Large Data Centers Research Articles

Utilization-prediction-aware virtual machine consolidation approach for energy-efficient cloud data centers

In the age of the information explosion, the energy demand for cloud data centers has increased markedly; hence, reducing the energy consumption of cloud data centers is essential. Dynamic virtual machine VM consolidation, as one of the effective methods for reducing energy energy consumption is extensively employed in large cloud data centers. It achieves the energy reductions by concentrating the workload of active hosts and switching idle hosts into low-power state; moreover, it improves the resource utilization of cloud data centers. However, the quality of service (QoS) guarantee is fundamental for maintaining dependable services between cloud providers and their customers in the cloud environment. Therefore, reducing the power costs while preserving the QoS guarantee are considered as the two main goals of this study. To efficiently address this problem, the proposed VM consolidation approach considers the current and future utilization of resources through the host overload detection (UP-POD) and host underload detection (UP-PUD). The future utilization of resources is accurately predicted using a Gray-Markov-based model. In the experiment, the proposed approach is applied for real-world workload traces in CloudSim and were compared with the existing benchmark algorithms. Simulation results show that the proposed approaches significantly reduce the number of VM migrations and energy consumption while maintaining the QoS guarantee.

Embedding individualized machine learning prediction models for energy efficient VM consolidation within Cloud data centers

The fast growth in demand for utility-based IT services has lead to the formation of large scale Cloud data centers. The electrical energy consumption of these data centers results in high operational costs and carbon dioxide emissions. Cloud data centers benefit from the use of virtualization technology to reduce their energy consumption. This technology enables a Cloud data center to allocate its physical resources (CPU, memory, hard disk, network bandwidth) on demand and balance loads between their physical hosts by live migration of Virtual Machines (VMs). However, the migration of VMs can result in Service Level Agreement Violations (SLAVs) and consequently low Quality of Service (QoS). Hence, in this paper, we propose an energy aware VM consolidation algorithm that minimizes SLAVs.Dynamic VM consolidation has three stages: a) Detecting over- and under-utilized hosts; b) Selecting one or more VMs for migration from those hosts; c) Finding destination hosts for the selected VMs. Therefore, the proposed VM consolidation algorithm contains different models for each stage. For the first stage, we developed different fine-tuned Machine Learning (ML) prediction models for individual VMs to predict the best time to trigger migrations from hosts. For the second stage, we lexicographically consider migration time and host CPU usage when selecting VMs to migrate. Finally, a new method based on the Best Fit Decreasing (BFD) algorithm was developed to select a destination host for the VMs being migrated. Our algorithm was compared to a baseline VM consolidation algorithm that used Local Regression for detecting over-utilized hosts, minimum migration time for the VM selection stage and power-aware best fit for the host selection stage. The comparison demonstrated that our VM consolidation algorithm improved energy consumption and SLAVs by 26% and 50%, respectively.

A two steps method of resources utilization predication for large Hadoop data center

SummaryWith the increase of data processing and Hadoop data center construction requirements, the performance of Hadoop data center is limited by inappropriate resources utilization. This paper introduces a new method to predict utilization for large‐scale Hadoop clusters. The new method adopts a two steps model, which includes Hadoop applications' performance simulation and resources utilization prediction. For performance simulation, a new simulator, which integrates baseline test and multilayered network model, is introduced and implemented. A resources utilization predictor is proposed in the second step. By analyzing the pattern of resources utilization, a single task model is proposed. A parallel‐batch‐task‐based (PBT) model, which represents the behavior of real Hadoop applications by integrating the single task model, is introduced. Two test scenarios are configured to verify the performance of our method. For the data center scenario, Terasort, Wordcount, and Hive are selected as benchmarks. In the virtual machines scenario, Terasort is used as benchmark. The experiments show that the error comparing between the simulator results and experimental environment results in most cases is less than 10%. The results confirm that we can locate the resource bottleneck for Hadoop clusters, meanwhile we can agilely configure clusters for applications with massive data.

