COMPUTER 0018-9162/16/$33.00 © 2016 IEEE PUBLISHED BY THE IEEE COMPUTER SOCIETY JULY 2016 53 RESEARCH FEATURE A Model-Based Approach to Designing Self- Aware IT Systems and Infrastructures Samuel Kounev, Nikolaus Huber, and Fabian Brosig, University of Würzburg Xiaoyun Zhu, Futurewei Technologies M odern IT system archi- tectures are becom- ing increasingly distributed, have loosely coupled services, and are often deployed on virtualized infrastruc- tures that abstract physical layers to improve system efciency. The ben- efts of distributed architectures and virtualized infrastructures come at the cost of higher system complexity and dynamics; the inherent semantic gap between application-level metrics and resource allocations at the phys- ical and virtual layers signifcantly increases the complexity of managing end-to-end application performance. To tackle this challenge, tech- niques for online performance prediction are needed that enable the continuous prediction of three per- formance aspects: application workload changes, the efects of these changes on system performance, and the expected impact of possible adaptation actions. Online performance prediction can be the basis for designing systems that proactively adapt to changing operating conditions, thus enabling self-aware performance and resource management. 1 (See the “Self-Aware Computing Systems” sidebar for more information.) We have developed a model-based approach to design- ing self-aware IT systems along with the Descartes Mod- eling Language (DML), 3 an architecture-level language that is central to online performance prediction and pro- active model-based system adaptation. We have applied our model-based design approach in several case stud- ies with realistic environments and in cooperation with industrial partners. 4,5 In an evaluation against a trigger-based approach (which relies on custom metrics and specifed thresholds to execute predefned reconfg- uration actions), our approach maintained acceptable Results of a fve-year research project and several industrial collaborations have produced tools that model the individual effects and complex dynamic interactions between an IT system’s application workload and resource contention at multiple levels in the execution environment. An evaluation shows signifcant resource efciency gains without sacrifcing the performance specifed in service-level agreements. RESEARCH FEATURE