Search

Since today's HPC and data center systems can comprise hundreds of thousands of servers and beyond, it is crucial to equip them with a network that provides high performance. New topologies proposed to achieve such performance need to be evaluated under different traffic conditions, aiming to closely replicate real-world scenarios. While most optimizations should be guided by common traffic patterns, it is essential to ensure that no pathological traffic pattern can compromise the entire system. Determining synthetic adversarial traffic patterns for a network typically relies on a thorough understanding of its topology and routing. In this paper, we address the problem of identifying a generic adversarial traffic pattern for low-diameter direct interconnection networks. We first focus on Random Regular Graphs (RRGs), which represent a typical case for these networks. Moreover, RRGs have been proposed as topologies for interconnection networks due to their superior scalability and expandability, among other advantages. We introduce Ant Mill, an adversarial traffic pattern for RRGs when using routes of minimal length. Secondly, we demonstrate that the Ant Mill traffic pattern is also adversarial in other low-diameter direct interconnection networks such as Slimfly, Dragonfly, and Projective networks. Ant Mill is thoroughly motivated and evaluated, enabling future studies of low-diameter direct interconnection networks to leverage its findings. [ABSTRACT FROM AUTHOR]

Thermal-aware (TA) task allocation is one of the most effective software-based dynamic thermal management techniques to minimize energy consumption in data centers (DCs). Compared to its counterparts, TA scheduling attains significant gains in energy consumption. However, the existing literature overlooks the heterogeneity of computing elements in terms of thermal constraints while allocating or migrating user jobs, which may significantly affect the reliability of racks and all the equipment therein. Moreover, the workload distribution among these racks/servers is not fair and efficient in terms of thermal footprints; it is potentially beneficial to determine the workload proportion for each computing node (rack/server) based on its marginal contribution in disturbing the thermal uniformity (TU) in a DC environment. To solve the said problems, we model the workload distribution in DCs as a coalition formation game with the Shapley Value (SV) solution concept. Also, we devise Shapley Workload (SW), a TA scheduling scheme based on the SV to optimize the TU and minimize the cooling cost of DCs. Specifically, the scheduling decisions are based on the ambient effect of the neighboring nodes, for the ambient temperature is affected by the following two factors: (1) the current temperature of computing components and (2) the physical organization of computing elements. This results in lower temperature values and better TU, consequently leading to lower cooling costs. Simulation results demonstrate that the proposed strategy greatly reduces the total energy consumption compared to the existing state-of-the-art. [ABSTRACT FROM AUTHOR]

This paper presents HiperView, a visual analytics framework monitoring and characterizing the health status of high-performance computing systems through a RESTful interface in real time. The primary objectives of this visual analytical system are: (1) to provide a graphical interface for tracking the health status of a large number of data center hosts in real-time statistics, (2) to help users visually analyze unusual behavior of a series of events that may have temporal and spatial correlation, and (3) to assist in performing preliminary troubleshooting and maintenance with a visual layout that reflects the actual physical locations. Two use cases were analyzed in detail to assess the effectiveness of the HiperView on a medium-scale, Redfish-enabled production high-performance computing system with a total of 10 racks and 467 hosts. The visualization apparatus has been proven to offer the necessary support for system automation and control. Our framework's visual components and interfaces are designed to potentially handle a larger-scale data center of thousands of hosts with hundreds of various health services per host. [ABSTRACT FROM AUTHOR]

The demand for higher computing power increases and, as a result, also leads to an increased demand for services hosted in cloud computing environments. It is known, for example, that in 2018 more than 4 billion people made daily access to these services through the Internet, corresponding to more than half of the world's population. To support such services, these clouds are made available by large data centers. These systems are responsible for the increasing consumption of electricity, given the increasing number of accesses, increasing the demand for greater communication capacity, processing and high availability. Since electricity is not always obtained from renewable resources, the relentless pursuit of cloud services can have a significant environmental impact. In this context, this paper proposes an integrated and dynamic strategy that demonstrates the impact of the availability of data center architecture equipment on energy consumption. For this, we used the technique of modeling colored Petri nets (CPN), responsible for quantifying the cost, environmental impact and availability of the electricity infrastructure of the data centers under analysis. Such proposed models are supported by the developed tool, where data center designers do not need to know CPN to compute the metrics of interest. A case study was proposed to show the applicability of the proposed strategy. Significant results were obtained, showing an increase in system availability of 100%, with equivalents operating cost and environmental impact. [ABSTRACT FROM AUTHOR]

