Daniel Versick

Reducing Energy Consumption by Load Aggregation with an Optimized Dynamic Live Migration of Virtual Machines

Energy consumption of data centers has been in-creasing continuously during the last years due to rising demands of computational power especially in current Grid- and Cloud-Computing systems. One promising approach of reducing this energy consumption is the consolidation of servers by virtualization. Many low loaded computer systems are virtualized and run on few physical servers for reducing the number of energy-consuming computers. At present this consolidation is usually done statically, thus, the administrator of a data center manually migrates many virtual machines with low load onto one physical server which may lead to overloading when the workload is rising unexpectedly. Dynamic server migration that adapts the number of running physical machines to the current workload overcomes these problems. Physical machines can be highly loaded and in case of further rising load virtual machines are migrated to other physical server systems that have been switched on. Such dynamic load aggregation approaches are rarely used and typically only consider few criteria for migration. This paper presents a classification of migration criteria for live migration of virtual machines in load aggregation environments and proposes an algorithm for combining many different kinds of migration criteria to a clustering-based metric. Thus, the novel load aggregation algorithm optimizes energy consumption as well as other migration criteria like runtime performance of applications.

Power consumption estimation of CPU and peripheral components in virtual machines

Energy consumption of IT increased continuously during the last decades. Numerous works have been accomplished for improving energy efficiency of hardware whereas software energy efficiency has been ignored for a long time. This contribution presents a novel approach for estimating energy consumption of computer systems in dependency of software-caused workloads in different execution environments. The system is the basis for automatic optimization of software execution in an energy-efficient way by finding the best-suiting host computer (and best-suiting peripheral devices). Thus, it opens novel ways to further improve energy-efficiency of IT systems by migrating software-caused load to an energy-efficient target. Exemplary, the approach is tested in a virtualized data center environment, where virtual machines are the applications. The presented approach is a vehicle for automatically computing an energy-efficient virtual machine placement. The paper presents a new algorithm for estimating virtual machine power consumption, which consists of CPU power consumption estimation as well as power usage estimation of peripheral components like hard disk drive and network interface controller. The accuracy of the presented approach is proved by means of measurements.

Energy consumption estimation of virtual machines

Energy consumption of IT increased continously during the last decades. Numerous works have been accomplished for improving energy efficiency of hardware whereas software energy efficiency has been ignoried for a long time. This contribution presents a novel approach for estimating energy consumption of applications in different execution environments. The system is the basis for automatic optimization of software execution in an energy-efficient way by finding the best-suiting host computer. Thus, it opens novel ways to further improve energy-efficiency of IT systems. Exemplary, the approach is tested in a virtualized data center environment, where virtual machines are the applications. The presented approach is a vehicle for automatically finding the most energy-efficient host machine for any virtual machine system. The paper presents a new algorithm for estimating virtual machine power consumption and shows the accuracy of the presented approach by means of measurements.

Integrated Performance Analysis of Computer Systems (IPACS). Benchmarks for Distributed Computer Systems

ABSTRACT The IPACS-Project (Integrated Performance Analysis of Computer Systems), which was founded by the German Federal Department of Education, Science, Research and Technology (BMBF), wants to define a new Basis for measuring system performance of distributed systems. Its objective is to develop methods for measuring system performance on High Performance Computers (HPC) based on low level benchmarks, compute kernels, open source- and commercial application benchmarks. Additionally, it covers the development of methods for performance modelling and prediction of commercial codes. A further significant element is the integration into a benchmark environment consisting of a web based repository and a distributed benchmark-execution framework that ensures an easy usability and enables a just-in-time analysis of benchmark results.

A Clustered GALS NoC Architecture with Communication-Aware Mapping

As processors migrate to multi- and many-core architectures, the role of the communication network becomes more important. Efficient communication architecture can drastically improve overall system performance. Taking into account the application behavior can facilitate system-level solutions that manage the communication cost. To address this issue, we propose a Clustered Globally Asynchronous Locally Synchronous Network-on-Chip (C-GALS NoC) communication architecture. C-GALS NoC is composed of local, synchronous clusters and a global asynchronous network. Additionally, we propose a cluster based communication-aware mapping algorithm (CAM) for mapping the application tasks to the C-GALS NoC, while minimizing the communication cost. The synergy of the C-GLAS NoC and the CAM algorithm results in a system-level mechanism that, according to our results, provides up to 2x and 3x, in performance and power improvement, respectively, in comparison with a regular GALS NoC. Finally, we demonstrate that C-GALS NoC is standard-cell compatible by synthesizing it using Design Compiler.

