You are on page 1of 18

Cluster Comput

DOI 10.1007/s10586-014-0404-x

Energy-efficient data replication in cloud computing datacenters
Dejene Boru · Dzmitry Kliazovich · Fabrizio Granelli ·
Pascal Bouvry · Albert Y. Zomaya

Received: 18 March 2014 / Revised: 4 July 2014 / Accepted: 24 September 2014
© Springer Science+Business Media New York 2015

Abstract Cloud computing is an emerging paradigm that Keywords Cloud computing · Data replication ·
provides computing, communication and storage resources Energy efficiency
as a service over a network. Communication resources often
become a bottleneck in service provisioning for many cloud
applications. Therefore, data replication which brings data 1 Introduction
(e.g., databases) closer to data consumers (e.g., cloud appli-
cations) is seen as a promising solution. It allows mini- Cloud computing is an emerging technology that attracts
mizing network delays and bandwidth usage. In this paper ICT service providers offering tremendous opportunities
we study data replication in cloud computing data cen- for online distribution of services. It offers computing
ters. Unlike other approaches available in the literature, we as a utility, sharing resources of scalable data centers
consider both energy efficiency and bandwidth consump- [1,2]. End users can benefit from the convenience of
tion of the system. This is in addition to the improved accessing data and services globally, from centrally man-
quality of service QoS obtained as a result of the reduced aged backups, high computational capacity and flexible
communication delays. The evaluation results, obtained billing strategies [3]. Cloud computing is also ecologi-
from both mathematical model and extensive simulations, cally friendly. It benefits from the efficient utilization of
help to unveil performance and energy efficiency trade- servers, data center power planning, large scale virtualiza-
offs as well as guide the design of future data replication tion, and optimized software stacks. Nevertheless, electric-
solutions. ity consumed by cloud data centers is still in the order
of thousands of megawatts [4]. In 2010, datacenters con-
sumed around 1.1–1.5 % of global electricity consumption
D. Boru
and between 1.7 and 2.2 % for U.S [5,6]. Pike Research fore-
CREATE-NET, Via alla Cascata 56/D, Trento, Italy
e-mail: dejene.oljira@create-net.org casts data center consumption of almost 140 TW h in 2020
[7].
D. Kliazovich (B) · P. Bouvry The growth of Internet services at an unprecedented rate
University of Luxembourg, 6 rue Coudenhove Kalergi,
Luxembourg, Luxembourg
requires the development of novel optimization techniques
e-mail: dzmitry.kliazovich@uni.lu at all levels to cope with escalation in energy consumption,
P. Bouvry
which in place would reduce operational costs and carbon
e-mail: pascal.bouvry@uni.lu emissions.
F. Granelli
In data centers, there is an over provisioning of com-
DISI - University of Trento, Via Sommarive 14, Trento, Italy puting, storage, power distribution and cooling resources
e-mail: granelli@disi.unitn.it to ensure high levels of reliability [8]. Cooling and power
distribution systems consume around 45 and 15 % of the
A. Y. Zomaya
School of Information Technologies, University of Sydney,
total energy respectively, while leaving roughly 40 % to the
Darlington, Australia IT equipment [9]. These 40 % are shared between com-
e-mail: albert.zomaya@sydeny.edu.eu puting servers and networking equipment. Depending on