A rack-aware scalable resource management system for Hadoop YARN

Big data have brought in an era of data exploration and utilisation with MapReduce computational paradigm as its major enabler. Though great efforts through the implementation of Hadoop have made computation scale to tens of thousands of commodity cluster processors, the centralised architecture of resource manager has adversely affected response time in large data centres. The developed model decouples the responsibilities of resource manager by providing another layer where each daemon called rack unit resource manager (RU_RM) carries out the responsibility of allocating resources to compute nodes within its local rack to ensure low latency for large files. The application was developed and tested with Hadoop workload benchmarks used for analysis. Two performance evaluation metrics (efficiency and average task-delay ratio) were used for comparison. Efficiency quantifies average cluster utilisation while average task-delay ratio measures average delay time. Results obtained showed that as file size increases, the developed model outperforms the existing framework.

Disaster Recovery in Cloud Computing Systems: An Overview

With the rapid growth of internet technologies, large-scale online services, such as data backup and data recovery are increasingly available. Since these large-scale online services require substantial networking, processing, and storage capacities, it has become a considerable challenge to design equally large-scale computing infrastructures that support these services cost-effectively. In response to this rising demand, cloud computing has been refined during the past decade and turned into a lucrative business for organizations that own large datacenters and offer their computing resources. Undoubtedly cloud computing provides tremendous benefits for data storage backup and data accessibility at a reasonable cost. This paper aims at surveying and analyzing the previous works proposed for disaster recovery in cloud computing. The discussion concentrates on investigating the positive aspects and the limitations of each proposal. Also examined are discussed the current challenges in handling data recovery in the cloud context and the impact of data backup plan on maintaining the data in the event of natural disasters. A summary of the leading research work is provided outlining their weaknesses and limitations in the area of disaster recovery in the cloud computing environment. An in-depth discussion of the current and future trends research in the area of disaster recovery in cloud computing is also offered. Several work research directions that ought to be explored are pointed out as well, which may help researchers to discover and further investigate those problems related to disaster recovery in the cloud environment that have remained unresolved.

The BNLBox Cloud Storage Service

Large scientific data centers have recently begun providing a number of different types of data storage in order to satisfy the various needs of their users. Users with interactive accounts, for example, might want a POSIX interface for easy access to the data from their interactive machines. Grid computing sites, on the other hand, likely need to provide an X509-based storage protocol, like SRM and GridFTP, since the data management system is built upon them. Meanwhile, an experiment producing large amounts of data typically demands a service that provides archival storage for the safe keeping of their unique data. To access these various types of data, users must use specific sets of commands tailored to their respective storage, making access to their data complex and difficult. BNLBox is an attempt to provide a unified and easy to use storage service for all BNL users, to store their important documents, code and data. It is a cloud storage system with an intuitive web interface for novice users. It provides an automated synchronization feature that enables users to upload data to their cloud storage without manual intervention, freeing them to focus on analysis rather than data management software. It provides a POSIX interface for local interactive users, which simplifies data access from batch jobs as well. At the same time, it also provides users with a straightforward mechanism for archiving large data sets for later processing. The storage space can be used for both code and data within the compute job environment. This paper will describe various aspects of the BNLBox storage service.

PAPSO: A Power-Aware VM Placement Technique Based on Particle Swarm Optimization

With the widespread usage of cloud computing to benefit from its services, cloud service providers have invested in constructing large scale data centers. Consequently, a tremendous increase in energy consumption has arisen in conjunction with its results, including a remarkable rise in costs of operating and cooling servers. Besides, increasing energy consumption has a significant impact on the environment due to emissions of carbon dioxide. Dynamic consolidation of Virtual Machines (VMs) into the minimal number of Physical Machines (PMs) is considered as one of the magic solutions to manage power consumption. The virtual machine placement problem is a critical issue for good VM consolidation. This paper proposes a Power-Aware technique depending on Particle Swarm Optimization (PAPSO) to determine the near-optimal placement for the migrated VMs. A discrete version of Particle Swarm Optimization (PSO) is adopted based on a decimal encoding to map the migrated VMs to the best appropriate PMs. Furthermore, an effective minimization fitness function is employed to reduce power consumption without violating the Service Level Agreement (SLA). Specifically, PAPSO consolidates the migrated VMs into the minimum number of PMs with a major constraint to decrease the number of overloaded hosts as much as possible. Therefore, the number of VM migrations can be reduced drastically by taking into consideration the main sources for VM migrations; overloaded hosts and underloaded ones. PAPSO is implemented in CloudSim and the experimental results on random workloads with different sizes of VMs and PMs show that PAPSO does not violate SLA and outperforms the Power-Aware Best Fit Decreasing algorithm (PABFD). It can reduce about 8.01%, 39.65%, 66.33%, and 11.87% on average in terms of consumed energy, number of VM migrations, number of host shutdowns and the combined metric Energy SLA Violation (ESV), respectively.

