Performance models, design and run time management of big data applications

Nowadays the big data paradigm is consolidating its central position in the industry, as well as in society at large. Lots of applications, across disparate domains, operate on huge amounts of data and offer great advantages both for business and research. As data intensive applications (DIAs) gain more and more importance over time, it is fundamental for developers and maintainers to have the support of tools that enhance their efforts since early design stages and until run time. The present dissertation takes this perspective and addresses some pivotal issues with a quantitative approach, particularly in terms of deadline guarantees to ensure quality of service (QoS). Technically interesting scenarios, such as cloud deployments supporting a mix of heterogeneous applications, pose a series of challenges when it comes to predicting performance and exploiting this information for optimal design and management. Performance models, with their potential for what if analyses and informed design choices about DIAs, can be a major tool for both users and providers, yet they bring about a trade-off between accuracy and efficiency that may be tough to generally address. The picture is further complicated by the adoption of the cloud technology, which means that assessing operating costs in advance becomes harder, but also that the contention observed in data centers strongly affects big data applications’ behavior. For all these reasons, ensuring QoS for novel DIAs is a difficult task that needs to be addressed in order to favor further development of the field. Over this background, the present dissertation takes two main routes towards facing such challenges. At first we describe and discuss a number of performance models based on various formalisms and techniques. Among these, there are both basic models aimed at predicting specific metrics, like response time or throughput, and more specialized extensions that target the impact on big data systems of some design decisions, e.g., privacy preserving mechanisms or cloud pricing models. On top of this, the proposed models are variously positioned across the spectrum between efficiency and accuracy, thus enabling different trade-offs depending on the main requirements at hand. This is relevant in the second main part of this dissertation, where performance prediction is at the core of some formulations for capacity allocation and cluster management. In order to obtain optimal solutions to these problems, in one case at design time and in the other at run time, we adopt both mathematical programming and several performance models, according to the different constraints on solving times and accuracy. More in detail, we propose performance models based on queueing networks (QNs), stochastic well formed nets (SWNs), and machine learning (ML). This variety is justified by the different uses of each methodology. ML provides algebraic formulas for execution times, which are perfectly fit to be added as constraints in our optimization problems’ mathematical programming formulations, thus yielding initial solutions in closed form. Since ML can reliably provide accurate predictions only in regions properly explored during the training phase, the optimal solution is searched via a simulation-optimization procedure based on analytical models like QNs or SWNs, which in contrast are quite insensitive to the parameter range of evaluation, being devised from first principles. These kind of models boast relative errors below 10 % on average when predicting response times. In terms of optimization, first of all we consider the design time problem of capacity allocation in a cloud environment. The design space is explored via both ML and simulation techniques, so as to choose the best virtual machine type in the catalog offered by cloud providers and, subsequently, determine the minimum cost configuration that satisfies QoS constraints. We show also how this optimization approach was applied during the design phase of a tax fraud detection product developed by industrial partners, i.e., NETF Big Blu. Afterwards we also considered the run time issue of finding the minimum tardiness schedule for a set of jobs when the current workload exceeds predictions and the deployed capacity is not enough to ensure the agreed upon QoS. Thanks to the varied efficiency of performance models, it is possible to solve the design time problem in a matter of hours, whilst run time instances are solved within minutes, consistently with the different requirements.