Today's Data Centers Research Articles

Intersection of the Cloud and Big Data

The two biggest trends in the data center today are cloud computing and big data. This column will examine the intersection of the two. Industry hype has resulted in nebulous definitions for each, so I'll start by defining terms.

Rethinking the Data Center Networking: Architecture, Network Protocols, and Resource Sharing

Large-scale data centers enable the new era of cloud computing and provide the core infrastructure to meet the computing and storage requirements for both enterprise information technology needs and cloud-based services. To support the ever-growing cloud computing needs, the number of servers in today's data centers are increasing exponentially, which in turn leads to enormous challenges in designing an efficient and cost-effective data center network. With data availability and security at stake, the issues with data center networks are more critical than ever. Motivated by these challenges and critical issues, many novel and creative research works have been proposed in recent years. In this paper, we investigate in data center networks and provide a general overview and analysis of the literature covering various research areas, including data center network interconnection architectures, network protocols for data center networks, and network resource sharing in multitenant cloud data centers. We start with an overview on data center networks and together with its requirements navigate the data center network designs. We then present the research literature related to the aforementioned research topics in the subsequent sections. Finally, we draw the conclusions.

Sharing Bandwidth by Allocating Switch Buffer in Data Center Networks

In today's data centers, the round trip propagation delay is quite small. Therefore, switch buffer sizes are much larger than the Bandwidth Delay Product (BDP). Based on this observation, in this paper we introduce a new transport protocol which provides bandwidth Sharing by Allocating switch Buffer (SAB) for data centers. SAB sets the congestion windows for flows based on the buffer size of the switches along the path. On one hand, as long as the total buffer allocated to all the flows is larger than the BDP, the network bandwidth can be fully utilized. On the other hand, since SAB only allocates the buffer space to flows, the totally injected traffic will not exceed the network capacity. Thus, SAB rarely loses packets. SAB also reduces flow completion time by allowing flows to reach their fair share of bandwidth quickly. The results of a series of experiments and simulations demonstrate that SAB has the features of fast convergence and rare packet loss. It reduces the latency of short flows and solves theTCP Incast and TCP Outcast problems.

Don't drop, detour!

Today's data centers must support a range of workloads with different demands. While existing approaches handle routine traffic smoothly, ephemeral but intense hotspots cause excessive packet loss and severely degrade performance. This loss occurs even though the congestion is typically highly localized, with spare buffer capacity available at nearby switches. We argue that switches should share buffer capacity to effectively handle this spot congestion without the latency or monetary hit of deploying large buffers at individual switches. We present detour-induced buffer sharing (DIBS), a mechanism that achieves a near lossless network without requiring additional buffers. Using DIBS, a congested switch detours packets randomly to neighboring switches to avoid dropping the packets. We implement DIBS in hardware, on software routers in a testbed, and in simulation, and we demonstrate that it reduces the 99th percentile of query completion time by 85%, with very little impact on background traffic.

Expandable and Cost-Effective Network Structures for Data Centers Using Dual-Port Servers

A fundamental goal of data center networking is to efficiently interconnect a large number of servers with the low equipment cost. Several server-centric network structures for data centers have been proposed. They, however, are not truly expandable and suffer a low degree of regularity and symmetry. Inspired by the commodity servers in today's data centers that come with dual port, we consider how to build expandable and cost-effective structures without expensive high-end switches and additional hardware on servers except the two NIC ports. In this paper, two such network structures, called HCN and BCN, are designed, both of which are of server degree 2. We also develop the low overhead and robust routing mechanisms for HCN and BCN. Although the server degree is only 2, HCN can be expanded very easily to encompass hundreds of thousands servers with the low diameter and high bisection width. Additionally, HCN offers a high degree of regularity, scalability, and symmetry, which conform to the modular designs of data centers. BCN is the largest known network structure for data centers with the server degree 2 and network diameter 7. Furthermore, BCN has many attractive features, including the low diameter, high bisection width, large number of node-disjoint paths for the one-to-one traffic, and good fault-tolerant ability. Mathematical analysis and comprehensive simulations show that HCN and BCN possess excellent topological properties and are viable network structures for data centers.

