(https://docslide.net/register.html)
(https://docslide.net/)
HOME (HTTPS://DOCSLIDE.NET/) LEADERSHIP (HTTPS://DOCSLIDE.NET/CATEGORY/LEADERSHIP-MANAGEMENT.HTML) TECHNOLOGY (HTTPS://DOCSLIDE.NET/CATEGORY/TECHNOLOGY.HTML) EDUCATION (HTTPS://DOCSLIDE.NET/CATEGORY/EDUCATION.HTML) MORE TOPICS (HTTPS://DOCSLIDE.NET/CATEGORY.HTML)
/
Home (https://docslide.net/) / Documents (https://docslide.net/category/documents.html) [IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.232013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - EDR: An energy-aware runtime load distribution system for data-intensive applications in the cloud (https://docslide.net/documents/ieee-2013-ieee-internationalconference-on-cluster-computing-cluster-indianapolis-58c19be5948d4.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - EDR: An energy-aware runtime load distribution system for data-intensive applications in the cloud REPORT (HTTPS://DOCSLIDE.NET/REPORT-COPYRIGHT/IEEE-2013-IEEE-INTERNATIONAL-CONFERENCE-ONDocuments (https://docslide.net/category/documents.html)
CLUSTER-COMPUTING-CLUSTER-INDIANAPOLIS-58C19BE5948D4/5750A5B21A28ABCF0CB3E6AD) Please download to view (https://docslide.net/download/link/ieee-2013-ieee-international-conference-on-cluster-computing-
RECOMMENDED
cluster-indianapolis-58c19be5948d4) EDR: An Energy-Aware Runtime Load Distribution System for Data-Intensive Applications in the Cloud Bo Li Virginia Tech
[email protected] Shuaiwen Leon Song Pacific Northwest National Lab
[email protected] Abstract-Data centers account for a growing percentage of US power consumption. Energy efficiency is now a first-class design constraint for the data centers that support cloud services. Service providers must distribute their data efficiently across multiple data centers. This includes creation of data replicas that provide multiple copies of data for efficient access. However, selecting replicas to maximize performance while minimizing energy waste is an open problem. State of the art replica selection approaches either do not address energy, lack scalability and/or are vulnerable to crashes due to use of a centralized coordinator. Therefore, we propose, develop and evaluate a simple cost-oriented decentralized replica selection system named EDR (Energy-Aware Distributed Running system), implemented with two distributed optimization algorithms. We demonstrate experimentally the cost differences in various replica selection scenarios and show that our novel approach is as fast as the best available decentralized approach DONAR,while additionally considering dynamic energy costs. We show that an average of 12% savings on total system energy costs can be achieved by using EDR for several data intensive applications. I. INTRODUCTION Cloud computing is built upon an infrastructure of geo graphically distributed data centers throughout the world. The most common way to protect against data loss in the cloud is through replication - or maintaining multiple copies of data [1]. When a user needs access to their data in the cloud, the system responds by selecting a copy of the data from various geographically distributed locations and providing the data to the end user. During this process (called replica selection), the system selects the copy of the data that it believes will result in the lowest latency (or fastest data transfer), least packet loss, etc. Service providers can use replica selection to distribute load generated by user requests, thus globally optimizing the use of available resources. Replica selection requires access to a coordinator that keeps track of where account data resides globally. As expected, for large systems with a large number of users, the data needed by a coordinator can grow unwieldy. Additionally, as data is moved dynamically at any time, the coordinator data must be kept up to date continuously. Replica selection coordinators are typically implemented using either a centralized or a distributed approach. Central ized coordinators are typically simpler and often faster than distributed implementations but suffer from a single point of failure [2] and poor scalability. Distributed coordinators 978-1-4799-0898-1/13/$31.00 ©2013 IEEE Ivona Bezakova Rochester Institute of Technology
[email protected] Kirk W. Cameron Virginia Tech
[email protected] are typically more complicated while potentially slower than centralized implementations but they can be more scalable [3]. Current replica selection implementations focus on opti mizing for bandwidth and latency and do not consider the cost of the power necessary to access a replica. However, according to the prediction from the 2007 EPA report [4], the energy used to power data centers will likely exceed 3% of the total US energy use by 2013. Therefore, power is now the biggest cost item in a data center [5][6] and the costs per kWh also vary widely by region globally [7]. Hence, all data does not cost the same from the data center operators' (Le. service providers) perspective. Substantial research has been conducted to optimize
(https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - AptStore: Dynamic storage management for hadoop (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis.html) Documents
(https://docslide.net/category/documents.html) (https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-58b1b797c86db.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - JUMMP: Job Uninterrupted Maneuverable MapReduce Platform (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-58b1b797c86db.html) Documents
(https://docslide.net/category/documents.html) (https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-589c00af97357.html)
can be achieved using our methods for the data intensive applications such as online video streaming and distributed file
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Modeldriven multisite workflow scheduling (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-589c00af97357.html)
sharing. The rest of this paper is organized as follows. In Section II, we discuss related work. Then, we present our
Documents
these data centers both locally [8] and globally [9] in attempts to reduce the number of machines powered on 24x7. To reduce costs globally, future data center operators will require replica selection that considers not only bandwidth capacity and network latency but also the dollar cost (e.g. power) of accessing to different replicas. In this paper, we propose a decentralized replica selection system named EDR, considering data transfer under varied regional power costs, bandwidth capacity and network latency of the data centers. We demonstrate experimentally the cost differences in various replica selection scenarios and show that our novel approach performs as well as the best available decentralized approach "DONAR" [3] while additionally considering dynamic energy costs. In particular, we show how an average of 12% savings