High-performance computing in a cloud environment may require massive data transfer among some of the virtual machines (VMs). These VMs are deployed in physical machines (hosts) of a data center. The data transfer among the communicating VMs may use the same shared communication links of the data center. Hence, it is important to have efficient bandwidth allocation policies for different data transfer requests (DTRs) which result in better utilization of bandwidth and fair allocation among the DTRs. In this paper, a few bandwidth allocation policies are proposed and their performances are analyzed. While designing these policies, the objective is the maximization of throughput and bandwidth utilization while minimizing the service time and turnaround time. Some of the policies are based on integer linear programming (ILP) which runs in exponential time while others are based on polynomial-time heuristics. Experimental results show that the performances of heuristic-based policies are comparable to those given by ILP-based exponential time policies. [ABSTRACT FROM AUTHOR]

In recent years, data centers have deployed many rapidly growing Internet-scale services. As the size of the data center continues to expand, it not only causes huge energy consumption, but also brings environmental problems such as huge carbon emissions. However, for an Internet-scale data center, it is a challenge to optimize power consumption while meeting the growing demands of Internet-scale applications. Moreover, current data centers usually have heterogeneous servers (with different power consumption and computing capacity), which causes the dynamic task placement problem more complicated. In this work, we first observe two kinds of heterogeneity from the analysis of a public trace that was collected from a Google data center. Considering the demands in energy consumption and performance, we model the dynamic task placement problem in a heterogeneous Internet-scale data center and propose a heuristic cost-aware algorithm to solve it. By simulating and comparing with the other two scheduling algorithms, our proposed heuristic algorithm can gain a well energy saving and keep application performance within acceptable limits. [ABSTRACT FROM AUTHOR]

The technology, cloud computing, in present days, is vastly used due to the services it provides and the ease with which they can be availed. The enormous development of the Internet technology is due to the advent of the concept of cloud. Along with its benefits, cloud computing brings along itself a detrimental side effect, i.e., carbon emission. This is due to the massive energy consumption in the cloud data centers. Reduction in energy consumption in cloud is thus one of the major challenges among the researchers. This work conducts a thorough study of the various techniques that help in minimization of energy consumption in data centers. It also explores and proposes approaches to reduce the same, eventually making the environment greener. In the proposed work, prediction mechanism has been adopted and implemented on the existing Minimization of Migration (MM) policy for large history data set, followed by dynamic thresholding mechanism in place of static thresholds. Rigorous simulations have been conducted, and the results show reduction in cloud data center energy consumption. [ABSTRACT FROM AUTHOR]

With the increasing demands for cloud computing services, the development of technologies based on virtualization in data centers was noticed. In the virtualized data center, the efficient mapping of virtual machines to physical machines is done using the consolidation technique. Due to the advantages of the server consolidation technique, a large body of research has been done in this field. A comprehensive study on the different server consolidation solutions has not been done yet, though. In this study, a systematic review has been done on a set of researches related to server consolidation. After investigating the considered researches, their proposed solutions were categorized into three groups based on the type of decision making for running the consolidation process. Groups involve static method, dynamic method (including threshold-based and periodic-based adaptation) and prediction-based dynamic method. Thereafter, we discussed handling the challenges presented in each research by investigating the proposed approach for developing consolidation technique. Then, the open issues in each study were expressed. Finally, the objectives, evaluation parameters, optimization methods and the affecting parameters of server consolidation in all studies were investigated and analyzed. [ABSTRACT FROM AUTHOR]