123

voice and video conferencing. solutions for communication equipment depend on (a) down- • Development of a data replication approach for joint opti- grading the operating frequency (or transmission rate) or (b) mization of energy consumption and bandwidth capacity powering down the entire device or its hardware compo- of data centers. ing the number of idle servers that can be put into a sleep where the cloud applications are running. In 2003. we propose a data replication tech- P = V 2 ∗ f. such formance. virtualization enables implemen- results obtained through simulations. Another technology which indirectly affects energy con- sumption is virtualization. data resources can consolidation scheduler—the policy which allows maximiz- be brought closer (replicated) to the physical infrastructure. DPM results in most of the energy sav- tion resources [14].15–18]. Both methods are applicable to computing servers as gaming. However.1 Energy efficiency There are two main approaches for making data center consume less energy: shutting the components down or scal. backup. After egy through mathematical modeling and using a packet- that. ciency [21]. applies only to the CPU. It is the most efficient if combined with the workload mance low latency service provisioning. ing servers and network switches. Section 4 provides evalua. ings. frequency f : ters.13]. the for- storage. • Modeling of energy consumption characteristics of data Similar to computing servers. memory. 2. ity. there are two main alternatives for ing down their performance. most of the energy-efficient center IT infrastructures. with different posed replication scenarios. Cluster Comput the data center load level. Server resources data replication. online office. These strategies optimize sys. In networking. shutdown (DNS) to further optimize energy consumption • Analysis of the tradeoff between performance. The second method in the literature [8. supplied voltage V . none of them focuses on power consumption P. bandwidth demand and cation requirements. Section 5 presents evaluation objectives. 123 . Reducing voltage or frequency reduces the power consump- ters as well as inside each datacenter. Virtualization is widely used in The rest of the paper is organized as follows: Sect. disks as well as peripheral devices continue consuming at their peak rates. agement. virtualization can be applied in both the comput- tion of the model outlining theoretical limits for the pro. In Sect. as power reduction tributions can be summarized as follows. nents in order to conserve energy. and may not necessarily have the goal of energy effi- topic. serviceabil- [13]. DVFS exploits the relation between distributed data centers. while system bus. network bandwidth and communica- tion delay both between geographically distributed data cen. 3 we develop a mathematical can be dynamically provisioned to a VM based on the appli- model for energy consumption. depends largely on the mer method is commonly referred to as dynamic power man- availability and efficiency of high-performance communica. work that proposed a power-aware interconnection network • Performance evaluation of the developed replication strat- utilized dynamic voltage scaling (DVS) links [10]. social networking. tion. Section 6 concludes tation of logically different addressing and forwarding mech- the paper and provides an outline for the future work on the anisms.12] and network switches [10. At the component level. agement (DPM) [11]. When applied to the servers. DVS technology was combined with dynamic network level cloud computing simulator. Both approaches are applicable making data center consume less energy: (a) shutting hard- to computing servers [11. and operating energy efficiency and replication techniques inside data cen. however. 2 current systems [20] and allows multiple virtual machines highlights relevant related works on energy efficiency and (VMs) to share the same physical server. GreenCloud [19]. To address this gap. as the average load of the system often stays below of replication strategies for data centers have been proposed 30 % in cloud computing systems [11]. and network switches. 2 Related works sumes 30–50 % of the total power used by the IT equipment [10]. the communication network con. For better reliability and high perfor. Power-aware networks • Optimization of communication delay to provide quality were first studied by Shang at el. our con. the first of user experience for cloud applications. corresponds to the dynamic voltage and frequency scaling tem bandwidth and data availability between geographically (DVFS) technology [12]. Specifically. reliability and energy consumption. Similar to DPM and DVFS power man- delay of cloud applications. ware components down or (b) scaling down hardware per- The performance of cloud computing applications. The effect of DVFS is limited. A large number mode. [10]. nique for cloud computing data centers which optimizes energy consumption.

which significance is determined with a set of weights Most of the cloud applications. t is current time.1 Cloud applications agement. tion in cluster of data grids is proposed. rely on tight interaction with databases. several data replication A cost-based data replication in cloud datacenter is pro- approaches have been proposed. utilization of communication resources and energy to guarantee the availability. Furthermore. a newly created data has the highest backbone network. thus exposing it to a single usage patterns and analysis of data popularity [25]. ter as well as between geographically distributed data centers. clustering of user locations and deploying replica closer to uted platforms and offered globally. lying communication costs both in terms of the energy and and (b) the optimization target. while data. resources can be replicated at redundant loca. For better reliability and the centroid of each cluster. an energy efficient data replication scheme for dat. which takes into account sys- network bandwidth. On top of that. ming and determines optimal points of replication based on a newly posted YouTube video attracts most of the visitors. 3. the popularity is not constant multiple data centers to minimize power consumption in the over time. Typically. Then. but not inside data centers. It periodically collects information from the cluster heads.2 Data replication location of replicas for each data object is determined by periodically processing a log of recent data accesses. and bandwidth in datacenter systems. Underutilized storage servers uted cloud computing system which supports replication of can be turned off to minimize energy consumption. a long-tailed gradual decay: ters. posed in [18]. builds a reliability model. To ensure data availability and reduce access delays for a file is computed in relationship with the access fre. several infrastructures. related to the traffic load. This approach is based on linear program.28] and social net- fic demand can bring significant power savings. Then. However.Cluster Comput 2. replica creation time points ing single point of failure. the authors suggest replication strategy across period of time. Then. demand. (1) Another optimization of data replication across data cen- ters is proposed in [17]. are required The approach presented in this paper is different from all to maintain data replicas. efficiency. such as storage devices and networking devices. The migration from an old site performance. dynamic data replica. The aim is to minimize data access where a0 is a maximum number of access events recorded delay by replicating data closer to data consumers. Cloud computing enables the deployment of immense IT replica site is determined by employing a weighted k-means services which are built on top of geographically distrib. However. The policy maker social networking. as the time passes it starts to lose popularity and Since power consumption of aggregation ports is linearly the audience [26]. Several studies of HTTP requests [27. to a new site is performed if the gain in quality of service of tions and using redundant infrastructures. an optimization based on the traf. For example. This involves an under. such as online office and selected according to the age of the reading. Data center infrastructures consume sig. To address expo. migration (communication cost) is higher than a predefined nential increase in data traffic [22] and optimization of energy threshold. are determined from data storage reliability function. This approach cre- ates a policy maker which is responsible for replica man. the cost of electricity differs at different geographical locations [24] making it another parameter to 3 System model consider in the process of data replication. In [8]. This work works [29] suggest using a power law distribution which has focuses on replication strategies between different data cen. network bandwidth and communi- nificant amounts of energy and remain underutilized [23]. quency of all other files in the system. In this section we present a model of geographically distrib- acenter storage is proposed. cation delay to define the employed replication strategy. ularity is measured as a number of access events in a given In [16]. To achieve load balancing. The model focuses on the performance of cloud appli- keeping one of the replica servers for every data object alive cations. the data center traffic demands and popularity of data objects. new replicas need replication approaches discussed above by (a) the scope of to be synchronized and any changes made at one of the sites data replication which is implemented both within a data cen- need to be reflected at other locations. the number of replicas location. further determines the popularity of a file based on the access Data queries can be fulfilled either locally or from a remote frequency. Moreover. This solution follows Database replication decisions can be based on the data a centralized design approach. This approach analyzes data storage failures Maintaining replicas at multiple sites clearly scales up the and data loss probability that are in the direct relationship and performance by reducing remote access delay and mitigat. and k is a coef- 123 . a(t) = a0 t −k . the access rate decays over time. Optimal after content publication. The pop- point of failure. In [15]. Underutilized resources can be exploited without additional costs. tem energy consumption. data replication can be used.