methodology on solving the energy-aware replica selection problem in Section III. After that, Section IV shows the evaluation of performance as well as the energy cost reduction of our EDR framework. Finally, in Section V, we conclude with a brief summary and describe our future work. II. RELATED WORK A. Energy Efficiency in a Data Center Energy efficiency in the cloud has been studied as an important issue by others. The effort of reducing energy cost has been taken through hardware, software, as well as network ing aspects [10]. For example, resource allocation [11] and scheduling algorithms considering QoS [12][13] can improve energy efficiency of data center as well as guarantee the quality of services. However, such work does not investigate the workload distribution of client requests among all the data centers, which can also affect the total energy cost even if the resources have already been optimally allocated. In particular, for the data-intensive services, the distribution of workload to each data center can significantly affect the energy cost. Zong et al. [14] apply a buffer-disk to schedule storage system tasks, so that energy consumption can be reduced by keeping a large number of data disks idle. But they only consider the relationship between disk state and power consumption regardless of different workload types for the disk. Kim et al.[15] try to reduce data center energy by introducing low power devices. But they do not consider the electricity prices at different geographical locations. Rao et al. [9] consider mul tiple electricity prices into their energy model for data centers. However, they do not consider the bandwidth capacity and do not provide a decentralized solution to their optimization model. In order to minimize the energy cost of data centers, Liu et al. [16] take workload and number of active servers in each data center into consideration. However, they assume that the single server energy consumption does not depend on the traffic load, which is not practical for modeling data-intensive applications in cloud. Smith and Sommerville [17] studied the impacts of vari ous types of applications on different subcomponents' energy consumption within a server. Furthermore, a linear relationship between data-intensive workload and energy consumption for hard disks in server systems is validated in work [18]. How ever, we cannot make the assumption that a linear relationship exists between workload and the energy consumption of other components in data centers. For instance, majority of the network devices are far from being energy proportional [19], [20]. B. Replica Selection in Cloud Ruiz-Alvarez and Humphrey [21] present an approach for selecting storage services in the cloud. However, this approach has some security issues for accessing multiple data centers at runtime so it is hard to be used for solving replica selection problems. Le et al. [22], Chen et al. [23], and Rajamani et al. [24] have studied the load distribution problem in the clusters and data centers. However, they have not considered the decentralized system architecture. Work [3] and [25] present decentralized replica selection systems for load distribution. The systems are validated to be working decently under a wild network environment. However, energy efficiency issues are not considered in their frameworks. III. METHODOLOGY In this section, we first present a simple energy cost model for data centers used by our runtime scheduler. Based on this model, we then formulate the replica selection problem as a convex optimization problem which minimizes total energy cost of data centers subject to bandwidth capacity and network latency of each data center. After that, we propose a simple de centralized framework named EDR where the distributed nodes cooperate with each other to solve the global optimization problem in parallel. Finally, we adapt two parallel algorithms into EDR and give a detailed analysis on their individual communication complexity and theoretical convergence rate. Some important modelrelated parameters in this section are summarized in Table I and mapped to the system service architecture in Fig. 1. TABLE I. C Set of all clients N Set of all replicas NOTATIONS Eg Total energy consumption of all the replicas in the cloud Ell Total energy consumption for replica n Pc.n Traffic load mapped from client c to replica n Pn Constraint sets on replica n Bn Bandwidth capacity on replica /1 T User·defined max tolerable network latency Rc Traffic load of the request from client c lell Network latency from client c to replica n lin Unit price (¢) of power in replica n all Weight value of replica n in consensus-based algorithm cyn,j311 Weight scalars for the energy consumption of servers and network devices in repl ica n 'Yn Parameter to correlate traffic load to network devices' energy consumption for replica n. y,/s value depends on the underlying device's architecture A. Energy Cost Model for Data Center In order to model the energy cost of data-intensive
(https://docslide.net/category/documents.html) (https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-58f3b2037af91.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Userfriendly metaworkflows in quantum chemistry (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-58f3b2037af91.html) Documents
(https://docslide.net/category/documents.html) (https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-58c1a379c02d3.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Making work queue cluster-friendly for data intensive scientific applications (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-58c1a379c02d3.html) Documents
(https://docslide.net/category/documents.html) (https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-589f15f56bba8.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Thermal aware automated load balancing for HPC applications (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-589f15f56bba8.html) Documents
(https://docslide.net/category/documents.html)