Software-defined network (SDN) can ease the implementation of QoS concept by introducing a flexible and manageable mechanism to overcome the TCP pacing in the network and data centers which have burst traffics. SDN's flexible structure lets the network manager control the resources while the traffic condition is changed. Therefore, it is possible to adopt an application flow rate based on its dynamic parameters such as delay and throughput. This paper studies a new dynamic QoS mechanism for SDNs. This mechanism improves the performance of TCP-based network traffic in the per-node model, compared with the classical end-to-end approach. In overall, we focused on two ideas: First, we alter the end-to-end holistic view of the TCP on per-node behavior. Second, we take advantage of a two-segment scheduler model. As a result, we are aware of congestion in the entire network nodes and are able to make a reaction to obtain the requested QoS. Simulation results show the effectiveness of the developed model. [ABSTRACT FROM AUTHOR]

Various topologies have been proposed for high-performance computing (HPC), i.e., fat-tree, Torus topology. Compared with conventional fat-tree topology, Torus performs much better when applied in HPC. Unfortunately, due to its wraparound links, Torus topology naturally has the tendency to trigger deadlock incidents inside the network. Researchers solve this problem by means of virtual channel, but this approach will also restrict the routing of message. In this paper, we propose a deadlock-free topology for HPC, called Mesh-of-Torus, which incarnates the good characteristics of Mesh and Torus topology. Comparing with mesh, Mesh-of-Torus has shorter network diameter. Furthermore, we have proposed a corresponding port assignment rules in consideration of complicated internal arbitration or scheduling mechanism incurred by the employment of virtual channel. Deadlock avoidance can be achieved when dimension-order routing algorithm and our port assignment rules are applied to Mesh-of-Torus. Finally, simulations and mathematical analysis have shown that Mesh-of-Torus outperforms Mesh in terms of average end-to-end latency and network load distribution. [ABSTRACT FROM AUTHOR]

Modern data center consists of thousands of servers, racks and switches. Complicated structure means it requires well-designed algorithms to utilize resources of data centers efficiently. Current virtual machine scheduling algorithms mainly focus on the initial allocation of virtual machines based on the CPU, memory and network bandwidth requirements. However, when tasks finished or lease expired, related virtual machines would be deleted from the system which would generate resource fragments. Such fragments lead to unbalanced resource utilization and decline of communication performance. This paper investigates the network influence on typical applications in data centers and proposed a self-adaptive network-aware virtual machine clustering and consolidation algorithm to maintain an optimal system-wide status. Our consolidation algorithm periodically checks whether consolidation is necessary and then clusters and consolidates virtual machines to lower communication cost with an online heuristic. We used two benchmarks in a real environment to examine network influence on different tasks. To evaluate the advantages of the proposed algorithm, we also built a cloud computing testbed. Real workload trace-driven simulations and testbed-based experiments showed that, our algorithm greatly shortened the average finish time of map-reduce tasks and reduced time delay of web applications. Simulation results showed that our algorithm considerably reduced the amount of high-delay jobs, lowered the average traffic passed through aggregate switches and improved the communication ability among virtual machines. [ABSTRACT FROM AUTHOR]

Virtual machine (VM) migration is a process of migrating VMs from one physical server to another. It provides several benefits to a data center in a variety of scenarios including improved performance, fault tolerance, manageability load balancing and power management. However, VM’s migration leads to performance degradation and service-level agreement (SLA) violations which cannot be ignored, particularly if critical business goals are to be met. In this paper, we propose an algorithm for VM’s placement and migration that considers different users quality of service requirement, in order to decrease energy consumption and SLA violations due to under utilization of data centers. The proposed work mainly focuses on a novel heuristics-based energy-aware resource allocation to allocate the user’s tasks in the form of cloudlets to the cloud resources that consumes minimal energy. In addition to that, it is incorporated with load balancing and constraint-based scheduling mechanism. The proposed work is implemented using the service-oriented-based architecture, and the same has been simulated using the CloudSim toolkit. In this paper, we compared our work with non-power-aware (NPA), dynamic voltage and frequency scaling (DVFS), single-threshold (ST) policies and minimization migration policy (MMP). The experiment results indicate that our approach saves about 83% of power comparing to the NPA system and 77% comparing to a system that apply only DVFS. However, if we compare these algorithms, which allow dynamic consolidation of VMs such as ST, it saves 53%, and finally, if we compare to MMP, it saves power between 22 and 38%. Similarly if we compare number of VM migration comparing to ST, it reduces 23 and 73% compared to MMP polices. [ABSTRACT FROM AUTHOR]

