Already hundreds of millions of PCs are found in homes, offering high computing capacity without being adequately utilized. This paper reveals the potential for energy saving in future home environments, which can be achieved by sharing resources, and concentrating 24/7 computation on a small number of PCs. We present three evaluation methods for assessing the expected performance. A newly created prototype is able to interconnect an arbitrary number of homes by using the free P2P-library FreePastry. The prototype is able to carry out task virtualization by sending virtual machines (VMs) from one home to another, most VMs being of size around 4 MB. We present measurement results from the prototype. We then describe a general model for download sharing, and compare performance results from an analytical model to results obtained from a discrete event simulator. The simulation results demonstrate that it is possible to reach almost optimal energy efficiency for this scenario.

This paper reveals the power saving potential of P2P file sharing in two cases; popular and unpopular files. For popular files, we derive, with regard to BitTorrent, an expression for the optimal time seeders should support leechers. For unpopular files, we extend an existing model by taking into account leechers' power consumption dependent on the load. Leechers are assumed to build a temporary cluster within the P2P-overlay. We determine the required number of active leechers to cope with a given load and compare results from an analytical model to simulation. We demonstrate that it is possible to reach almost optimal energy efficiency for the download scenario by comparing the local case without cooperation with the distributed case where leechers cooperate.

The ability to live migrate virtual machines (VMs) between physical servers without any perceivable service interruption is pivotal for building more energy efficient Cloud Computing infrastructures in the future. Nevertheless, energy efficiency is not worth the effort if quality metrics (e.g., QoS, QoE) are severely decreased by, e.g., dynamic consolidation using live migration. In this work, we present results for a low level approach by patching KVM to implement Xen's live migration stop conditions, and a high level approach by monitoring the progress and service level state of a live migration and estimating the impact on the service level during the remaining migration time. We compare these approaches with an unpatched, vanilla KVM version and different sets of parameters used for live migration. Our service level management approaches offer superior QoS during migration. Especially, they allow to migrate also highly utilized VMs with comparably small influence on QoS.

Avast majority of servers in classical data centers of all scales are underutilized for a significant amount of time. These servers operate at a very low rate of efficiency and consume huge amounts of energy. In this work, we investigate how dynamic consolidation and workload forecasting can be exploited to increase the energy efficiency of a virtualized, heterogeneous server infrastructure. We base our evaluation on real utilization traces of a production system operated by the University of Vienna's central IT department. The traces contain the CPU utilization of more than 30 VMs over a period of four weeks. These VMs offer all kinds of services to students, staff and other visitors. We use these traces to investigate a business infrastructure scenario, where energy costs are just one of several parts of operational costs. We present a novel cost model using configurable penalties for the most important operational cost categories. We compare the total costs of a bin-packing related heuristic and a new genetic VM mapping algorithm used for dynamic consolidation (GA) and load balancing (LB). We tradeoff forecasting against resource reserves in combination with shorter measurement intervals. We demonstrate the flexibility of a genetic algorithm. The GA and LB approaches are directly influenced by penalizing cost parameters. Our cost model allows easy adaption by infrastructure operators to implement custom priorities and optimization goals.

The ability to live migrate virtual machines (VMs) between physical servers without any perceivable service interruption is pivotal for building more energy efficient Cloud Computing infrastructures in the future. Nevertheless, energy efficiency is not worth the effort if quality metrics (e.g., QoS, QoE) are severely decreased by, e.g., dynamic consolidation using live migration. We identify the most significant utilization metrics to predict the service level during live migrations for a web server scenario. We show important correlations, give reasons and draw conclusions for systems using live migration for yielding higher energy efficiency. We also give reasons for extending the current hypervisors' capabilities regarding VM utilization collection and reporting. We present the effects of live migration on service levels for different workload scenarios. In particular, we demonstrate that live migration should be done preventively. This anticipates disproportional high service level degradation due to live migration. We examine the most important utilization metrics for predicting the service level by both stepwise and exhaustive regression. As a result, we can explain 90% of the service level variance during live migration with a single variable, using more variables yields 95%.