appli cations in the cloud, we build a correlation between energy consumption and the workload from the clients. Since the primary goal of the model is to help clients make runtime decision on how to efficiently distribute workload among replicas, we want to keep the model simple and only reflect the major energy consuming components which are significantly impacted by the traffic loads. Therefore, we make the following assumptions for modeling data center energy consumption: 1). From a system component-level modeling perspective, it can be nontrivial to describe the total energy consumption of a single server node, which normally consists of CPU, memory, hard disk, motherboard, accelerators, system and CPU fans, etc. Detailed system-level modeling has been conducted in work such as [26][27], which often involves modeling on off chip time and power affected by workload and hardware frequency. In our case, since we are primarily targeting the data and network intensive applications, we can assume they provide relatively consistent workload intensity for individual servers during each period of execution time. Thus, we can assume the power consumption for each replica is constant and the time spent on each replica is linear to the workload. Consequently, it is reasonable to make the assumption that the relationship between energy consumption of each replica and its workload is Ii near as well. 2). Network devices' energy is also a significant contributor to the overall data center energy consumption. It depends on several factors, including traffic load, temperature, quality of service (QoS) policies and floor space. In this work, we only consider the most influential factor- traffic load [28]. Restrepo et al.[29] have studied data centers' network energy profiles and categorized the
(https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-58d3e5c4dc162.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - FlexQuery: An online query system for interactive remote visual data exploration at large scale (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-58d3e5c4dc162.html) Documents
(https://docslide.net/category/documents.html)
relationships between network energy and traffic loads. For instance, switch architectures such as Batcher and Crossbar [30] generally follow the "Linear" relations while the "Cubic" relationship often corresponds to common data intensive workload on network devices in the cloud. Y/1 value heavily depends on the underlying network architecture. In this work, we only consider the energy cost from network devices such as NICs, routers and swtiches while ignoring the energy spent on the optical transmitters. 3). Infrastructure energy consumption such as cooling is another major part of the data center operating costs. Some report [4] indicates that cooling itself can contribute as high as 33% of the entire data center's energy consumption. However, cooling energy can be very complicated to model because it is not directly impacted by the workload. Based on PUE [31], we can treat this part of the energy cost as a fraction of the total system energy consumption. Since it will not affect the runtime scheduling decision, we simply ignore its effects in the fi nal model. Based on the assumptions above, we can build a simple en ergy cost model for data centers only considering three major energy-consuming components: server nodes, network devices (routers, switches, etc), and the cooling system. According to our previous discussion on the cooling
(https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-58b9d1530f2d1.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - HPC runtime support for fast and power efficient locking and synchronization (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-58b9d1530f2d1.html)
energy in Section III-A(3), we can build a weighted combination of linear (for servers) and degree Yn polynomial (for network
Documents
devices) relationships between energy consumption and network traffic load in our model. The total energy consumption of
(https://docslide.net/category/documents.html)
all the replicas can be modeled as: Eg = L Un' (an L Pe,n + .Bn(L Pc,n?") (1) N e e where an and.Bn are weight scalars for the energy consumption of servers and network devices in replica n. The goal of our problem is to minimize 9 for the clients' requests to the data centers. The global optimization problem can be formulated as: minimize PC,11 subject to Eg = L En N
(https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-585ce516e506e.html)
scalability can also be achieved through EDR's decentralized architecture. If the replica selection work is assigned to a
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Lit: A high performance massive data computing framework based on CPU/GPU cluster (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-585ce516e506e.html)
single central agent, the crash of such agent can cause the failure of the entire Replica Layer (N): PC,n Client Layer(C): Fig.
Documents
in(P) = L Pc.n - Bn 0, Vn E N e he(P) = L Pe.n - Rc = 0, Vc E C N ee,n(P) = Ic,n - T 0, Vn EN, c E C (2) where En = Un . (an Ie Pc,n + .Bn(Ie Pc,nyr,,) is the energy consumption of replica n, fi,(P) is the bandwidth capacity constraint of replica n, hc(P) is the request constraint of client c, and ec.n(P) represents the network latency constraint from client c to replica n. The problem turns out to be a degree Yn polynomial objective function (convex function) with several linear equality and inequality constraints. B. EDR System Architecture The EDR system is built on top of the common data center infrastructure without additional devices, shown in Fig.l. In the system, each replica keeps listening to the clients' incoming requests. Once the requests are received, the replicas will start cooperating with each other to solve the global optimization problem. In EDR, the replica selection service is transparent to the clients, which means the clients do not need to know which replica(s) they are communicating with. This is decided at run time by EDR. Higher system reliability and better