This paper focuses on energy consumption which is a major problem in the dark silicon era. As energy consumption becomes a key issue for operation and maintenance of cloud data centers, cloud computing providers are becoming significantly concerned. Here, we show how spin-transfer torque random access memory (STT-RAM) can be used as an on-chip L2 cache to obtain lower energy compared to conventional L2 caches, like SRAM. High density, fast read access and non-volatility make STT-RAM a significant technology for on-chip memories. Previous studies have mainly studied specific schemes based on common applications and do not provide a thorough analysis of emerging scale-out applications with multiple design options. Here, we discuss different outlooks consisting of performance and energy efficiency in cloud processors by running emerging scale-out workloads. Experiment results on the CloudSuite benchmarks show that the proposed method reduces energy by 51% (on average) and improves energy delay product by 37% (on average) where instruction per cycle degradation is only 22% (on average) compared to the SRAM method. [ABSTRACT FROM AUTHOR]

Data center network (DCN) is the core of cloud computing and accounts for 40% energy spend when compared to cooling system, power distribution and conversion of the whole data center (DC) facility. It is essential to reduce the energy consumption of DCN to ensure energy-efficient (green) data center can be achieved. An analysis of DC performance and efficiency emphasizing the effect of bandwidth provisioning and throughput on energy proportionality of two most common switch-centric DCN topologies: three-tier (3T) and fat tree (FT) based on the amount of actual energy that is turned into computing power are presented. Energy consumption of switch-centric DCNs by realistic simulations is analyzed using GreenCloud simulator. Power-related metrics were derived and adapted for the information technology equipment processes within the DCN. These metrics are acknowledged as subset of the major metrics of power usage effectiveness and data center infrastructure efficiency, known to DCs. This study suggests that although in overall FT consumes more energy, it spends less energy for transmission of a single bit of information, outperforming 3T. [ABSTRACT FROM AUTHOR]

Energy consumption has become a critical design factor in today's data centers. In recent years, extensive research has been done to address power-performance trade-off in data centers considering both IT equipments and cooling infrastructures (e.g., thermal-aware task scheduling, server consolidation, load balancing and geographical load balancing to name a few). This paper introduces a design-time technique that targets energy-efficient design of a green data center farm in Iran. Workload predictions, geographical maps of wind speed and solar radiation, data center and renewable resources configurations are used as a priori to design an energy-efficient data center farm for Internet services. The proposed problem is mathematically formulated as a nonlinear optimization problem and is effectively solved using a coordinate descent-based method. We also show that with some minor modification, our proposed technique can be applied at run-time for the purpose of change management. The experimental results show that the proposed method can lead to 11.6 % cost saving on average over conventional approaches. [ABSTRACT FROM AUTHOR]

Increasing demand for acquiring diverse range of services has led to the establishment of huge energy hungry cloud data centers all around the world. Cloud providers face with major concerns to reduce their energy consumption while ensuring high quality of service based on the Service Level Agreement (SLA). Consolidation is proposed as one of the most effective techniques for online energy saving in cloud environments with dynamic workloads. This paper proposes novel proactive online resource management policies to optimize energy, SLA, and number of migrations in cloud data centers. More precisely, this paper proposes new prediction algorithm for determination of overloaded hosts as well as novel multi-criteria decision making techniques to select virtual machines. The results of simulations using CloudSim simulator shows up to 98.11 % reduction in the output metric which is representative of energy consumption, SLA violation, and number of migrations, in comparison with state of the art. [ABSTRACT FROM AUTHOR]

Thermal cameras provide fine-grained thermal information that enables monitoring and autonomic thermal management in large datacenters. The real-time thermal monitor network employing thermal cameras is proposed to cooperatively localize hotspots and extract their characteristics (i.e., temperature, size, and shape). These characteristics are adopted to classify the causes of hotspots and make energy-efficient thermal management decisions such as job migration. Specifically, a sculpturing algorithm for extracting and reconstructing shape characteristics of hotspots is proposed to minimize the network overhead. Experimental results show the validity of all the algorithms proposed in this paper. [ABSTRACT FROM AUTHOR]