Sensor Selection for IT Infrastructure Monitoring

Abstract

Abstract. Supervisory control is the main means to assure a high level performance and availability of large IT infrastructures. Applied control theory is used in physical and virtualization based clustering, autonomic-, self-healing and cloud computing, but similar problems arise in any distributed environment. The selection of a compact, but sufficiently characteristic set of control variables is one of the core problems both for design and run-time complexity. Most results in the literature are based on a single algorithm for variable selection, but our measurements indicate that no single algorithm can generate faithful estimates for all the different operational domains. We propose to use a combination of different model extraction techniques on benchmark-like data logs. The main advantages of this multi-paradigm approach are twofold: it provides good parameter estimators for predictive control in a simple way; and supports the identification of the actual operational domain facilitating context-aware adaptive control, diagnostics and repair.

Qualitative performance control in supervised IT infrastructures

“…in operational domain recognition of all three time horizons, with a large proportion of inaccurate predictions originating from the degrading state, which again could be avoided by implementing a more sophisticated decision algorithm. The above introduced results differ from those shown in our previous work [18] for a practical reason, namely the lack of additional database back-ends and their supporting middleware. As a crucial effect, the stochastics of TPC-W are not scattered on several backends, but influence directly the single underlying instance.…”

“…While an in-depth comparison is out of the scope of this paper, our application experiences during the work presented here, the one reported on in [18] and others conducted during an industrial pilot project clearly show that mRMR is almost always a better choice for variable selection in this setting than the classic, linear approaches, and in contrast to feature extraction methods, maintains the original meaning of variables. We employ neural networks as predictive classifiers.…”

“…A multitiered, distributed On-Line Transaction Processing (OLTP) system serves as demonstrator of the methodology, where control can influence the system by adding and removing servers to and from its tiers. This paper is based upon the results reported in [18], where we presented an approach for solving the variable selection problem. This work fits into the line of papers building system monitoring around an approximate transfer function predicting output QoS metrics from input workload and internal system state value measurements.…”

“…Separation may be necessary to the point where the transfer function has to be composed of functions from very different function classes -this was our main message in [18]. This paper, in turn, investigates how the tech-niques of [18] aid prediction-based qualitative identification of the actual operational domain in order to support diagnosis and triggering corrective actions.…”

“…As in our previous work [18], we use the bioinformatics inspired minimum Redundancy Maximum Relevancy (mRMR) algorithm [19] for feature selection. Although greedy by nature, mRMR selects an optimal set of variables instead of individual "best" ones, thus it aims at covering every defining aspect of the examined system with a minimal set of features.…”