Heterogeneous Job Allocation Scheduler for Hadoop MapReduce Using Dynamic Grouping Integrated Neighboring Search

MapReduce is a crucial framework in the cloud computing architecture, and is implemented by Apache Hadoop and other cloud computing platforms. The resources required for executing jobs in a large data center vary according to the job types. In general, there are two types of jobs, CPU-bound and I/O-bound, which require different resources but run simultaneously in the same cluster. The default job scheduling policy of Hadoop is first-come-first-served and therefore, may cause unbalanced resource utilization. Considering various job workloads, numerous job allocation schedulers were proposed in the literature. However, those schedulers encountered the data locality problem or unreasonable job execution performance. This study proposes a job scheduler based on a dynamic grouping integrated neighboring search strategy, which can balance the resource utilization and improve the performance and data locality in heterogeneous computing environments.

ATLAS Data Carousel

The ATLAS experiment at CERN’s LHC stores detector and simulation data in raw and derived data formats across more than 150 Grid sites world-wide, currently in total about 200PB on disk and 250PB on tape. Data have different access characteristics due to various computational workflows, and can be accessed from different media, such as remote I/O, disk cache on hard disk drives or SSDs. Also, larger data centers provide the majority of offline storage capability via tape systems. For the HighLuminosity LHC (HL-LHC), the estimated data storage requirements are several factors bigger than the present forecast of available resources, based on a flat budget assumption. On the computing side, ATLAS Distributed Computing was very successful in the last years with high performance and high throughput computing integration and in using opportunistic computing resources for the Monte Carlo simulation. On the other hand, equivalent opportunistic storage does not exist. ATLAS started the Data Carousel project to increase the usage of less expensive storage, i.e. tapes or even commercial storage, so it is not limited to tape technologies exclusively. Data Carousel orchestrates data processing between workload management, data management, and storage services with the bulk data resident on offline storage. The processing is executed by staging and promptly processing a sliding window of inputs onto faster buffer storage, such that only a small percentage of input data are available at any one time. With this project, we aim to demonstrate that this is the natural way to dramatically reduce our storage cost. The first phase of the project was started in the fall of 2018 and was related to I/O tests of the sites archiving systems. Phase II now requires a tight integration of the workload and data management systems. Additionally, the Data Carousel studies the feasibility to run multiple computing workflows from tape. The project is progressing very well and the results presented in this document will be used before the LHC Run 3.

Innovations in Terahertz Interconnects: High-Speed Data Transport Over Fully Electrical Terahertz Waveguide Links

The continued increase in computing capability that has occurred through the technology scaling foretold by Moore [1] and Dennard et al. [2] has led to a commensurate increase in the input-output requirements of computing systems. This phenomenon has been most significantly experienced in terms of the interconnect technologies in large data centers (DCs) and high-performance computing (HPC) applications. Historical data have shown a two to three times increase in aggregate data rates every two years [3]. This space covers a variety of different interconnect length scales. In a comprehensive survey by Thraskias et al. [4], high-speed interconnects are grouped according to the following taxonomy.

Real-Time Data Center's Telemetry Reduction and Reconstruction Using Markov Chain Models

Large-scale data centers are composed of thousands of servers organized in interconnected racks to offer services to users. These data centers continuously generate large amounts of telemetry data streams (e.g., hardware utilization metrics) used for multiple purposes, including resource management, workload characterization, resource utilization prediction, capacity planning, and real-time analytics. These telemetry streams require costly bandwidth utilization and storage space, particularly at medium-long term for large data centers. This paper addresses this problem by proposing and evaluating a system to efficiently reduce bandwidth and storage for telemetry data through real-time modeling using Markov chain based methods. Our proposed solution was evaluated using real telemetry datasets and compared with polynomial regression methods for reducing and reconstructing data. Experimental results show that data can be lossy compressed up to 75% for bandwidth utilization and 95.33% for storage space, with reconstruction accuracy close to 92%.