As soon as the database reply is received. Figure 2 presents a plot of Eq. and dupdate is the time required to update database. At the time (t) end of the execution. For ρ = 1. 1 Cloud computing datacenter Central DB Wide-Area Network (WAN) Datacenters 2. which consists of three layers of net- 123 . N Datacenter DB Core Network Aggregation Network Access Network Rack Rack DB Central DB Access Datacenter DB Access Rack DB Access a0 every data item they access. the update rate can be expressed as a fraction of 3. Cluster Comput Fig. the Fig. The access rate can be obtained as the first derivative from where dr eq is a time required for the workload description to Eq. it queries a database and waits for the database reply to arrive. the workload execution is started.31]. As a result. It begins with the user request arrival at the data- center gateway. modify them with some probability updating the database. 1] controls relation between the access rate ra (t) Fig. dexec is a da ra (t) = . Large-scale cloud computing systems are composed of geo- ru (t) = ρ · ra (t). (2) workload execution time which is defined by the size of the dt computing work of the workload and computing speed of the Whenever cloud applications access data items. (1). ddb is a one-way communi- cation delay between the server and the database. cloud applications modify three tier fat tree [31]. The data- base querying delay corresponds to the round-trip time and depends on the database location. arrive at the computing server. For this. (4) type of the content [27]. The most widely used data center topology is the and the update rate. some workloads will send a modified data item back to the database for the update. After being scheduled it is forwarded through the data center network to the selected computing resource a(t) for execution. while for ρ = 0 these modifica- tions is never performed. they can server. …. 2 Data access distribution total delay associated with the workload execution in data- centers can be computed as follows: ficient typically in the range [30. At the server. Figure 3 presents the timeline of a workload execution in data center. (3) graphically distributed across the globe data centers (see where ρ[0. Therefore.2 Cloud computing system architecture the access rate. which depends on the ddc = dr eq + 2 · ddb + dexec + dupdate .(1). 1). 3. the workload can request data item if it is needed for its execution.

In addition. about two-thirds of its peak power consumption. A module called replica manager (RM) is located at the Eq. In a similar fashion. relation from f . in and out datacenter. It periodically analyzes data access statistics to which can adjust operating frequency when server is under- identify which data items are the most suitable for replica. utilization level at which the server attains asymptotic. (b) minimize utilization of network bandwidth and higher capital investment and operation costs. I/O resources and other peripherals operational even when base (Datacenter DB). (6) 2 the database and updated in all the replica sites.4 Storage requirements of data replication ered cloud computing system in terms of energy consump- tion. where server. the power consump- frequently used data items from the Central DB. While accessing data items cloud applications can    Ppeak − P f i xed l3 modify them. The access rate (or popular. (6) forms the basis for DVFS power management Central DB. which is used to replicate the most no computations are performed. 3. 05]. To account of it. request or forwards it up to the Central DB. aggregation and access. each data center hosts a local database. to the fact that servers must keep memory modules. (5) and update statistics makes it possible to project data center and (6).3 Energy consumption of computing servers layer is where the computing servers.35]: rack-level database (Rack DB). This implies a cubic maintained for each data item. a  [0. the rack. The power consumption of a server depends on its CPU uti- Central database (Central DB) is located in the wide-area lization. Voltage statistics showing the number of accesses and updates are reduction requires frequency downshift. and a is a requested data or forwards the request to the Datacenter DB. Rack DB either replies with the consumed at the peak load.2. and the datacenter is stored. When data is queried. These modifications have to be sent back to Ps (l) = P f i xed + 1 + l 3 − e− a . i. To estimate 123 . are arranged in racks. The core layer (c) minimize communication delays encountered in the data provides packet switching backplane for all the flows going center network. Eq. an idle server consumes network and hosts all the data required by the cloud applica. The availability of access Figure 4 plots server power consumption given in Eqs. Then. which can result in sumption. Ps (l) = P f i xed + 1 + l − e− a . One of the pitfalls of data replication is the increased usage The objective is to (a) minimize system-level energy con.Cluster Comput Fig. Ppeak power rack-level database server. physically attached to the network. The aggregation layer integrates con- nections and traffic flows from multiple racks. the V is voltage and f is an operating frequency [36]. of storage and networking resources. In addition. The access 3. tion scales with offered CPU load according to the following each rack hosts at least one server capable of running local equation [34. 3 Workload execution Database End of workload timeline request execution Start workload execution Request enters to data center Done dreq ddb dexec dupdate Scheduling and Database query Workload execution Update database sending to server work switches: core. the information about requesting CPU power consumption is proportional to V 2 f . (5) All database requests produced by the cloud applications 2 running at the computing servers are first directed to the where P f i xed is an idle power consumption. disks. which is used for subsequent   Ppeak − P f i xed  l  replication from the Datacenter DB. tion and at which replication sites. The following subsections present a model of the consid. bandwidth usage and energy consumption. called datacenter data. utilized to conserve operational power consumption [10]. usage of network bandwidth and communication delays. (5) can be rewritten as ity) is measured as the number of access events per period follows: of time.e. As reported in [32–34]. For most of the CPUs. l is a server load. This is due tions. To speed up database access and reduce access latency. the Datacenter DB either satisfies the close to linear power consumption versus the offered load.