I. An illustration of a general service-oriented cloud with multiple replicas and clients. Fig. 2. The EDR server side components diagram. replica selection system. It is unlikely to happen in a decen tralized environment with efficient fault tolerance mechanisms unless all the replicas malfunction. However, the decentralized solution does not always outperform the centralized method. For a lower runtime calculation workload (e.g. fewer clients and smaller workload in each request), the communication and synchronization overhead may result in a performance degradation for the decentralized system. Fortunately, users and companies care more about the energy consumption during the peak service hours, which dominate the entire operating cost. Still, selecting high-performance distributed algorithms used for solving the global optimization problems in EDR is essential for minimizing the energy cost during the decision making phase at runtime. The detailed descriptions of the selected distributed algorithms used in EDR can be found in Section III-D. C. EDR System Design The replica selection system involves the client side and the server side. The programs of both sides are designed as multithreaded programs using TCPjIP sockets for communication. The structure of the server side program is illustrated in Fig. 2. The ClientListener thread keeps listening to the new requests from clients. The ReplicaListener thread keeps listening to the requests of solution information from other replicas. The FileDownload thread handles the sending of requested files to the clients. Once a new client request comes, it communicates with the ClientListener thread first and then waits for the solution of how to distribute its requested load. Once the solution is reached, the client side will create new threads to
(https://docslide.net/category/documents.html) (https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-58a8bf90b9097.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Forcedirected geographical load balancing and scheduling for batch jobs in distributed datacenters (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-58a8bf90b9097.html) Documents
(https://docslide.net/category/documents.html)
communicate with all the replicas at the same time to download the computed amount of load. In the current EDR design, we guarantee the reliability of the system by using a combination of time-out mechanism and ring fault-tolerance structure. The ReplicaListener thread is used to communicate between replicas. Once a replica malfunctions, the other replicas will know and then remove this dead replica from their "active member lists" and the ring structure. After that, EDR will perform the runtime scheduling again based on the new ring of replicas. D. Solving Global Optimization Problem: LDDM v.s. CDPSM Solving an optimization problem with constraints in a distributed environment is not as easy as on a single node. In order to solve (2) in a distributed manner, we consider two methods as the candidates: Lagrangian duaJ decompo sition method (LDDM) [32] and consensusbased distributed projected subgradient method (CDPSM) [33]. LDDM is an iterative method for solving the convex optimization problem in paraJlel. Other than considering forming the dual problem, CDPSM presents a consensus mechanism to solve the convex optimization problem, whose objective function is sum of several local objective functions, through distributed agents. Both of them can be adapted to solve our constrained convex optimization problem in parallel. In this paper, we implement both algorithms and then compare their communication com plexity and convergence speed. 1) Consensus-based Distributed Projected Subgradient Method (CDPSM): This method is originally proposed to solve constrained optimization problems in multi-agents net works. In our paper, we adapt this method to our EDR system. The
(https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-58b8df466a8db.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Write bandwidth optimization of online Erasure Code based cluster file system (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-58b8df466a8db.html) Documents
(https://docslide.net/category/documents.html)
objective function Eg in our replica selection problem is the sum of functions which are local objective functions for replicas, in the form of Eg = IN En. Each replica works on solving its own 10caJ optimization problem En which is subject to the 10caJ constraints Pe,n E Pn, where Pn is a subset of the constraint sets that have 10caJ variables of replica n. The optimization problem in replica n can be formulated as: minimize En Pe,n subject to P,n E P n The main idea of this algorithm is to use a consensus mechanism among distributed replicas to split the computation work. Each distributed replica keeps working on solving a subproblem of the globaJ problem. The consensus mechanism can combine solutions of subproblems to form the global optimization solution. Given Pe.n is the solution to the global optimization problem, each replica n starts by estimating (Pe,n ICE C n E N}n E Pn and updating its solution Pe,n iteratively by cooperating with other replicas. The consensus and projection procedure for iteratively estimating can be denoted by the following equation: N P.n(k + 1) = ProjpJL. a . P{n(k) - dk . gn(k)t (3) j=i where a are the weights of all the replicas, dk > ° is the step size, and gn(k) is the subgradient on its 10caJ objective function En. Since the objective function of our problem is twice differentiable, we could use gradient instead of subgradient as gn(k). The symbol ProjpJ·]+ denotes the operation of
(https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-588d4e835a620.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - LVFS: A scalable big data scientific storage system (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-588d4e835a620.html)
projection. We have: Projp,,[p,nt = arg min IIP,n - Pe,nll Pc,f/EPn By projecting the solution Pe.n back into its own local
Documents
con straint set Pn, the algorithm guarantees that in each iteration the solution is feasible. Based on this method, every
(https://docslide.net/category/documents.html)
replica in our system keeps running to handle client requests and the consensus mecha nism. We can present the algorithm for each replica as follows: Algorithm 1 Algorithm of CDPSM 1: Initialization: Set the unit price of replica i. 2: repeat 3: Collect the clients' requests from clients. 4: Collect the solution Pe.n from other replicas. 5: Get the consensus solution Ve.n = In anP,n' where In an = I 6: Update solution by Pe,n = Ve,n - d . g(Ve,n), where d is step size and g(Ve,n) is gradient value of function En at Ve,n. 7: Project Pe,n to the constraint sets following the project rule PxJP,n]+ 8: until