Who Is Your Neighbor: Net I/O Performance Interference in Virtualized Clouds

User-perceived performance continues to be the most important QoS indicator in cloud-based data centers today. Effective allocation of virtual machines (VMs) to handle both CPU intensive and I/O intensive workloads is a crucial performance management capability in virtualized clouds. Although a fair amount of researches have dedicated to measuring and scheduling jobs among VMs, there still lacks of in-depth understanding of performance factors that impact the efficiency and effectiveness of resource multiplexing and scheduling among VMs. In this paper, we present the experimental research on performance interference in parallel processing of CPU-intensive and network-intensive workloads on Xen virtual machine monitor (VMM). Based on our study, we conclude with five key findings which are critical for effective performance management and tuning in virtualized clouds. First, colocating network-intensive workloads in isolated VMs incurs high overheads of switches and events in Dom0 and VMM. Second, colocating CPU-intensive workloads in isolated VMs incurs high CPU contention due to fast I/O processing in I/O channel. Third, running CPU-intensive and network-intensive workloads in conjunction incurs the least resource contention, delivering higher aggregate performance. Fourth, performance of network-intensive workload is insensitive to CPU assignment among VMs, whereas adaptive CPU assignment among VMs is critical to CPU-intensive workload. The more CPUs pinned on Dom0 the worse performance is achieved by CPU-intensive workload. Last, due to fast I/O processing in I/O channel, limitation on grant table is a potential bottleneck in Xen. We argue that identifying the factors that impact the total demand of exchanged memory pages is important to the in-depth understanding of interference costs in Dom0 and VMM.

Maximizing the detection probability of overheating server components with sensor placement based on thermal dynamics

Server overheating has become a well-known issue in today's data centers that host a large number of high-density servers. The current practice of server overheating detection is to monitor the server inlet temperature with the temperature sensor on the server enclosure, or the CPU temperature with on-die thermal sensors. However, this is in contrast to the fact that different components in a server may have different overheating thresholds, which are closely related to their respective thermal failure rates and expected lifetimes. Moreover, the thermal correlation between the inlet (or CPU) and other server components can be different for every server model. As a result, relying on the single inlet or CPU temperature for server overheating detection is over-simplistic, which may lead to either degraded detection performance or false alarms that can result in excessive cooling power, leading to unnecessarily low inlet temperature.In this paper, we propose a model-based approach that leverages thermal dynamics to intelligently choose sensor placement locations for precise overheating server component detection. We first formulate the detection problem as a constrained optimization problem. We then adopt Computational Fluid Dynamics (CFD) to establish the thermal model and analyze the thermal status of the server enclosure under various overheating conditions, such as inlet overheating, fan failures and CPU overloading. Based on the CFD analysis, we apply data fusion and advanced optimization techniques to find a near-optimal solution for sensor placement locations, such that the probability of detecting different overheating components is significantly improved. Our empirical results on a real rack server testbed demonstrate the detection performance of our solution. Extensive simulation results also show that the proposed solution outperforms other commonly used overheating monitoring solutions in terms of detection probability and error rate.

Towards full network virtualization in horizontal IaaS federation: security issues

Horizontal IaaS federation exploits datacenter for federation of IaaS provider by supplying virtual nodes (e.g. virtual machines, virtual switches, and virtual routers) and virtual links. Today's datacenters for cloud computing do not supply full network virtualization in terms of user-level network management and user-agreed network topology. The datacenters lack the basic security services required for the collocation of tenants' virtual networks. The network virtualization research projects from academia and industry support full network virtualization but lack the basic security services required for the collocation of tenants' virtual networks. This paper investigates the security issues in four areas namely, (a) monolithic IaaS cloud, (b) network virtualization research projects, (c) datacenter network virtualization and (d) virtual resources to incorporate full network virtualization environment in horizontal IaaS federation. Further, it presents the security related qualitative comparisons of datacenters, network virtualization research projects and virtual resources to incorporate full network virtualization in horizontal IaaS federation.

