Skip to Main content Skip to Navigation

Control Theory for Computing Systems: Application to big-data cloud services & location privacy protection

Abstract : This thesis aims at investigating techniques to build and control efficient, dependable and privacy-preserving computing systems. Ad-hoc service configuration require a high level of expertise which could benefit from automation in many ways. A control algorithm can handle bigger and more complex systems, even when they are extremely sensitive to variations in their environment. However, applying control to computing systems raises several challenges, e.g. no physics governs the applications. On one hand, the mathematical framework provided by control theory can be used to improve automation and robustness of computing systems. Moreover, the control theory provides by definition mathematical guarantees that its objectives will be fulfilled. On the other hand, the specific challenges of such use cases enable to expand the control theory itself. The approach taken in this work is to explore in details two application computing systems: location privacy and cloud services. A third use-case on the use of control for machine learning algorithm is presented in appendix. Those use-cases are complementary in the nature of their technologies, scale and end-users. The widespread of mobile devices has fostered the broadcasting and collection of users’ location data. It enables users to benefit from a personalized service and service providers or any other third party to derive useful information from the mobility databases, whereas it also exposes highly sensi- tive personal data. To overcome this privacy breach, algorithms have been developed that modify the user’s mobility data, hopefully to hide some sensitive information, called Location Privacy Protection Mechanisms (LPPMs). However, those tools are not easily configurable by non experts and are static processes that do not adapt to the user’s mobility. We develop two tools, one for already collected databases and one for online usage, that, by tuning the LPPMs, guarantee to the users objective-driven levels of privacy protection and of service utility preservation. First, we present an automated tool able to choose and configure LPPMs to protect already collected databases while ensuring a trade-off between privacy protection and database processing quality. Second, we present the first formulation of the location privacy challenge in control theory terms (plant and control, disturbance and per- formance signals), and a feedback controller to serve as a proof of concept. In both cases, design, implementation and validation has been done through experiments using data of real users. The surge in data generation of the last decades, the so-called bigdata, has lead to the development of frameworks able to analyze them, such as the well known MapReduce. Advances in computing practices have also settled the cloud paradigms (online ready-to-use resources to rent) as premium solution for all kind of users. In this work, we focus on performance of MapReduce jobs running on clouds and thus develop advanced monitoring techniques of the jobs execution time and the platform availability; by tuning the resource cluster size and realizing admission control, in spite of the unpre- dictable client workload. In order to deal with the non linearities of the MapReduce system, a robust adaptive feedback controller has been designed. To reduce the cluster utilization and costs, we present a new event-based triggering mechanism formulation combined with an optimal predictive controller. Evaluation is done on a MapReduce benchmark suite running on a large-scale cluster, and using real jobs workloads. Learning algorithms are now prevalent in both the research and industry worlds. While they show impressive results in terms of performance, other aspects has been neglected so far, such as automa- tion, robustness or privacy. Machine learning algorithms control is investigated in two complementary ways: robustness regarding noise in the dataset, and the parametrization of the algorithms, with the introduction of feedback action. Results are validated using classic datasets and task-specific ones.
Complete list of metadata

Cited literature [255 references]  Display  Hide  Download
Contributor : Sophie Cerf <>
Submitted on : Tuesday, August 27, 2019 - 3:58:07 PM
Last modification on : Thursday, November 19, 2020 - 1:02:12 PM


  • HAL Id : tel-02272258, version 1


Sophie Cerf. Control Theory for Computing Systems: Application to big-data cloud services & location privacy protection. Systems and Control [cs.SY]. UNIVERSITÉ GRENOBLE ALPES, 2019. English. ⟨tel-02272258⟩



Record views


Files downloads