Pe,n do not change. The size of solution Pe,n in each replica is O(ICI·IN!). The consensus mechanism requires distributed replicas to request the solutions from other replicas. So the communication com plexity of each iteration is of size O(ICI·INI·IN - II· IN!) which is approximately O(ICI' INI3), where C is the number of clients and N is the number of replicas. 2) Lagrangian Dual Decomposition Method (LDDM): Since there are dependencies in the global variables among replicas, we need to decouple them in order to solve the prob lem in parallel. LDDM provides us with a way to solve such problem. Given the original problem (2), we can formulate the Lagrangian duaJ problem from the global optimization problem as: minimize Pe,n N C L(Pe,n,fl) = L. En + L.fli· he(P) n=l e=1 subject to fn(P) = L. Pe,n - Bn 0, Vn E N ee,n(P) = te,n - T 0, Vn E N (4) By using the Lagrangian multiplier fl, the equality con straints that have the global coupling variables of the original problem, are transformed into the objective function of its dual problem (4). So for the replicas in our system, each of them just needs to solve the local optimization problem and update fl by the clients periodically. The local optimization subproblem is defined as (in replica n): C minimize En + L. fli . Pe,n PC,II c=l subject to L. Peon - Bn ° C te,n - T 0, Ve E C (5) where {Pe,n IC E C} are the local variables in replica n. The task of updating fl is assigned to the clients since the equality constraints in the original problem (2) are associated with each client request. The updating of J-l is done by solving the problem (6). Gradient method can be used to solve such linear programming problem. J-l can be any real number.
(https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-58c4b458a77ab.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) Parallelizing windowed stream joins in a shared-nothing cluster (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-58c4b458a77ab.html) Documents
(https://docslide.net/category/documents.html)
minimize gCJ-i) = inf L(Pe.n, J-l) tt PC,II subject to J-l E RC We implement the algorithm as below: Algorithm 2 Algorithm of LDDM (Replica n) l : Initialization: Set the unit price of replica i. (6) 2: Collect the clients' requests and their values of J-l and inform the other replicas. 3: repeat 4: Solve the local optimization problem (5). 5: Send solution Pc,n to each client c. 6: Request the new J-lc from the client c. 7: Stops if {Pe,n ICE C} do not change. 8: until Pe.n do not change. To achieve higher performance for distributed algorithms, both low communication complexity and high algorithm con vergence rate are required. Comparing with CDPSM, the system implemented with the LDDM has lower complexity. Its runtime coordination is between pairs of clients and replicas, so there is little communication among the replicas. The size of the solution of each replica is O(IC!). The communication complexity of each iteration is O(ICI . IN!), which is lower than the complexity of using CDPSM shown in the previous subsection. In theory, the LDDM also has higher convergence rate than CDPSM. Fig. 5 shows the comparison of simulated convergence rates of these two methods. To explain the con cept, we conduct a simulation experiment with three replicas using MatLab. For solving the same optimization problem, the CDPSM converges slower than the LDDM. So theoretically speaking, the LDDM is expected to have higher performance for solving our problem. This theory will be validated by the experiments of running data intensive applications on real world machines in Section IV. Additionally, to solve convex minImization problem, the step size we choose in the algorithm can affect the
(https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-585cf1290dead.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Zput: A speedy data uploading approach for the Hadoop Distributed File System (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-585cf1290dead.html) Documents
(https://docslide.net/category/documents.html)
conver gence speed or even determine if the algorithm can converge successfully. To guarantee the fairness of the comparison, we choose to use constant step size for both algorithms in this paper. IV. EXPERIMENT RESULTS AND ANALYSIS In this section, we first present a system which can emulate the behaviors of data centers in cloud in terms of energy consumption. Then, we use two types of data-intensive applica tions, video streaming and distributed file services, to evaluate data centers' performance and energy cost with EDR. Finally, we demonstrate experimentally the cost differences in various replica selection scenarios and show that EDR performs as well as DONAR while additionally taking system energy costs into consideration. Olslrlbuled ProjeCled Sut;.gractenl Method (Performance) Lalr.'flljanDualo..cQmp,vj'A4W';VJ' Fig. 5. Simulation results for CDPSM and LDDM methods in our EDR system. Different convergence rates have been shown here for comparison. A. System Setup and Assumptions 1) Assumptions: In the following experiments, we are going to use a single cluster node to emulate a real replica. We assume that, for dataintensive applications, energy cost model of a single cluster node is very similar to that of a data center. It can be proved as below. Assuming we have workload p, based on equation (1) we can describe the energy consumption(Es) of a single
(https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-58d6320ebd461.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - On service migration in the cloud to facilitate mobile accesses (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-58d6320ebd461.html)
cluster machine as: Es = ap + {3p3 (7) We assume Yn =3 here for data intensive applications. If we are using a data center
Documents
to handle P client requests, the task can be split into Pi where 2:1 Pi = p, N is the number of nodes involved with this task
(https://docslide.net/category/documents.html)
in this data center. So the energy consumption(Ed) of this data center for request p is: N N N N Ed = L. (a Pi + {3pt) = a L. Pi+{3 L. Pt = ap+{3 L. Pt (S) i=1 i=1 i=1 i=1 In reality, the energy consumption of network devices is much lower than that of servers in a data center. Therefore, we can assume that the value of {3 is much smaller than a in equation (1). So we can have Es ::::: Ed. Based on this assumption, it is reasonable for us to use a cluster node to model the energy behaviors of a real replica in cloud. 2) System Setup: We use eight nodes of our SystemG cluster as replicas to conduct our experiments. The SystemG cluster is a 22.S TFLOPS supercomputer providing a research platform for development of high performance software and simulation tools. Each node is equipped with two quad-core 2.S GHz Intel Xeon Processors, an S GB RAM, and a 6MB cache. SystemG is also equipped with both Ethernet and Infiniband adapters and switches. In this experiment, we use the Domin ion PX Intelligent Power Distribution Units to dynamically profile power consumption of controlled machines. The power sampling rate is approximately 50 times/sec. The model parameters used in this section are defined as follows: 1) For the electricity prices (¢/kwh) in this study, we random generate an integer number between 1 and 20 for each of the S replicas in every experiment. This is to simulate various power prices of data centers in different geographical locations; 2) the bandwidth cap for our SystemG Ethernet is approximately 100 MB/s; 3) we set the user-defined max tolerable network latency T as l.Sms, the worst case scenario for one full-size frame ISIS bytes under heavy workload on SystemG; 4) according to the measurement on SystemG, we set the scalars an = 1, and {3n = 0.01. In our experiments, we