the bandwidth is used for sending job . (7) per rack [30]. capacity C [G B] and N data objects stored in the cloud sys- An availability of per-server bandwidth is one of the core tem each of the size Si [G B]. compression. while the consumption of network ports can scale uplink flows are those directed from the computing servers with the volume of the forwarded traffic as follows: towards the core switches. network bandwidth is used for propagat- r =1 ing database requests and when applications need to update modified data items: where Pchassis is a power related to switch chassis. As a result.6 0. all communications inside data center can be sumption of switch chassis and line cards remain constant broadly categorized to the uplink and downlink types. 1] is a port utilization which can be defined as item. there is no need to keep links operating at the peak rates. and T is a measurement interval. nrp is number of ports operating at rate r size of data request. the number of nodes required for most widely used three-tier fat tree topology (see Fig. The over time. This further reduces the per server bandwidth components contribute to the switch energy consumption.2 0. its uplink bandwidth appears to be applying a number of advanced data replication techniques. in Sect. (8) In the uplink. An aggregation switch offers 12 ports to the access 3.9 1 Server load 3. down to 277 Mbps for fully loaded connections. oversubscribed by a factor of 48 · 1G / 20 G = 2. assume a uniform storage database for different replication strategies. a rack switch serving 48 servers C each connected with 1 Gb/s link has only two 10 Gb/s links System storage requirements can be further reduced by in the uplink. while between bursts links remain idle. the power con. link rates can No DVFS DVFS enabled be adapted to satisfy long-term traffic demands using IEEE 802.1.1 0. Downgrading port rate is especially useful as almost all of the links are never fully utilized for a long duration. the oversubscription ratio at the aggrega- port transceivers. 4 Server power consumption In this section we analyze network capacity of data centers and bandwidth requirements of cloud applications that access the involved storage requirements. Another bandwidth multiplexing occurs at the aggregation layer.5 0. 0 0. n c is number of Bul = Nser v Ra Sr eq + Ru Sdata .7 0. For the three tier architecture with 8-way Equal Cost Multipath Rout- Network switches are hardware devices which consist of the ing (ECMP) [30].3az Energy Efficient Ethernet (EEE) [50] standard.3 0. Plinecar d is   the power consumed by a single line card. line cards. Several studies characterize energy consumption of network According to the model of cloud applications described switches [40–42].38]. 3. Conversely. According to [43] and [44]. and switch chassis [39]. follows: In the downlink. All these tion layer is 1. R Pswitch = Pchassis +n c ∗ Plinecar d + nrp ∗ Ppr ∗ u rp . which such as data deduplication. the downlink flows are those from the core switches to the computing servers. and erasure-coding also limits the per server available bandwidth to 416 Mb/s.4 0.5. Then.6 Bandwidth model Fig.8 0. Ppr is a power drawn by a port where Nser v is the number of computing servers. The for each data object is ri . When idle. Typical link utilization is only 3–5 % [48]. and Sdata is the size of the updated data and u rp ∈ [0. aggrega-  N tion.5 Energy consumption of network switches layer and is connected to all the core layer switches.4. (10) line cards plugged into switch. Each port can be configured to operate at different rates P idle and its power consumption scales exponentially with the Power [Kwh] increase in the transmission rate [45–47]. Instead. and access switches as well as the number of servers i ri Si Ns = . Ra and Ru are data access and update rates respectively. Network packets arrive in bursts [49]. Cluster Comput P peak where B p (t) is an instantaneous throughput at the port’s link at the time t. [37. C p is the link capacity. 1) maintaining the cloud data is given by: imposes strict limits on the number of hosted core. Sr eq is the running at rate r . if the replication factor requirements affecting design of modern data centers. For example.