Responding rapidly to service level violations using virtual appliances

One of the key goals in the data center today is providing storage services with service-level objectives (SLOs) for performance metrics such as latency and throughput. Meeting such SLOs is challenging due to the dynamism observed in these environments. In this position paper, we propose dynamic instantiation of virtual appliances, that is, virtual machines with storage functionality, as a mechanism to meet storage SLOs efficiently. In order for dynamic instantiation to be realistic for rapidlychanging environments, it should be automated. Therefore, an important goal of this paper is to show that such automation is feasible. We do so through a caching case study. Specifically, we build the automation framework for dynamically instantiating virtual caching appliances. This framework identifies sets of interfering workloads that can benefit from caching, determines the cache-size requirements of workloads, non-disruptively migrates the application to use the cache, and warms the cache to quickly return to acceptable service levels. We show through an experiment that this approach addresses SLO violations while using resources efficiently.

Network Traffic Analysis with Cloud Platform

High-speed network poses challenges on heavy traffic measurement and analysis, which becomes more serious due to the scale of modern data centers. Cloud platform is an alternative to crack the nut because it gives concurrent considerations to both high-performance computing and high-extensible data storage. Therefore, we use Hadoop to construct a trial cloud platform to tackle the challenges derived from the heavy traffic, and perform network traffic analysis and target packet dissection. Our testbed evaluates the performance against different processing strategies, and shows the statistic results. According to the experimental results, the Hadoop-based platform can fit to process massive traffic data very well, and it's appropriate to manage today's data centers with cloud computing services.

Finishing flows quickly with preemptive scheduling

Today's data centers face extreme challenges in providing low latency. However, fair sharing, a principle commonly adopted in current congestion control protocols, is far from optimal for satisfying latency requirements. We propose Preemptive Distributed Quick ( PDQ ) flow scheduling, a protocol designed to complete flows quickly and meet flow deadlines. PDQ enables flow preemption to approximate a range of scheduling disciplines. For example, PDQ can emulate a shortest job first algorithm to give priority to the short flows by pausing the contending flows. PDQ borrows ideas from centralized scheduling disciplines and implements them in a fully distributed manner, making it scalable to today's data centers. Further, we develop a multipath version of PDQ to exploit path diversity. Through extensive packet-level and flow-level simulation, we demonstrate that PDQ significantly outperforms TCP, RCP and D 3 in data center environments. We further show that PDQ is stable, resilient to packet loss, and preserves nearly all its performance gains even given inaccurate flow information.

Improving resource efficiency in data centers using reputation-based resource selection

Today data centers are consuming a lot of energy but not very efficiently. Much of energy is wasted. There are several types of energy waste at different levels including infrastructure-, machine- and system-level waste. The former two levels have been improved significantly in the last few years, however, few efforts have been put on the last level, especially the resource waste caused by failures in a data center. In this paper, we attack the problem proactively by leveraging a reputation-based resource selection scheme to reduce the number of resubmissions of tasks, resulting from the failure during the course of their execution. To capture the characteristics of resources, we introduce Opera, an OPEn ReputAtion model. Opera characterizes itself with two important novelties: a vector representation of the reputation and the just-in-time feature that represents the real-time system status, which, to our knowledge, has never been considered in conventional reputation systems. To demonstrate the effectiveness of Opera, we have integrated the Opera trust model into the scheduler of Hadoop. The experimental results showed that Opera enables the scheduler to select appropriate nodes which helped to reduce not only the number of re-executed tasks but also the execution time of Hadoop's jobs under the presence of failures and heavy workload up to 59% and 32%, respectively. This improvements, in turn, can improve the energy efficiency of the whole system and the network by 16.17% and 53.32% for the sort application respectively.