(https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-58c92fcd147f2.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - K MapReduce: A scalable tool for dataprocessing and search/ensemble applications on large-scale supercomputers (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-58c92fcd147f2.html) Documents
(https://docslide.net/category/documents.html)
use two types of data-intensive ap plications: the video streaming and the distributed file service. 240 235 replica! 230 225 220 215 -'------------ 240 235 .. 1 230 i 225 & 220 TIme(sec) replicaS 215 -'------------ 1 11 21 31 41 51 61 71 81 91 101 TimEic) 240 235 230 225 220 215 240 235 1 230 i 225 & 220 215 replica2 n ::: " ;;: ;ri ;;; ;:: iii Time{secl 1 11 21 31 41 51 61 71 Timelser) 240 240 235 replica3 235 230 230 225 225 220 220 215 215 a: 0 1 11 21 31 41 51 61 71 81 91 101 1 n n 11 21 31 41 51 61 71 81 91 nme{secl Time{secl 240 240 235 replica7 235 .. .. ! 230 1 230 225 i 225 & & 220 220 215 215 81 91 1 11 21 31 41 51 61 71 81 91 101 1 11 21 31 41 51 61 71 81 91 101 Time(sec) Time(secl Fig. 3. Runtime power profile for individual replica using COPSM (distributed file service) 225 " 223 1221 i 219 & 217 replica! 215 -'------------ 1 11 21 31 41 51 61 71 Time{secl 225 223 replicaS 221 219 217 215 1 11 21 31 41 51 61 71 81 Time(secl 225 replica2 225 " 223 " 223 1221 1221 219 & 219 & 217 217 215 215 1 11 21 31 41 51 61 71 81 91 Time(secl 225 225 " 223 replica6 " 223 1221 1221 i 219 & 219 & 217 217 215 215 1 11 21 31 41 51 61 71 81 91 101 Time(secl 225 replica3 replica4 " 223 1221 219 & 217 215 1 11 21 31 41 51 61 71 81 91 1 11 21 31 41 51 61 71 81 Time{secl Time{sec) 225 replica7 " 223 replicaS 1221 i 219 & 217 215 n n n n n n n n iii n n n n N m " m 0 n 1 11 21 31 41 51 61 71 81 91 101 n n Time(secl Time(secl Fig. 4. Runtime power profile for individual replica using LOOM (distributed file service) The size per request is different for these two applications. We set the size per request for the video streaming is approx imately 100 MBytes and for the distributed file service it is approximately 10 MBytes. B. Performance and Power Analysis In this subsection, we use a data intensive application (distributed file service) as an example to study the power and performance characteristics of our EOR system implemented with COPSM and LOOM. The power profiles of 8 replicas running with distributed file service application using our EOR system are shown in Fig 3 (COPSM) and Fig 4 (LOOM). In most cases, system energy is consumed by both replica selection phase (including local solution calculation and global solution synchronization), and the file transferring phase after the selection. The "valleys" shown in these two figures represent the time when only replica selection process is running or system is listening to the new requests. The "peaks" represent the time when replicas are accepting new clients' requests or transferring files to the previous clients. The execution time of each replica shown in the figures depends on both assigned workload and the solution calculation+synchronization time. We can observe that, when handling the same number of client requests, EOR system implemented with LOOM runs faster than the system with COPSM (for most of the individual cases, LOOM finishes earlier). It validates that LOOM has a lower communication complexity and better convergence rate than COPSM. Also, the average power of using LOOM is lower than that of using COPSM. The reason is that compared to LOOM, COPSM needs to coordinate with all other replicas and clients at every iteration in order to make runtime scheduling decision, which results in constant higher workload intensity. This aJso indi cates that COPSM's system complexity is higher than LOOM. In Fig 4, we can also observe that the power consumption of replica 3 and 5 remain constantly low during the entire execution. This is because neither of these two replicas has been selected as the downloading targets by EOR at runtime. C. Energy Cost Analysis In order to show that EOR can effectively reduce the total energy cost of the data centers, we conduct the experiments with 8 replicas and real time client requests. The pattern of data-intensive requests follows Youtube commercial workload patterns [34]. Based on the traffic data, we evaluate the total en ergy cost of all 8 replicas running with LOOM- and COPSM based EOR. And then we compare the results with that of using baseline algorithm, Round-Robin. For example, Fig. 6 and 7 show the energy cost of each of the 8 replicas running with video streaming and distributed file service applications under three different algorithms. The electricity prices (¢jkwh) for No.1 to No.8 replicas are: 1,8,1,6,1,5,2,3. This is randomly generated according to Section IV-A(2). Fig.6 and 7 show that the traffic loads are successfully distributed among replicas under EDR runtime scheduling. Here we take the application of video streaming as an example, shown in Fig. 6. The pattern of distribution is determined by our energy cost model at runtime considering varying electricity prices, bandwidth capacity of each replica and network latency. In Fig.6, most of the traffic load is assigned to replica 3, 5, and 7 primarily due to the relatively lower electricity prices, but also related to the bandwidth cap and network latency. 450000 400000 350000 300000 '* 250000 8 200000 i:i 150000 100DOO W 50000 4 Replica .LOOM .COPSM • Round Robin Fig. 6. Energy cost of each replica for the video streaming application under three different scheduling approaches. ..,. c 250000 200000 ! 