.

(9) T t Cp T ∗ Cp t database objects and propagating data item updates between 123 . t+T 1 t+T B p (t) 1 descriptions to computing servers for execution. receiving up = dt = B p (t) dt.

(8)). E s = Ps (l) · (2 · ddb + dexec ) · Ra · T. data equal to 1 Gb/s. the residual bandwidth l and in the uplink R l available at each capacity for each segment (access. Ra is an average For a three-tier data center with Nser v servers. update between Datacenter DBs and Rack DBs Taking into account the model of cloud applications (Sect. S K L M agg For the wide-area network it corresponds to the update E dc = Es + E kcor e + El + access Em . and (16) define bandwidth scription at lower layers. and data updates. Based on the aforementioned and by adapting tier of the network can be obtained as follows: Eq.3) and network switches (Sect. while Eqs. and Br ep is the bandwidth obtain total energy consumption of data center IT equipment required to update all the replicas. Port utilization and traffic volumes are pro- aggregation and core tiers respectively. Cagg and Ccor e are the capacities at the access. database access rate. 3. (14). (8). the energy consumption of the access switches can l = BC l − B . (14) from replica databases. (13) the workload execution and database query delays and can where Ndc is the number of Datacenter DBs and Nrack is the be obtained as follows: number of Rack DBs in each data center. (20) between Central DB and Datacenter DBs s=1 k=1 l=1 m=1 Br ep. Caccess is portional to the size of job descriptions. E m access are the energy consumptions of while for the network inside data center it corresponds to the k core. (5). (15). as propagation delay inside datacenter is in the order of micro seconds. Rdl dl dl be computed as follows: (17) l = BC l+1 − B . dexec is the workload execution time. Equations (10) and (11) allow com- links in modern datacenters. (see Eq. as follows: Br ep is different on different segments of the downlink. (21) Now. BCaccess Nser v At any moment of time the residual bandwidth left not in (22) use in the data center can be computed as follows:  cor e agg access  where P f i xed corresponds to the power consumption of the Rdl = min Rdl . If data query is satisfied BCaccess = Nser v · Caccess . Nacc access.4). Nser v /Naccess is the number of  cor e agg access  servers per rack. Ppaccess and Bdl /BCaccess are power con- Rup = min Rul . cor e) is an index indicating a tier Naccess BCaccess level. El . ddb becomes smaller. (15) becomes a job of the network to deliver the update after com- puting server becomes available for executing other tasks. It will be different for every tier of the data center network required to query and receive a data item from the data- due to bandwidth oversubscription involved. having computed the bandwidth required by running applications and their data base interactions.dc = Nser v · Nrack · Ru · Sdata . aggregation. ·Pp · · · T. BCcor e = Nagg · Ncor e · Ccor e . Commonly. agg while E kcor e .1). The delay associated with the database update is not included as it BCagg = 2 · Naccess · Cagg .7 Database access and energy consumption   Bdl = Nser v · Ra · S job + Sdata + Br ep . (19) sumption and port utilization of an access link. The delay ddb depends on the database location network capacities at each tier can be obtained as follows: and employed replication strategy. and m access switches respectively. Rul ul ul  Nser v Bdl E access = P access f i xed + · Ppaccess · +2 where l ∈ (access. Rdl . base. and core) of in the downlink Rdl ul the network. Rul . data requests. a workload obtained according to Eq. while Cagg and Ccor e correspond to 10 Gb/s traffic. Sdata is the size servers (Sect. the corresponding execution. while Pp agg 123 . (12) where E s is the energy consumed by a computing server s. we can of the requested data object in bits. Rdl . and T is a total time of the workload Nagg aggregation and Ncor e core switches. 3.wan = Nser v · Ndc · Ru · Sdata .Cluster Comput data replicas: 3. Therefore. energy consumption depends on the amount of traversing traffic and utilization of network ports where Caccess . puting traffic requirements in the uplink and the downlink The uplink capacity is always limited due to over sub- respectively. (11) Having the model of energy consumption for computing where S job is the size of the job description. (16) For network switches. l aggregation. Rul . ddb is the time ity. we can obtain where Ps (l) is a power consumed by the server executing residual bandwidth by subtracting it from the network capac. The expression l + 1 refers to the tier located above the  agg Bul Naccess tier l. (18) switch chassis and line cards. agg. 3. the load of individual servers becomes proportional to Br ep.