Virtualized data centers where several virtual machines (VMs) are hosted per server are becoming more popular due to Cloud Computing. As a consequence of energy efficiency concerns, the exact combination of VMs running on a specific server will most likely change over time. We present experimental results how to use the energy/power consumption logs of a power monitored server as a side-channel that allows us to recognize the exact combination of VMs it currently hosts to a high degree. For classification, we use a maximum log-likelihood approach, which works well for comparably small training and test set sizes. We also show to which degree a specific VM can be recognized, regardless of other VMs currently running on the same server, and show false negative/positive rates. To cross-validate our results, we have used a Kolmogorov-Smirnov test, resulting in comparable quality of recognition within shorter time. In order to clarify whether our approach is generalizable and yields reproducible results, we have set up a second experimental infrastructure in Lyon, using a different hardware platform and power measurement device. We have obtained similar results and have experimented with different CPU frequency scaling governors, yielding comparable quality of recognition. As a result, energy consumption data of servers must be protected carefully, as it is potentially valuable information for an attacker trying to track down a VM to mount further attack steps.

Introduction. Resource monitoring holds a very important position within the energy efficient computing paradigm. The underlying idea is to monitor the energy consumption behavior of appliances in different scenarios to develop consumption signatures. The signatures can help to deduce detailed consumption information which can be used to improve energy savings. There are many contemporary works being done in this direction where researchers investigate resource monitoring systems and their application. Additionally, efforts are being made to utilize collected samples of monitored information in useful ways [1]. Though the overall idea is very strong, there are certain research challenges which need to be overcome before turning this vision into reality. One of the most widely discussed among these is the privacy implications of such systems [3] and usability of collected information in a constructive manner. Detection theory through hypothesis testing is a tool that can be used to advance state-ofthe-art in privacy in energy monitoring. Compromising privacy can be thought of as deducing the profile of a user from observed data. In the context of process energy data monitoring, a profile corresponds to selecting one out of a finite set of N possible processes that are ran on a physical machine. Privacy concerns are raised when a malicious entity builds a set of empirical probability mass functions (p.m.f.), one for each process. Each p.m.f. captures the statistics of instantaneously consumed energy of the process. This task can be performed easily offline by taking multiple observations from each process that is running separately on every machine. A privacy breach exists, if, during the time a process runs on a physical machine, the malicious entity takes observations of the energy consumption level. Various factors may inherently limit the amount of observations taken. The question for the entity that seeks to compomise privacy is to identify the process that is running with good accuracy. Virtualization comes into stage to the support of privacy preservation; By appropriately mixing two or more processes (and thus p.m.f’s) on a virtualized machine, the privacy is protected, in the sense that the individual processes are made indistinguishable. This is also one of the core research questions in EuroNF SJRP SPEC where we want to address this interrelation between energy consumption monitoring and their impact on user privacy.

Home environments promise high potential in terms of resource sharing and energy saving. More and more home computers are running on an always-on basis (e.g., media centers or file sharing clients). Such home environments have not been sufficiently analyzed regarding the possibility of aggregating home user resources in an energy-efficient way. This article describes a future home environment in which available hardware resources (e.g., CPU cycles, disk space, or network capacity) are shared energy efficiently and balanced among end users. Furthermore, the article provides an overview of different virtualization methods that are needed in future home environments to enable cooperation of home networks. Virtualization-related requirements are discussed in detail and virtualization methods and concepts are compared to each other with respect to their usability in the architecture.

Already, hundreds of millions of PCs are found in homes, offering high computing capacity without being adequately utilized. This paper reveals the potential for energy saving in future home environments, which can be achieved by sharing resources, and concentrating 24/7 computation on a small number of PCs. We present three evaluation methods for assessing the expected performance. A newly created prototype is able to interconnect an arbitrary number of homes by using the free P2P library FreePastry. The prototype is able to carry out task virtualization by sending virtual machines (VMs) from one home to another, most VMs being of size around 4 MB. We present measurement results from the prototype. We then describe a general model for download sharing, and compare performance results from an analytical model to results obtained from a discrete event simulator. The simulation results demonstrate that it is possible to reach almost optimal energy efficiency for this scenario.