A parametric CFD study of computer room air handling bypass in air-cooled data centers

Computer room air handling (CRAH) bypass (BP) method is an application that reduces the total fan power in air-cooled data centers by lowering CRAH fan speed. BP fans supply the complementary fraction of server airflow from warmer room air through lower-resistance floor openings into the plenum. Hence, chillers need to operate at lower temperatures less efficiently, which poses an optimization problem. The reduced-order modeling tools in the literature adequately predict fan power. However, they assume well-mixed temperature at rack inlets to deal with this optimization problem and fail to quantify the impact of ignoring temperature non-uniformities. This study includes an experimental verification of computational fluid dynamics (CFD) modeling for CRAH BP method in predicting temperatures in a data center test-cell. The subsequent parametric study uses both CFD and reduced-order modeling tools on a more representative quadrant of a large data center to investigate induced and forced CRAH BP in both enclosed and open aisle configurations as well as comparing various plenum heights, IT load and server airflow rate. Results indicate where reduced-order modeling tools perform reasonably well and where the need for CFD emerges and identify various favorable data center designs and operating conditions for CRAH BP method.

Study of Failure Hazard Degree in Large Data Centers

Study of Failure Hazard Degree in Large Data Centers

Design and Research of SDN Unified Controller in Large Data Center

Firstly, this paper introduces the current situation and requirements of software defined network (SDN) controller in large data center, and then analyzes the mainstream SDN controller system, and then summarizes the development trends of the data center and SDN controller system. Next, this paper mainly introduces the design architecture and implementation technology of the SDN unified controller. Finally, it summarizes the advantages of the SDN unified controller system and discusses the challenges of future work on the project.

Virtual Machine Migration Planning in Software-Defined Networks

In this paper, we examine the problem of how to schedule the migrations and how to allocate network resources for migration when multiple VMs need to be migrated at the same time. We consider the problem in the Software-defined Network (SDN) context since it provides flexible control on routing. More specifically, we propose a method that computes the optimal migration sequence and network bandwidth used for each migration. We formulate this problem as a mixed integer programming, which is NP-hard. To make it computationally feasible for large scale data centers, we propose an approximation scheme via linear approximation plus fully polynomial time approximation, and obtain its theoretical performance bound. Through extensive simulations, we demonstrate that our fully polynomial time approximation (FPTA) algorithm has a good performance compared with the optimal solution and two state of-the-art algorithms. That is, our proposed FPTA algorithm approaches to the optimal solution with less than 10% variation and much less computation time. Meanwhile, it reduces the total migration time and the service downtime by up to 40% and 20% compared with the state-of-the-art algorithms, respectively.

Long-term projections of the hourly electricity consumption in Danish municipalities

To assess future challenges for the energy system, the transmission - and distribution grids, long-term projections of the spatial distribution of electricity consumption with an hourly resolution is important. Based on 2015 data from hourly meters in Denmark, we develop a model that converts national projections of the annual electricity consumption to projections of the hourly consumption in Danish municipalities. Due to different weights and consumption profiles for categories of customers, both the level and the hourly consumption profile differ between municipalities; also future changes will differ.To illustrate future changes, the latest official projection by the Danish Transmission System Operator is used, showing a moderate increase in classic consumption and a considerable increase in total consumption due to electrification of the heating and transport sectors. In addition, the projection includes a number of large datacentres. Electric vehicles and individual heat pumps will determine the future consumption profile in municipalities, while large datacentres mainly determine the level of consumption at specific nodes in the grid. That is, electric vehicles and heat pumps challenge both the energy system, the transmission and the local distribution grids, while large datacentres mainly challenge the energy system and specific nodes in the overall transmission grid.

Learning Latent Events From Network Message Logs

We consider the problem of separating error messages generated in large distributed data center networks into error events. In such networks, each error event leads to a stream of messages generated by hardware and software components affected by the event. These messages are stored in a giant message log. We consider the unsupervised learning problem of identifying the signatures of events that generated these messages; here, the signature of an error event refers to the mixture of messages generated by the event. One of the main contributions of the paper is a novel mapping of our problem which transforms it into a problem of topic discovery in documents. Events in our problem correspond to topics and messages in our problem correspond to words in the topic discovery problem. However, there is no direct analog of documents. Therefore, we use a non-parametric change-point detection algorithm, which has linear computational complexity in the number of messages, to divide the message log into smaller subsets called episodes, which serve as the equivalents of documents. After this mapping has been done, we use a well-known algorithm for topic discovery, called LDA, to solve our problem. We theoretically analyze the change-point detection algorithm, and show that it is consistent and has low sample complexity. We also demonstrate the scalability of our algorithm on a real data set consisting of $97$ million messages collected over a period of $15$ days, from a distributed data center network which supports the operations of a large wireless service provider.

