PRISMA: Efficient Algorithms and Methods for Online Extraction of Performance Models in Virtualized Environments
Final Report Abstract
During the course of the PRISMA project, we developed and published efficient algorithms and methods for automatic extraction of architecture-level performance models of virtualization platforms and their hosted applications during system operation. Many complex performance effects and influences in virtualized environments (e.g., mutual influences of the fine-granular system components and layers such as OS, virtualization, middleware, application logic, I/O subsystem, caching and communication protocols) are only observable during system operation when the system is running in the real production environment under real production workloads as opposed to running in a controlled testing environment with artificial workloads or synthetic benchmarks. Therefore, we envisioned a novel class of virtualization platforms and virtual appliances that integrate the developed model extraction algorithms in their architecture. The term virtual appliance (VA) refers to a prepackaged virtual machine image containing a software stack designed to run on a virtualization platform. In pursuit of this goal, we developed a reference architecture for online learning of performance models. While doing so, the following two main research challenges were tackled: • How to extract models based on monitoring data collected at run time with no possibility to conduct static code analysis of the application, no control of the executed application workloads, and only limited flexibility to vary the system configuration during operation? • How to automatically identify and quantify parametric dependencies and cope with a potentially very large search space of possible dependencies? We addressed the first research question by proposing an agent-based architecture, in which the agents were already embedded in VAs in order to minimize the effort for the developer. Furthermore, each agent was equipped with a limited scope, as the visibilities of each entity in a virtualized environment are quite limited. On instantiation of a VA, the contained agent starts to monitor the application serving real production workloads and it automatically builds a submodel describing the observed performance behavior of the application and platform layers inside the VA. The agent continuously updates the model skeleton to reflect dynamic changes, for instance, in the configuration or in the workload of the application. The virtualization platform then composes the submodels from different VAs and agents from underlying infrastructure layers into an end-to-end performance model. The resulting end-to-end performance model of the virtualized system can then be used for online resource management. This agent-based architecture circumvents the need for static code analysis, while the continuous updates enable the model to update itself if the executed configuration or the workload changes. This way, the longer the application runs, the better the performance model will reflect the real-world system behavior. The second research question was concerned with the extraction of parametric dependencies. We addressed it by drawing parallels to some techniques from machine learning. The first step is to identify which model variables are dependent on which input parameters. The main problem is the exponentially growing search space of possible dependencies. Our approach tackles this problem by leveraging feature selection techniques from machine learning. The identification of parametric dependencies between different variables of the monitoring stream can be framed as a classic application of feature selection. Therefore, we proposed a generic algorithm for the automated identification of parametric dependencies on monitoring streams and applied three different heuristics to filter the resulting dependencies. In the second step, we proposed a meta-selector selecting the most appropriate regression technique for characterizing every dependency based on the characteristics of the available data. Summarizing, the works conducted during the course of this project form valuable contributions to the respective research fields. Although targeted at virtualized environments, some aspects like the use of agentbased architectures for performance model extraction of distributed software systems, the proposed model merging algorithm, and the introduced techniques for the identification and characterization of parametric dependencies can be transferred to non-virtualized domains or even to other disciplines of computer science. Therefore, as part of our future work, we will try to further broaden the scope of the contributions in order to accelerate the progress of the entire field.
Publications
-
Online Learning of Runtime Models for Performance and Resource Management in Data Centers. In Samuel Kounev, Jeffrey O. Kephart, Aleksandar Milenkoski, and Xiaoyun Zhu, editors, Self-Aware Computing Systems. Springer Verlag, Berlin Heidelberg, Germany, 2017
Jürgen Walter, Antinisca Di Marco, Simon Spinner, Paola Inverardi, and Samuel Kounev
-
Self-Tuning Resource Demand Estimation. In Proceedings of the 14th IEEE International Conference on Autonomic Computing (ICAC 2017), July 2017
Johannes Grohmann, Nikolas Herbst, Simon Spinner, and Samuel Kounev
-
Black-box learning of parametric dependencies for performance models. In Proceedings of 13th International Workshop on Models@run.time (MRT), co-located with ACM/IEEE 21st International Conference on Model Driven Engineering Languages and Systems (MOD-ELS 2018), CEUR Workshop Proceedings, October 2018
Vanessa Ackermann, Johannes Grohmann, Simon Eismann, and Samuel Kounev
-
Modeling of Parametric Dependencies for Performance Prediction of Component-based Software Systems at Run-time. In 2018 IEEE International Conference on Software Architecture (ICSA), pages 135–144, April 2018
Simon Eismann, Jürgen Walter, Jóakim von Kistowski, and Samuel Kounev
-
Detecting Parametric Dependencies for Performance Models Using Feature Selection Techniques. In Proceedings of the 27th IEEE International Symposium on the Modelling, Analysis, and Simulation of Computer and Telecommunication Systems, MASCOTS ’19, October 2019
Johannes Grohmann, Simon Eismann, Sven Elflein, Manar Mazkatli, Jóakim von Kistowski, and Samuel Kounev
-
Online model learning for self-aware computing infrastructures. Journal of Systems and Software, 147:1 – 16, 2019
Simon Spinner, Johannes Grohmann, Simon Eismann, and Samuel Kounev