Transdisciplinary Challenges of Scientific Cloud Computing

Cloud Computing implements the next generation distributed system architecture by offering on-demand compute and storage resources for arbitrary customers in a scalable and self-manageable fashion (Mell & Grance, 2011). Information Technology (IT) resources, platforms and services are made available at virtually unlimited scale for everybody, everywhere, and anytime. Cloud Computing therefore not only leads to an outsourcing of enterprise IT services, it furthermore fosters a breathing IT infrastructure which adapts itself to the current size and demand of any organization. Thus, it supports the trend of IT to higher automation, smaller costs, and increased service levels (Gibson & Kasravi, 2012). The recent trend for Cloud Computing led to a variety of new commercial providers and to a wide industrial uptake of the according technology. During the last years, the market was subject to a tremendous growth of up to 36%, the size is estimated to be 19.5 billion USD in 2016 (Columbus, 2013). It is undebatable that, even in the light of unsolved data protection and trust issues, Cloud Computing made its way into the default tool chain of modern IT-based environments. The NIST reference architecture specification (Liu et al., 2012) defines five major actors in Cloud Computing: cloud consumer, cloud provider, cloud carrier, cloud auditor and cloud broker. For most scenarios, it is sufficient to focus on the relationship between provider and consumer. Both have a particular balance in their shared control of the elastic cloud resources, expressed by three well-known service models infrastructure-as-a-service (IaaS), platform-as-a-service (PaaS), and software-as-a-service (SaaS). As orthogonal aspect to the service model, cloud infrastructures can follow the public cloud or private cloud deployment model. Hybrid setups may also exist, were private cloud installations perform dynamic offloading to public cloud offerings in case of scalability issues (Sotomayor, Montero, Llorente, & Foster, 2009). On the lowest service level, the IaaS-based cloud service provisioning, the consumer gets maximum control of the cloud resources by the means of virtualization. Cloud systems in this category consist of virtualized hardware resources at the provider side, managed through web interfaces or remote APIs by the cloud consumer. The provisioning architecture has to be flexible and highly scalable, virtual resources are directly and dynamically instantiated, migrated, checkpointed, and destroyed by the consumers. The provider has the responsibility to manage the according hardware and low-level software scalability, but leaves the application and operating system responsibility completely to the user. Popular examples are the Amazon Elastic Compute Cloud (EC2) and the Microsoft Azure environment.

The CÆSARA architecture for power and thermal-aware placement of virtual machines

Energy consumption of data centers increased continously during the last decades. As current techniques for improving their energy efficiency are usually limited to one type of data center components, it is difficult to employ them for a holistic optimization which may utilize synergies between all components. This paper shortly presents architecture and working principles of the novel energy management system CÆSARA that applies a comprehensive view for evaluating an energy-efficient virtual machine placement in current virtualization environments for minimizing the number of running energy-consuming physical machines. In contrast to existing solutions, our placement algorithm does not only aggregate load to a minimum number of servers - it also discovers which servers in conjunction with their infrastructure act as most energy-efficient migration targets. Thus, the CÆSARA architecture unifies two worlds which are usually separated when finding an energy-efficient configuration: the inner computer world which can be characterized by load values and the infrastructure world which is characterized by physical parameters. This paper presents the architecture of CÆSARA, defines properties for an energy-efficient virtual machine placement algorithm, and sketches how the integration of infrastructural energy consumption can be employed by using a thermodynamic approach. Exemplarily, this is realized by the example of server cooling equipment.

Environment for I/O performance measurement of distributed and local secondary storage

Whilst the performance gap between main memory and secondary storage of computer systems increased rapidly during the last years, especially the analysis of secondary storage performance becomes very important. For that purpose this paper introduces a plugin-based framework providing an integrated execution environment for new benchmarks and offering analysis tools for the complex result output of particularly I/O measurements. As an example of this environment we describe the PRIOmark, an I/O benchmark that characterizes file system performance of single and distributed systems by means of two different file system interfaces. Additionally, the I/O performance of three file systems, two file system interfaces as well as different I/O workloads are compared to show the flexibility of the presented PRIOmark.

A mathematical model for the transitional region between cache hierarchy levels

The knowledge of internal structures of the cache-memory hierarchy and its performance is very important in modern computer systems. Therefor, this paper introduces a mathematical model that describes the transition between Level 1 and Level 2 cache of current processors. The theoretical predictions are proved by measurements for two Intel CPUs and an UltraSparc II system.