and error-prone Learning of dependencies using standard monitoring Idea data collected during production Increase model accuracy and expressiveness, Benefit additional step towards autonomic model learning 23 Johannes Grohmann – On Learning Parametric Dependencies

In a nutshell Introduction Related Work Approach Evaluation Conclusion Manual identification of parametric dependencies is Problem not always possible, time-intensive and error-prone Learning of dependencies using standard monitoring Idea data collected during production Increase model accuracy and expressiveness, Benefit additional step towards autonomic model learning Use feature selection techniques for detecting, Action regression for characterizing the dependencies 24 Johannes Grohmann – On Learning Parametric Dependencies

APPROACH 25 Johannes Grohmann – On Learning Parametric Dependencies

Required monitoring information Introduction Related Work Approach Evaluation Conclusion  Monitoring data per invocation through Kieker [vHWH12] monitoring • Parameter values and types • Return value and type Identification parameters • Resource demand 26 Johannes Grohmann – On Learning Parametric Dependencies

Required monitoring information Introduction Related Work Approach Evaluation Conclusion  Monitoring data per invocation through Kieker [vHWH12] monitoring • Parameter values and types • Return value and type Identification parameters • Resource demand • Method signature Parameter-related information • Entity 27 Johannes Grohmann – On Learning Parametric Dependencies

Required monitoring information Introduction Related Work Approach Evaluation Conclusion  Monitoring data per invocation through Kieker [vHWH12] monitoring • Parameter values and types • Return value and type Identification parameters • Resource demand • Method signature Parameter-related information • Entity • Trace id • Execution order index ( EOI ) Trace reconstruction • Execution stack size ( ESS ) 28 Johannes Grohmann – On Learning Parametric Dependencies