as well as the aggregation and access switches are 10 Gb/s. Being proportional to both links.42 In this section we perform evaluation of the system model Computing server 301 developed in Sect. interconnected by 4 core and 8 aggregation switches. Table 1 Datacenter topology tion of an aggregation network link. small portion of 3–15 % is consumed by the port transceivers. power consumption of communication ports can be adjusted The values for power consumption are derived from [52] and in network switches based on the load of the forwarded traf.024 servers arranged into 32 racks The bandwidth consumption is typically low in the uplink. and only a of their CPUs with the offered computing load. The update requests can be large in size. With DVFS. 3. peripheral slots. peak CPU consumption [51] and 171 W consumed by other 4. they are sent only at the fraction of the access rate. Energy consumption of network switches is almost con- ment DVFS [12] and DPM [11] power management tech. Similarly. and power supply unit [34]. The propagation by the size of the data items and the data access rate (see Sect. 3. aggregation and access segments consumption of 301 W is composed of 130 W allocated for a of the datacenter network requiring replication.3 μs where 2 · Naccess /Nagg is the number of aggregation switch links connected to racks. Parameter Value Similarly. servers can scale power consumption is consumed by switches’ chassis and line cards. Parameter Power Consumption [W] Chassis Line cards Port 4 Model evaluation Gateway. we assume a uniform distribution of jobs among the consumption of an idle server is bound and corresponds to computing servers as well as traffic in the data center net. 1. mother board.3 μs. aggregation switches 1558 1212 27 Access switches 146 – 0. while Ppcor e and Bul /BCcor e are the power consumption and port utilization of a core network Table 2 Power Consumption of Datacenter Hardware link. The main performance indicators are: data center energy consumption. fic. The DPM technology allows enabling a sleep mode in idle servers and switches.3 μs E cor e = f i xed + Nagg P cor e · Ppcor e · · T.2 Bandwidth consumption Table 1 summarizes data center setup parameters.5 for details). The net. 3. stant for different transmission rates as 85–97 % of the power niques. Both computing servers and network switches imple. the minimum Fig. disks. Having only 123 . The server peak energy ing capacities of the core. the size of a data item and the update rate. [53]. The uplink is used for sending database queries and database work links interconnecting the core and aggregation switches update requests. 198 W. core. fan. work. (23) BCcor e Gateway link 100 Gb/s. 3. point to the datacenter through a gateway switch.1 Setup scenario devices like memory. 50 ms Core network link 10 Gb/s. which is Figure 5 presents the downlink system bandwidth require- connected to all the core layer switches with 100 Gb/s. The topology is comprised of 1. (24) BCcor e Access network link 1 Gb/s. the energy consumption of the aggregation and core switches can be computed as follows: Gateway nodes 1  Core switches 4 agg Naccess agg Bdl E agg = P f i xed + 2 · · Pp · + Ncor e · Aggregation switches 8 Nagg BCagg Access (rack) switches 32  Bul Computing servers 1. 4. delay of all these links is set to 3. As the only component Considering three tier data center architecture presented in which scales with the load is the CPU power. available network band- width and communication delay. the bandwidth Table 2 presents the power consumption profiles of data consumption grows fast and easily overcomes correspond- center servers and network switches. There is only one entry 3.3 μs   Bdl Aggregation network link 10 Gb/s. the required bandwidth is mainly determined servers to the top-of-rack switches is 1 Gb/s. 50 ms ments with no database updates.024 Ppcor e · · T. The bandwidth of the access links connecting computing In the downlink. However. Cluster Comput and Bul /BCaccess are power consumption and port utiliza.

8 1 erable when the size of data access is large and frequency of Server load data updates is high. data to be acknowledged. as while waiting for database data to arrive. 7 Energy consumption of computing servers DB replication high data update rates can exceed the capac- ity of the gateway link. the energy con- and access switches.03 Ra=0. Figure 6 shows bandwidth required for propagating 150 replica updates in the downlink from Central DB to Data- center DB and from Datacenter DB to Rack DBs.4 1200 Ru=0.04 1.2 0.02 1200 Ra=0. sumed by the computing servers as well as core.Cluster Comput 1.5 3 6 12 24 48 96 Data size [MB] Fig. while doing computing work as well els of network equipment.8 Acess Network 1024Gb/s Access network 1024 Gb/s 1050 Ru=0. data size larger than 125 MB 300 will cause congestion in the access segment of the network clearly indicating the limits. Datacenter DB tion from Central DB to the Datacenter DB in order to avoid Central DB 600 the bottleneck. remain always server load. Finally. the energy delay transmissions while waiting for previously transmitted consumed by IT equipment is composed of the energy con. The result suggests that devising 123 .000 Ra=1.400 1350 Ra=0.6 0. Similarly. 3. for the case of Datacenter Fig. Unlike in the case of computing servers.0 Bandwidth [Gb/s] Ru=0. 6 Bandwidth required for updating replicas 100 Gb/s at the gateway link would trigger replication even for the small data items of less than 12 MB (or 8 Ethernet 750 Rack DB packets) for the access rate of 1 Hz requiring data replica. The load under DVFS power saving. the bandwidth of Energy [Kwh] 450 the aggregation network of 640 Gb/s will be exceeded after 78 MB and will require additional data replication from Data- center DB to Rack DBs. 7. Energy consumption of the computing sumption of network switches is less sensitive to variation in servers is presented in Fig. Figure 8 reports the obtained energy consumption lev- during both phases.01 Ra=0. such as switch chassis and line cards. 3 for details).2 Ru=0. The minimum the bandwidth is consumed at both the core and aggregation querying time corresponds to the round-trip communication layers. In particular. while other hardware com- obtained energy consumption increases with the increase in ponents.4 0. The servers execute cloud appli. It is mainly due to the fact cations which perform a certain amount of computing job and that only port level power consumption scales with the traffic make a single database query for successful completion. The charts reveal that even if replication is used and the data access is 0 localized the burden on network bandwidth becomes consid.05 Bandwidth [Gb/s ] 900 800 Aggregation Network 640Gb/s 750 Aggregation network 640 Gb/s 600 600 400 Core Network 320Gb/s 450 Core network 320 Gb/s 200 300 Datacenter gateway 100 Gb/s 150 0 15 30 45 60 75 90 105 120 135 150 0 Data size [MB] 1. The bandwidth provided by the core network of 320 Gb/s will be exceeded with data items larger than 40 MB for the access rate of 1 Hz. the amount of forwarded traffic. aggregation. However. in real systems communication 4. 5 Downlink bandwidth demand Fig. This is due to the fact that energy is consumed active.3 Energy consumption delays are larger and are the subject to queuing delays on congested links and protocol-related procedures which often According to the model presented in Sect.6 Ru=0. When updating data on Rack DBs. 0. delay between the computing server and the database (see Fig.

