Markus Hedwig

Automated control for elastic n-tier workloads based on empirical modeling

Elastic n-tier applications have non-stationary workloads that require adaptive control of resources allocated to them. This presents not only an opportunity in pay-as-you-use clouds, but also a challenge to dynamically allocate virtual machines appropriately. Previous approaches based on control theory, queuing networks, and machine learning work well for some situations, but each model has its own limitations due to inaccuracies in performance prediction. In this paper we propose a multi-model controller, which integrates adaptation decisions from several models, choosing the best. The focus of our work is an empirical model, based on detailed measurement data from previous application runs. The main advantage of the empirical model is that it returns high quality performance predictions based on measured data. For new application scenarios, we use other models or heuristics as a starting point, and all performance data are continuously incorporated into the empirical models knowledge base. Using a prototype implementation of the multi-model controller, a cloud testbed, and an n-tier benchmark (RUBBoS), we evaluated and validated the advantages of the empirical model. For example, measured data show that it is more effective to add two nodes as a group, one for each tier, when two tiers approach saturation simultaneously.

Experimental evaluation of N-tier systems: Observation and analysis of multi-bottlenecks

In many areas such as e-commerce, mission-critical N-tier applications have grown increasingly complex. They are characterized by non-stationary workloads (e.g., peak load several times the sustained load) and complex dependencies among the component servers. We have studied N-tier applications through a large number of experiments using the RUBiS and RUBBoS benchmarks. We apply statistical methods such as kernel density estimation, adaptive filtering, and change detection through multiple-model hypothesis tests to analyze more than 200GB of recorded data. Beyond the usual single-bottlenecks, we have observed more intricate bottleneck phenomena. For instance, in several configurations all system components show average resource utilization significantly below saturation, but overall throughput is limited despite addition of more resources. More concretely, our analysis shows experimental evidence of multi-bottleneck cases with low average resource utilization where several resources saturate alternatively, indicating a clear lack of independence in their utilization. Our data corroborates the increasing awareness of the need for more sophisticated analytical performance models to describe N-tier applications that do not rely on independent resource utilization assumptions. We also present a preliminary taxonomy of multi-bottlenecks found in our experimentally observed data.

Automated News Reading: Stock Price Prediction Based on Financial News Using Context-Specific Features

We examine whether stock price effects can be automatically predicted analyzing unstructured textual information in financial news. Accordingly, we enhance existing text mining methods to evaluate the information content of financial news as an instrument for investment decisions. The main contribution of this paper is the usage of more expressive features to represent text and the employment of market feedback as part of our word selection process. In our study, we show that a robust Feature Selection allows lifting classification accuracies significantly above previous approaches when combined with complex feature types. That is because our approach allows selecting semantically relevant features and thus, reduces the problem of over-fitting when applying a machine learning approach. The methodology can be transferred to any other application area providing textual information and corresponding effect data.

CloudXplor: a tool for configuration planning in clouds based on empirical data

Configuration planning for modern information systems is a highly challenging task due to the implications of various factors such as the cloud paradigm, multi-bottleneck workloads, and Green IT efforts. Nonetheless, there is currently little or no support to help decision makers find sustainable configurations that are systematically designed according to economic principles (e.g., profit maximization). This paper explicitly addresses this shortcoming and presents a novel approach to configuration planning in clouds based on empirical data. The main contribution of this paper is our unique approach to configuration planning based on an iterative and interactive data refinement process. More concretely, our methodology correlates economic goals with sound technical data to derive intuitive domain insights. We have implemented our methodology as the CloudXplor Tool to provide a proof of concept and exemplify a concrete use case. CloudXplor, which can be modularly embedded in generic resource management frameworks, illustrates the benefits of empirical configuration planning. In general, this paper is a working example on how to navigate large quantities of technical data to provide a solid foundation for economical decisions.

Bottleneck detection using statistical intervention analysis

The complexity of todays large-scale enterprise applications demands system administrators to monitor enormous amounts of metrics, and reconfigure their hardware as well as software at run-time without thorough understanding of monitoring results. The Elba project is designed to achieve an automated iterative staging to mitigate the risk of violating Service Level Objectives (SLOs). As part of Elba we undertake performance characterization of system to detect bottlenecks in their configurations. In this paper, we introduce our concrete bottleneck detection approach used in Elba, and then show its robustness and accuracy in various configurations scenarios. We utilize a wellknown benchmark application, RUBiS (Rice University Bidding System), to evaluate the classifier with respect to successful identification of different bottlenecks.