150000 8 100000 50000 Replica • LOOM .COPSM • Round Robin Fig. 7. Energy cost of each replica for the distributed file service application under three different scheduling approaches. Fig.8 shows the total energy consumption and cost of all eight replicas running with LDDM- and CDPSM based EDR and the baseline system running with Round-Robin method for two data intensive applications. Fig.8 (a) shows that LDDM outperforms both CDPSM and Round-Robin approaches in terms of total energy cost. It is because LDDM can converge faster than CDPSM in general, which means less communi cation and synchronization overhead. Fig.8(b) shows a very interesting phenomenon: for the video streaming case, CDPSM actually consumes less total energy Goules) than LDDM. This result still makes sense because our objective function is to minimize the total energy cost (cents) instead of total energy consumption. Also, even though CDPSM requires additional energy for computation and communication, it still outper forms the round-table method because CDPSM does provide the global optimization solution for workload distribution. The observations from Fig.8 are consistent with the other 40 runs under various configurations using EDR. Through all the runs, the LDDM-based EDR can save an average of 12% energy cost compared to the Round-Robin method, while CDPSM-based EDR can save an average of 22.64% energy consumption. looOOOOm - 'OOOOODt f ::: l= -11- t::: H- -= 400000 '" 150000 _ _ - - :!E1OOOOO i200 ]50000 _
LOOM CDPSM Round Robin 0 LOOM CDPSM
Round Robin • Video Streaming .DFS • Video Streminll .OFS ) Fig. 8. The total data center energy cost and energy consumption comparisons under three different scheduling approaches for two data intensive applications. D. System Performance Analysis While the decentralized architecture may bring the issue of communication overhead to the replica selection system, we validate the performance of EDR with another efficient decentralized replica selection system, DONAR [3]. Unlike EDR, DONAR does not consider energy cost reduction at runtime scheduling. In this experiment, we use three replicas in EDR and three mapping nodes in DONAR. These mapping nodes function as distributed coordinators to split the load into different replicas. The requests also follow the pattern of Youtube.com. The system performance results for EDR and DONAR are shown in Fig.9. 24 48 72 96 120 144 168 192 Request Count Fig. 9. System Performance of DONAR and EDR in terms of response time while number of client requests scale. From Fig. 9, we can observe that the performance of EDR is very close to DONAR, which has been validated to be as efficient as the centralized system [3] . The response time per request is less than 200 ms. And the response time increases close to linearly when the client requests increase. As we mentioned previously, EDR is implemented with LDDM which has the communication complexity of O(ICI . 1M), and for DONAR it is O(ICI . 1M . IMI) where IMI is the number of the mapping nodes. Therefore, with the increasing system size IMI, EDR will eventually outperform DONAR in a large scale cloud system. V. CONCLUSION AND FUTURE WORK In this paper, we propose EDR, an energy-aware runtime scheduling system for data-intensive applications in the cloud. EDR provides a decentralized architecture to solve the replica selection problem and considers not only the bandwidth ca pacity and network latency but also the total energy cost of the entire cloud when forming the data center energy cost model. Our experiments prove that EDR can effectively reduce the total energy cost with a comparable efficiency to the best available decentralized replica selection system named DONAR. In future, we plan to port EDR to a large scale real world commercial cloud such as Amazon EC2, and also with more restrictions other than bandwidth capacity and latency. REFERENCES [I] A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel, "The cost of a cloud: research problems in data center networks," SIGCOMM Comput. Commun. Rev., vol. 39, no. I, pp. 68-73, Dec. 2008. [Online]. Available: http://doi.acm.orgjlO.1145/1496091.1496103 [2] M. Litzkow, M. Livny, and M. Mutka, "Condor-a hunter of idle workstations," in Distributed Computing Systems, 1988., 8th International Conference on, jun 1988, pp. 104 -I l l . [3] P. Wendell, J. W. Jiang, M. J. Freedman, and J. Rexford, "Donar: decentralized server selection for cloud services," SIGCOMM Comput. Commun. Rev., vol. 41, pp. 231-242, August 2010. [Online]. Available: http://doi.acm.org/IO.1145/2043164.1851211 [4] the U.S. EPA ENERGY (2007) http://www.energystar.gov. http://www.energystar.gov STAR [Online]. Program. Available: [5] J. G. Koomey, C. Belady, M. Patterson, and A. Santos, "Assessing trends over time in performance, costs, and energy use for servers," 2009. [6] J. Koomey, "Growth in data center electricity use 2005 to 2010," Tech. Rep., July 2011. [Online]. Available: http://fulltextreports.com/2011/08/04/growth-in-data-center electricity-use-2005-to-2010/ [7] A. Qureshi, "Plugging Into Energy Market Diversity," in 7th ACM Workshop on Hot Topics in Networks (HotNets), Calgary, Canada, October 2008. [8] R. Bianchini and R. Rajamony, "Power and energy management for server systems," Complller, vol. 37, no. II, pp. 68 - 76, nov. 2004. [9] L. Rao, X. Liu, L. Xie, and W. Liu, "Minimizing electricity cost: Optimization of distributed internet data centers in a multi-electricity market environment," in INFOCOM, 2010 Proceedings IEEE, march 2010, pp. I -9. [10] A. Berl, E. Gelenbe, M. Di Girolamo, G. Giuliani, H. De Meer, M. Q. Dang, and K. Pentikousis, "Energy-efficient cloud computing," The Complller.!ournal, vol. 53, no. 7, pp. 1045-1051,2009. [Online]. Avail able: http://comjnl.oxfordjournals.orgjcgi/doi/10.1093/comjnl/bxpO80 [II] R. Urgaonkar, U. Kozat, K. Igarashi, and M. Neely, "Dynamic resource allocation and power management in virtualized data centers," in Network Operations and Management Symposium (NOMS), 2010 IEEE, april 2010, pp. 479 -486. [12] R. Buyya, A. Beloglazov, and J. H. Abawajy, "Energyefficient man agement of data center resources for cloud computing: A vision, architectural elements, and open challenges," CoRR, vol. abs/1006.0308, 2010. [13] A. Beloglazov and R. Buyya, "Energy efficient resource management in virtualized cloud data centers," in Proceedings of the 2010 10th IEEEjACM International Conference on Cluster, Cloud and Grid Compwing, ser. CCGRID '10. Washington, DC, USA: IEEE Computer Society, 2010, pp. 826-831. [Online]. Available: http://dx.doi.orgjlO.1109/CCGRID.201O.46 [14] Z. Zong, M. Briggs, N. O'Connor, and X. Qin, "An energy-efficient framework for large-scale parallel storage systems," in Parallel and Distributed Processing Symposium, 2007. IPDPS 2007. IEEE Interna tional, march 2007, pp. I -7. [15] H. S. Kim, D. I. Shin, Y. J. Yu, H. Eom, and H. Y. Yeom, 'Towards energy proportional cloud for data processing franleworks," in Proceedings of the First USENIX conference on Sustainable in/ormation technology, ser. SustainIT'IO. Berkeley, CA, USA: USENIX Association, 2010, pp. 4--4. [Online]. Available: http://dl.acm.orgjcitation.cfm?id=1863159.1863163 [16] Z. Liu, M. Lin, A. Wierman, S. H. Low, and L. L. Andrew, "Greening geographical load balancing," in Proceedings of the ACM SIGMETRICS joint international conference on Measurement and modeling of compwer systems, ser. SIGMETRICS ' II. New York, NY, USA: ACM, 2011, pp. 233-244. [Online]. Available: http://doi.acm.orgjlO.1145/1993744.1993767 [17] 1. W. Smith and I. Sommerville, "Workload classification & software energy measurement for efficient scheduling on private cloud platforms," CoRR, vol. abs/1105.2584, 2011. [18] A. Lewis, S. Ghosh, and N.