results presented in previous sections. In addi- width demand in the upcoming intervals using models pre. RM computes data replication functionality. which includes the consumption of both the servers and network switches. Ns2 platform [54] for TCP/IP network simulation. A module called RM located at the Central DB periodi. 3. 5 Replication algorithm To ensure energy efficiency and performance of cloud appli- cations. and the downlink residual bandwidth. To achieve consistency with modeling suitable candidates for replication. GreenCloud is a cloud com- access and update rates in previous intervals and makes an puting simulator which captures data center communication estimate for their future values. Generally. data access becomes localized improving performance of cloud applications. while for the Rack DB replication the size of 120 MB can be achieved as the downlink traffic becomes restricted to the access segment of the network.6. network switches and the datacenter network are monitored to determine the most communication links. is maintained. it processes at the packet level. we propose a replication algorithm which takes into account power consumption and bandwidth required for data access. along with the 6 Simulation results number of requests. In addition. and based on the historical observations of the access frequency data can be replicated to the Datacenter DB and Rack DBs. and rack ID. tion. The only exception is the gateway link. For all replication scenarios. congestion levels in consumed by computing hardware. datacenter ID. Figure 9 reports the tradeoff between datacenter energy consumption. Due to the core network saturation. the core layer reaches saturation earlier. the maximum size of the data segment in Central DB and Datacenter DB scenarios is equal to 39 MB. Every data object is available at the Central DB. Cluster Comput power saving modes that shut down entire hardware compo- nents of a switch would allow substantial savings. For every access to meta data information. the benefit of Rack DB replication is two-fold: on one hand net- work traffic can be restricted to the access network. which has lower nominal power consumption and higher network capacity. the simulation scenario 123 . As a result. GreenCloud offers fine-grained modeling of the energy sented in Sects.5 and 3. which available bandwidth remains constant for Datacenter DB and Rack DB replication scenarios since data queries are processed at the replica databases and only data updates are routed from Central DB to Datacenter DB. It is based on the widely known becomes possible to compute energy consumption and band. For performance evaluation purposes we developed the cally analyzes this meta data information to identify which GreenCloud simulator [19] and extended it with the required data objects need to be replicated and where. residual bandwidth decreases for all network segments with increase of the load. while on the other. it has to be noted that applying such kind of approaches will affect network connectivity and may result in system perfor- mance degradation as datacenter load cannot be accurately predicted. being the smallest of the datacenter network segments with the capacity of 320 GB/s. It is important to note that the relative variation of the consumed energy is much smaller than the drop in available network bandwidth. However. which includes data object ID. With these parameters.