Empirical analysis of database server scalability using an N-tier benchmark with read-intensive workload

The performance evaluation of database servers in N-tier applications is a serious challenge due to requirements such as non-stationary complex workloads and global consistency management when replicating database servers. We conducted an experimental evaluation of database server scalability and bottleneck identification in N-tier applications using the RUBBoS benchmark. Our experiments are comprised of a full scale-out mesh with up to nine database servers and three application servers. Additionally, the fourtier system was run in a variety of configurations, including two database management systems (MySQL and PostgreSQL), two hardware node types (normal and low-cost), and two database replication techniques (C-JDBC and MySQL Cluster). In this paper we present the analysis of results generated with a read-intensive interaction pattern (browse-only workload) in the client emulator. These empirical data can be divided into two kinds. First, for a relatively small number of servers, we find simple hardware resource bottlenecks. Consequently, system throughput increases with an increasing number of database (and application) servers. Second, when sufficient hardware resources are available, non-obvious database related bottlenecks have been found that limit system throughput. While the first kind of bottlenecks shows that there are similarities between database and application/web server scalability, the second kind of bottlenecks shows that database servers have significantly higher sophistication and complexity that require in-depth evaluation and analysis.

Risk-Aware Service Level Agreement Design for Enterprise Information Systems

Effective information systems have become key to sustainable business practices. However rising costs of operation and a growing system complexity are driving the search for a more efficient delivery of corporate computing. New technologies in the emerging service world such as cloud computing provide powerful alternatives to traditional IT operation concepts. Nonetheless, executive decision makers still have reservations about migrating to this new technology. In addition to security concerns, a key issue is the still prevailing lack of strict SLAs in these service offerings. In fact, service providers hesitate to offer strictly binding SLAs because assessing economic risk exposure is a major challenge. In this paper, we present a novel model for sustainable SLA design for enterprise information systems. Our model combines various state-of-the-art concepts from the field of system management and balances the failure risk with the cost of operation. More concretely, our model helps IT decision makers to understand the relationship between the operation cost and the service quality. Consequently, the minimum economic price of a service and the corresponding operation strategy, given the customer requirements and the infrastructure characteristics, can be determined based on our approach.

Datacenter Investment Support System (DAISY)

While the cloud computing paradigm provides the technological foundation for previously unmatched efficiency in information systems, its economic implications are complex and significant. Bringing the benefits of this state-of-the-art technology to corporate datacenters requires a systematic cost analysis of current best-practice in conjunction with the new possibilities. We address this challenge and present a novel resource management model based on marginal cost of computing. DAISY (Datacenter Investment Support System) aids in deriving the optimal investment strategy for large enterprise computing infrastructures taking into account all buy and rental options. Our model-based framework is evaluated through the analysis of different time discrete and continuous scenarios. Our main contribution is the thorough economic analysis of the dependencies between static and adaptive modes of operation for modern computing infrastructure. Our results clearly show the ambivalent economic character of computing as a commodity and that further research is necessary to guarantee sustainable economic decisions.

A new perspective on experimental analysis of N-tier systems: Evaluating database scalability, multi-bottlenecks, and economical operation

Economical configuration planning, component performance evaluation, and analysis of bottleneck phenomena in N-tier applications are serious challenges due to design requirements such as non-stationary workloads, complex non-modular relationships, and global consistency management when replicating database servers, for instance. We have conducted an extensive experimental evaluation of N-tier applications, which adopts a purely empirical approach the aforementioned challenges, using the RUBBoS benchmark. As part of the analysis of our exceptionally rich dataset, we have experimentally investigated database server scalability, bottleneck phenomena identification, and iterative data refinement for configuration planning. The experiments detailed in this paper are comprised of a full scale-out mesh with up to nine database servers and three application servers. Additionally, the four-tier system was run in a variety of configurations, including two database management systems (MySQL and PostgreSQL), two hardware node types (normal and low-cost), two replication strategies (wait-all and wait-first—which approximates primary/ secondary), and two database replication techniques (C-JDBC and MySQL Cluster). Herein, we present an analysis survey of results mainly generated with a read/write mix pattern in the client emulator.

Integrated Modeling of Technical and Business Aspects in Service Networks

The current trend towards a global services economy provides significant opportunities and challenges. For establishing complex services and delivering competitive advantages, several service providers have to work together. This collaboration creates a service network as an organizational form to be managed by a so-called service integrator. Within a service network, multiple dependencies between the resulting service and the contributions of the various service providers exist, on both technical and business aspects. In addition to the functional aspects, the non-functional service properties and respective service levels are of great importance. Successful joint management of the technical and business dependencies is a key prerequisite for the successful management of service networks.