Minimizing the thermal impact of computing equipment upgrades in data centers

Upgrading of today's air-cooled data centers (DCs) with high-performance computing, networking, and storage equipment is a challenging task due to typically higher power needs and more adverse cooling requirements of new equipment. To cope with the increase of the power and heat load in DCs, load spreading is commonly applied. This technique distributes the thermal load of new equipment over multiple racks if the power requirement and heat generation of new equipment exceeds the rack's capacity. We present a novel load spreading technique that allows upgrading of the computing equipment with minimal thermal impact on the existing optimized DC cooling environment. Our approach is based on an abstract heat-flow model of the DC, whose parameters are determined by performing a measurement campaign in the DC and with support of computational fluid dynamics simulations. The optimum placement of the new equipment in the racks of the DC is found by applying a particle swarm optimization technique to this model. The effectiveness of our method was assessed based on experiments performed in a production DC. The results show that our holistic approach for optimizing the placement of the upgraded computing equipment in the DC outperforms the conventional load spreading technique.

Cosmic rays don't strike twice

Main memory is one of the leading hardware causes for machine crashes in today's datacenters. Designing, evaluating and modeling systems that are resilient against memory errors requires a good understanding of the underlying characteristics of errors in DRAM in the field. While there have recently been a few first studies on DRAM errors in production systems, these have been too limited in either the size of the data set or the granularity of the data to conclusively answer many of the open questions on DRAM errors. Such questions include, for example, the prevalence of soft errors compared to hard errors, or the analysis of typical patterns of hard errors. In this paper, we study data on DRAM errors collected on a diverse range of production systems in total covering nearly 300 terabyte-years of main memory. As a first contribution, we provide a detailed analytical study of DRAM error characteristics, including both hard and soft errors. We find that a large fraction of DRAM errors in the field can be attributed to hard errors and we provide a detailed analytical study of their characteristics. As a second contribution, the paper uses the results from the measurement study to identify a number of promising directions for designing more resilient systems and evaluates the potential of different protection mechanisms in the light of realistic error patterns. One of our findings is that simple page retirement policies might be able to mask a large number of DRAM errors in production systems, while sacrificing only a negligible fraction of the total DRAM in the system.

Cosmic rays don't strike twice

Main memory is one of the leading hardware causes for machine crashes in today's datacenters. Designing, evaluating and modeling systems that are resilient against memory errors requires a good understanding of the underlying characteristics of errors in DRAM in the field. While there have recently been a few first studies on DRAM errors in production systems, these have been too limited in either the size of the data set or the granularity of the data to conclusively answer many of the open questions on DRAM errors. Such questions include, for example, the prevalence of soft errors compared to hard errors, or the analysis of typical patterns of hard errors. In this paper, we study data on DRAM errors collected on a diverse range of production systems in total covering nearly 300 terabyte-years of main memory. As a first contribution, we provide a detailed analytical study of DRAM error characteristics, including both hard and soft errors. We find that a large fraction of DRAM errors in the field can be attributed to hard errors and we provide a detailed analytical study of their characteristics. As a second contribution, the paper uses the results from the measurement study to identify a number of promising directions for designing more resilient systems and evaluates the potential of different protection mechanisms in the light of realistic error patterns. One of our findings is that simple page retirement policies might be able to mask a large number of DRAM errors in production systems, while sacrificing only a negligible fraction of the total DRAM in the system.

Electric grid balancing through lowcost workload migration

Energy production must continuously match demand on the electric grid. A deficiency can lead to service disruptions, and a surplus can place tremendous stress on grid components, potentially causing major blackouts. To manage this balance, grid operators must increase or lower power generation, with only a few minutes to react. The grid balancing problem has also impeded the pace of integrating bountiful renewable resources (e.g., wind), whose generation is intermittent. An emerging plan to mitigate this problem is demand response, i.e., for grid operators to alter the electricity usage behavior of the masses through real-time price signals. But due to prohibitively high infrastructure costs and societal-scale adoption, tangible demand response mechanisms have so far been elusive. We believe that altering the usage patterns of a multitude of data centers can be a tangible, albeit initial, step towards affecting demand response. Growing in both density and size, today's data center designs are shaped by the increasing awareness of energy costs and carbon footprint. We posit that shifting computational workloads (and thus, demand) across geographic regions to match electricity supply may help balance the grid. In this paper we will first present a real grid balancing problem experienced in the Pacfic Northwest. We then propose a symbiotic relationship between data centers and grid operators by showing that mutual cost benefits can be accessible. Finally, we argue for a low cost workload migration mechanism, and pose overarching challenges in designing this framework.