Superconducting optoelectronic loop neurons

Superconducting optoelectronic hardware has been proposed for large-scale neural computing. In this work, we expand upon the circuit and network designs previously introduced. We investigate circuits using superconducting single-photon detectors and Josephson junctions to perform signal reception, synaptic weighting, and integration. Designs are presented for synapses and neurons that perform integration of rate-coded signals as well as detect coincidence events for temporal coding. A neuron with a single integration loop can receive input from thousands of synaptic connections, and many such loops can be employed for dendritic processing. We show that a synaptic weight can be modified via a superconducting flux-storage loop inductively coupled to the current bias of the synapse. Synapses with hundreds of stable states are designed. Spike-timing-dependent plasticity can be implemented using two photons to strengthen and two photons to weaken the synaptic weight via Hebbian-type learning rules. In addition to the synaptic receiver and plasticity circuits, we describe an amplifier chain that converts the current pulse generated when a neuron reaches threshold to a voltage pulse sufficient to produce light from a semiconductor diode. This light is the signal used to communicate between neurons in the network. We analyze the performance of the elements in the amplifier chain to calculate the energy consumption per photon created. The speed of the amplification sequence allows neuronal firing up to at least 20 MHz, independent of connectivity. We consider these neurons in network configurations to investigate near-term technological potential and long-term physical limitations. By modeling the physical size of superconducting optoelectronic neurons, we calculate the area of these networks. A system with 8100 neurons and 330 430 total synapses will fit on a 1×1cm2 die. Systems of millions of neurons with hundreds of millions of synapses will fit on a 300 mm wafer. For multiwafer assemblies, communication at light speed enables a neuronal pool the size of a large data center (105m2) comprised of trillions of neurons with coherent oscillations at 1 MHz.

Energy Efficient Level by Level Scheduling for Multiple Workflows in Cloud

Cloud computing provides paradigms for hosting a large number of services by providing configurable computing resources on demand, causing multiple applications to be hosted simultaneously in the form of multiple workflows which requires large scale data centers. The reduction in energy consumption is one of the major concerns of large-scale data centers. In this article, the authors develop a multiple workflow scheduling heuristic with the aim to reduce energy consumption along with the execution time. In the proposed approach, first tasks are considered level by level following the precedence constraints as a result of the tasks multiple DAGs could run in parallel. The tasks at each level are considered in the order of their ranking and scheduled on the efficient processor based on their estimated finish time where idle slots of resources are used to reduce the overall makespan and energy consumption. Finally, DVFS is applied during idle slots and the communication phase to reduce energy consumption further by scaling the frequency to appropriate level.

Efficient Provisioning of Security Service Function Chaining Using Network Security Defense Patterns

Network functions virtualization intertwined with software-defined networking opens up great opportunities for flexible provisioning and composition of network functions, known as network service chaining. In the cloud, this allows providers to create service chains tuned to each application type while optimizing resources’ utilization. This is particularly useful to accommodate different tenants’ applications with different security needs. However, considering security provisioning from the single perspective of resources optimization may lead to deployment solutions that do not comply with well-known security-related best practices and recommendations. In this paper, we propose network security defense patterns (NSDP) aimed at leveraging the best practice and know-how from the security experts and at capturing various security constraints to efficiently select compliant security functions’ deployment options. The placement problem being a NP-Hard problem to solve, we also propose a scalable networking and computing resources aware optimization framework to efficiently provision different NSDPs. We further show the feasibility of implementing NSDPs in the cloud infrastructure through the integration of our approach into an open source cloud framework, namely OpenStack, in our test laboratory. The simulation results show the effectiveness of our approach in selecting an optimal placement of the security functions for large data centers with hundreds of thousands of computing nodes, while complying with the predefined security constraints and improving the scalability compared to the current placement algorithms.