Meine Dissertation befasst sich mit software-gesteuerter Steigerung der Energie-Effizienz von Rechenzentren. Deren Anteil am weltweiten Gesamtstrombedarf wurde auf 1-2%geschatzt, mit stark steigender Tendenz. Server verursachen oft innerhalb von 3 Jahren Stromkosten, die die Anschaffungskosten ubersteigen. Die Steigerung der Effizienz aller Komponenten eines Rechenzentrums ist daher von hoher okonomischer und okologischer Bedeutung. Meine Dissertation befasst sich speziell mit dem effizienten Betrieb der Server. Ein Grosteil wird sehr ineffizient genutzt, Auslastungsbereiche von 10-20% sind der Normalfall, bei gleichzeitig hohem Strombedarf. In den letzten Jahren wurde im Bereich der Green Data Centers bereits Erhebliches an Forschung geleistet, etwa bei Kuhltechniken. Viele Fragestellungen sind jedoch derzeit nur unzureichend oder gar nicht gelost. Dazu zahlt, inwiefern eine virtualisierte und heterogene Server-Infrastruktur moglichst stromsparend betrieben werden kann, ohne dass Dienstqualitat und damit Umsatzziele Schaden nehmen. Ein Grosteil der bestehenden Arbeiten beschaftigt sich mit homogenen Cluster-Infrastrukturen, deren Rahmenbedingungen nicht annahernd mit Business-Infrastrukturen vergleichbar sind. Hier durfen verringerte Stromkosten im Allgemeinen nicht durch Umsatzeinbusen zunichte gemacht werden. Insbesondere ist ein automatischer Trade-Off zwischen mehreren Kostenfaktoren, von denen einer der Energiebedarf ist, nur unzureichend erforscht. In meiner Arbeit werden mathematische Modelle und Algorithmen zur Steigerung der Energie-Effizienz von Rechenzentren erforscht und bewertet. Es soll immer nur so viel an stromverbrauchender Hardware online sein, wie zur Bewaltigung der momentan anfallenden Arbeitslast notwendig ist. Bei sinkender Arbeitslast wird die Infrastruktur konsolidiert und nicht benotigte Server abgedreht. Bei steigender Arbeitslast werden zusatzliche Server aufgedreht, und die Infrastruktur skaliert. Idealerweise geschieht dies vorausschauend anhand von Prognosen zur Arbeitslastentwicklung. Die Arbeitslast, gekapselt in VMs, wird in beiden Fallen per Live Migration auf andere Server verschoben. Die Frage, welche VM auf welchem Server laufen soll, sodass in Summe moglichst wenig Strom verbraucht wird und gewisse Nebenbedingungen nicht verletzt werden (etwa SLAs), ist ein kombinatorisches Optimierungsproblem in mehreren Variablen. Dieses muss regelmasig neu gelost werden, da sich etwa der Ressourcenbedarf der VMs andert. Weiters sind Server hinsichtlich ihrer Ausstattung und ihres Strombedarfs nicht homogen. Aufgrund der Komplexitat ist eine exakte Losung praktisch unmoglich. Eine Heuristik aus verwandten Problemklassen (vector packing) wird angepasst, ein meta-heuristischer Ansatz aus der Natur (Genetische Algorithmen) umformuliert. Ein einfach konfigurierbares Kostenmodell wird formuliert, um Energieeinsparungen gegenuber der Dienstqualitat abzuwagen. Die Losungsansatze werden mit Load-Balancing verglichen. Zusatzlich werden die Forecasting-Methoden SARIMA und Holt-Winters evaluiert. Weiters werden Modelle entwickelt, die den negativen Einfluss einer Live Migration auf die Dienstqualitat voraussagen konnen, und Ansatze evaluiert, die diesen Einfluss verringern. Abschliesend wird untersucht, inwiefern das Protokollieren des Energieverbrauchs Auswirkungen auf Aspekte der Security und Privacy haben kann.