SHIP: A Scalable Hierarchical Power Control Architecture for Large-Scale Data Centers

In today's data centers, precisely controlling server power consumption is an essential way to avoid system failures caused by power capacity overload or overheating due to increasingly high server density. While various power control strategies have been recently proposed, existing solutions are not scalable to control the power consumption of an entire large-scale data center, because these solutions are designed only for a single server or a rack enclosure. In a modern data center, however, power control needs to be enforced at three levels: rack enclosure, power distribution unit, and the entire data center, due to the physical and contractual power limits at each level. This paper presents SHIP, a highly scalable hierarchical power control architecture for large-scale data centers. SHIP is designed based on well-established control theory for analytical assurance of control accuracy and system stability. Empirical results on a physical testbed show that our control solution can provide precise power control, as well as power differentiations for optimized system performance and desired server priorities. In addition, our extensive simulation results based on a real trace file demonstrate the efficacy of our control solution in large-scale data centers composed of 5,415 servers.

Virt-LM

Virtualization technology has been widely applied in data centers and IT infrastructures, with advantages of server consolidation and live migration. Through live migration, data centers could flexibly move virtual machines among different physical machines to balance workloads, reduce energy consumption and enhance service availability. Today's data centers can grow to a huge scale. This implies that frequent live migration would be desirable for the economic use of hardware resources. Then, the performance of the live migration strategy will be an issue. So, we need a reliant evaluation method to choose the software and hardware environments that will produce the best live migration performance. However, there is not a complete live migration benchmark available currently. In addition, the existing evaluation methodologies select different metrics, different workloads and different test means. Thus, it is difficult to compare their results. In this paper we first survey the current research and their evaluation methods on live migration. We then summarize the critical issues for the live migration evaluation and also raise other unreported potential problems. We propose our solutions and present an implementation in our live migration benchmark -- Virt-LM. This is a benchmark for comparing live migration performance among different software and hardware environments in a data center scenario. We detail its design and provide some experimental results to validate its effectiveness.

Better never than late

The soft real-time nature of large scale web applications in today's datacenters, combined with their distributed workflow, leads to deadlines being associated with the datacenter application traffic. A network flow is useful, and contributes to application throughput and operator revenue if, and only if, it completes within its deadline. Today's transport pro- tocols (TCP included), given their Internet origins, are agnostic to such flow deadlines. Instead, they strive to share network resources fairly. We show that this can hurt application performance. Motivated by these observations, and other (previously known) deficiencies of TCP in the datacenter environment, this paper presents the design and implementation of D3, a deadline-aware control protocol that is customized for the datacenter environment. D3 uses explicit rate control to apportion bandwidth according to flow deadlines. Evaluation from a 19-node, two-tier datacenter testbed shows that D3, even without any deadline information, easily outper- forms TCP in terms of short flow latency and burst tolerance. Further, by utilizing deadline information, D3 effectively doubles the peak load that the datacenter network cansupport.

Scalable and Cost-Effective Interconnection of Data-Center Servers Using Dual Server Ports

The goal of data-center networking is to interconnect a large number of server machines with low equipment cost while providing high network capacity and high bisection width. It is well understood that the current practice where servers are connected by a tree hierarchy of network switches cannot meet these requirements. In this paper, we explore a new server-interconnection structure. We observe that the commodity server machines used in today's data centers usually come with two built-in Ethernet ports, one for network connection and the other left for backup purposes. We believe that if both ports are actively used in network connections, we can build a scalable, cost-effective interconnection structure without either the expensive higher-level large switches or any additional hardware on servers. We design such a networking structure called FiConn. Although the server node degree is only 2 in this structure, we have proven that FiConn is highly scalable to encompass hundreds of thousands of servers with low diameter and high bisection width. We have developed a low-overhead traffic-aware routing mechanism to improve effective link utilization based on dynamic traffic state. We have also proposed how to incrementally deploy FiConn.