-F. Tzeng, "Run-time energy consumption estimation based on workload in server systems," in Proceedings of the 2008 conference on Power aware computing and systems, ser. HotPower'08. Berkeley, CA, USA: USENIX Association, 2008, pp. 4-4. [Online]. Available: http://dl.acm .orgjcitation .cfm?id= 1855610.1855614 [19] P. Mahadevan, P. Sharma, S. Banerjee, and P. Ranganathan, "Energy aware network operations," in INFOCOM Workshops 2009, IEEE, april 2009, pp. I -6. [20] S. Seetharaman, "Energy conservation in multi-tenant networks through power virtualization," in Proceedings of the 2010 international conference on Power aware computing and systems, ser. HotPower'lO. Berkeley, CA, USA: USENIX Association, 2010, pp. 1-8. [Online]. Available: http://dl.acm.orgjcitation.cfm?id=I924920.1924924 [21] A. Ruiz-Alvarez and M. Humphrey, "An automated approach to cloud storage service selection," in Proceedings of the 2nd international workshop on Scientific cloud compilling, ser. ScienceCloud ' II. New York, NY, USA: ACM, 2011, pp. 39-48. [Online]. Available: http://doi.acm.orgjlO.1145/ 1996109. /996117 [22] K. Le, R. Bianchini, M. Martonosi, and T. Nguyen, "Cost-and energy aware load distribution across data centers," Proceedings of HotPower, 2009. [23] G. Chen, W. He, J. Liu, S. Nath, L. Rigas, L. Xiao, and F. Zhao, "Energy-aware server provisioning and load dispatching for connection-intensive internet services," in Proceedings of the 5th USfcrJIX Symposium on Networked Systems Design and Implementation, ser. NSDI'08. Berkeley, CA, USA: USENIX Association, 2008, pp. 337-350. [Online]. Available: http://dl.acm.orgjcitation.cfm? id=1387589.1387613 [24] K. Rajanlani and C. Lefurgy, "On evaluating request-distribution schemes for saving energy in server clusters," in Proceedings of the 2003 IEEE International Symposium on Performance Analysis of Systems and Software, ser. ISPASS '03. Washington, DC, USA: IEEE Computer Society, 2003, pp. 111-122. [Online]. Available: http://dl.acm.orgjcitation.cfm?id=1153924.1154555 [25] F. Dabek, R. Cox, F. Kaashoek, and R. Morris, "Vivaldi: a decentralized network coordinate system," in Proceedings of the 2004 conference on Applications, technologies, architeclilres, and prolOcols for computer communications, ser. SIGCOMM '04. New York, NY, USA: ACM, 2004, pp. 1526. [Online]. Available: http://doi.acm.orgjlO.1145/1015467.1015471 [26] S. Song, c.-Y. Su, R. Ge, A. Vishnu, and K. W. Canleron, "Iso-energy efficiency: An approach to power-constrained parallel computation," in Proceedings of the 201 J IEEE International Parallel & Distributed Processing Symposium, ser. IPDPS ' II. Washington, DC, USA: IEEE Computer Society, 2011, pp. 128-139. [Online]. Available: http://dx.doi.orgjI0.1109{IPDPS.2011.22 [27] S. Song and K. W. Cameron, "System-level power-performance effi ciency modeling for emergent gpu architectures," in PACT, 2012, pp. 473-474. [28] S. Nedevschi, L. Popa, G. Iannaccone, S. Ratnasamy, and D. Wetherall, "Reducing network energy consumption via sleeping and rate adaptation," in Proceedings of the 5th USENIX Symposium on Networked Systems Design and Implementation, ser. NSDI'08. Berke ley, CA, USA: USENIX Association, 2008, pp. 323-336. [Online]. Available: http://dl.acm.org/citation.cfm?id=1387589.1387612 [29] J. Restrepo, C. Gruber, and C. Machuca, "Energy profile aware routing," in Communications Workshops, 2009. ICC Workshops 2009. IEEE International Conference on, june 2009, pp. I -5. [30] T. Ye, L. Benini, and G. De Micheli, "Analysis of power consumption on switch fabrics in network routers," in Design AU/omation Conference, 2002. Proceedings. 39th, 2002, pp. 524 - 529. [31] T. G. Grid, "The Green Grid Data Center Power
(https://docslide.net/documents/ieee-2013-ieeeinternational-conference-on-cluster-computingcluster-indianapolis-58c1b2bbc64ad.html)
[IEEE 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Indianapolis, IN, USA (2013.09.23-2013.09.27)] 2013 IEEE International Conference on Cluster Computing (CLUSTER) - Fast data analysis with integrated statistical metadata in scientific datasets (https://docslide.net/documents/ieee2013-ieee-international-conferenceon-cluster-computing-clusterindianapolis-58c1b2bbc64ad.html) Documents
(https://docslide.net/category/documents.html)
View more (https://docslide.net/search? q=%5BIEEE+2013+IEEE+International+Conference+on+Cluster+Co +Indianapolis%2C+IN%2C+USA+%282013.09.232013.09.27%29%5D+2013+IEEE+International+Conference+on+Clu +EDR%3A+An+energyaware+runtime+load+distribution+system+for+dataintensive+applications+in+the+cloud)
Efficiency Metrics: PUE and DCiE," Tech. Rep., 2007. [32] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and distribwed computa tion: numerical methods. Upper Saddle River, NJ, USA: Prentice-Hall, Inc., 1989. [33] A. Nedic, A. Ozdaglar, and P. Parrilo, "Constrained consensus and optimization in multi-agent networks," AUlOmatic Control, IEEE Trans actions on, vol. 55, no. 4, pp. 922 -938, april 2010. [34] P. Gill, M. Arlitt, Z. Li, and A. Mahanti, "Youtube traffic characteriza tion: a view from the edge," in Proceedings of the 7th ACM SIGCOMM conference on Internet measurement, ser. IMC '07. New York, NY, USA: ACM, 2007, pp. 15-28.
Download (https://docslide.net/download/link/ieee2013-ieee-international-conference-oncluster-computing-cluster-indianapolis58c19be5948d4) (https://docslide.net/documents/ieee-2013-ieee-international-conference-on-cluster-computing-clusterindianapolis-58c19be5948d4.html) Facebook
Twitter
Google+
Email
More
DESCRIPTION
TOP RELATED
(https://docslide.net/documents/ieee-2013-ieee-international-conference-on-cluster-computing-cluster-indianapolis-58ccb73558c9b.html) (https://docslide.net/documents/ieee-2013-ieee-international-conference-o (https
[IEEE 2013 IEEE INTERNATIONAL …
COMPANY
[IEEE 2013 IEEE INTERNATIONAL …
CONTACT & LEGAL
[IEEE 2013 IEEE INTERNATIONAL …
OPENING HOURS
(https://facebook.com/d (https://twitter.com (https://goo