3. The duration of each simulation is 60 minutes. and c Rack DB 600 80 80 Energy [Kwh] Energy [Kwh] replication scenarios 450 60 450 60 300 40 300 40 20 150 150 20 0 0 0 0 0 0. 20 The main metrics selected for performance evaluation are (a) energy consumption at the component and system levels.6 0. The second is that energy consumption decreases as as a scheduling decision is taken for a newly arrived work.6 0. as well as the 30 replication threshold vary in different simulation runs.8 1 Access rate [s -1 ] (c) Rack DB 123 .6 0. network bandwidth 5 and access delay characteristics of the system.6 0. The reason is that communication delay is included Fig. 0 0. The sizes of data items.2 0. can be observed in the obtained results. data access and update rates. it is sent over the data center network to the selected tions.4 0.8 1 0 0. 15 (b) network bandwidth and (c) communication delay. The first trend is that The workload generation events are exponentially distrib.2 0. data become available closer to the computing server loca- load. energy consumption increases with the increase in the data uted in time to mimic typical process of user arrival.4 0.2 0. The size 40 of a workload description and database queries is limited to 35 1.1 Effect of data size Access rate [s-1] Fig. DNS power sav- E n er g y [ K w h ] 25 ing scheme is enabled both for servers and switches in all simulation scenarios. 9 Energy and residual 96 700 96 Residula bandwidth [Gb/s] Residual bandwidth [Gb/s] bandwidth for a Central DB.500 bytes and fits into a single Ethernet packet.Cluster Comput 45 Central DB Datacenter DB Rack DB server for execution.4 0. The workload execution and data query- ing follow the timeline diagram presented in Fig. Two trends Table 1 and energy consumption profiles from Table 2. As soon size. 10 The following subsections report the effect of data size and the update rate on energy consumption. Each server accesses one data item every was selected accordingly using topology setup presented in 300 ms and makes no updates of the database.8 1 6. b 600 Datacenter DB. 8 Energy consumption of network switches Figure 10 presents the measurements of energy consump- tion of computing servers for data item size varied from 10 to 40 MB.2 0.8 1 -1 Access rate [s ] Access rate [s -1 ] (a) Central DB (b) Datacenter DB 96 600 Legend Residual bandwidth [Gb/s] 80 Gateway Energy [Kwh] 450 Core 60 Aggr 300 Access 40 Energy 150 20 0 0 0 0.4 0.

aggregation Fig. 12a) and Datacenter DB 10 20 30 40 (Fig. It should be 0 noted that for a Central DB (Fig. Fig. As expected. Figure 12 shows the downlink bandwidth requirement 60 for different data sizes. 12c). delay becomes smaller for replicas located closer to servers Energy consumption of network switches scales similarly and for all the replication scenarios an increase in the size of with the increase in data size (see Fig. as data traf- Energy [Kwh] fic mostly remains constrained in the access part of the 90 network. 10 Energy consumption of servers and access) of the network. the reported bandwidth is Data size [MB] consumed at all the segments (gateway. 11 Energy consumption of 30 network switches 30 25 25 Energy [Kwh] Energy [Kwh] 20 20 15 15 10 10 5 5 0 0 10 20 30 40 10 20 30 40 Data size [MB] Data size [MB] (a) Central DB (b) Datacenter DB Legend 15 Access Energy [Kwh] Aggr 10 Core Gateway 5 0 10 20 30 40 Data size [MB] (c) Rack DB 123 . which prevents servers to enter into the sleep mode. which is the effect of the exponential arrival of the incoming tasks. but age time elapsed from the moment of sending data request can be reduced by shortening round-trip times to the data. core. network. The band- 30 width slightly varies with the time. access base. 11). and having the requested data arrived. These Figure 13 reports data access delays measured as an aver- delays become larger with the increase in data item size. the access is localized and the reported bandwidth is consumed only in the access part of the into the execution time of the cloud application (see Fig. 3). 12b) replication scenarios. 120 the consumption can be reduced considerably. For the rack replication (Rack DB). The consumption data objects increases data access delay. Cluster Comput Central DB with no replication (Central DB) is higher than for other repli- 150 Datacenter DB cation scenarios as all the layers become actively involved Rack DB into traffic forwarding. Bandwidth demand remains high for large data sizes for all replication scenarios. while for the Rack DB repli- cation (Fig.

replicat. As expected. DB and from Datacenter DB to Rack DBs are propagated. 14. energy consumption grows with the increase in the update rate due to longer awake The simulation results presented above show that for cloud periods.3 Hz and different update rates. Switches at all layers are involved into forwarding applications. b Datacenter DB. database updates from Central DB to Datacenter delays speeding up execution. For the update rate equal to the access rate.8 energy consumption of the system. In the uplink. we kept the size of the data item of 6 MB and access rate of 0. While in the energy. 123 . conserve bandwidth and minimize communication downlink.2 ing servers as the role of the servers is just to send modified data item to the Central DB at the end of the workload exe- 0 cution.4 send updates to the database. update rate variations do not affect energy consumption of comput- 0.6 the interval [0. and c Rack 35 30 DB replication scenarios Bandwidth [Gb/s] Bandwidth [Gb/s] 30 25 25 20 20 15 15 10 10 0 10 20 30 40 50 0 10 20 30 40 50 Time [sec] Time [sec] (a) Central DB (b) Datacenter DB 45 Legend 40 10MB 20MB Bandwidth [Gb/s] 35 30MB 30 40MB 25 20 15 10 0 10 20 30 40 50 Time [sec] (c) Rack DB 10MB 6. 12 Bandwidth 35 40 consumption for a Central DB. cloud applications modify every accessed data item and 0. that perform database updates rarely.3 Hz fixed while varying the number of updates requested by the cloud applications in 0. database update traffic.Cluster Comput Fig. Ra ] . they forward replica ing data closer to the computing servers always helps to save updates sent from the servers to the Central DB.2 Effect of data update rate 1 20MB 30MB To better understand the impact of database updates on 40MB Data access delay[sec] 0. 13 Data access delay switches for fixed data size of 6 MB and access rate of 0. As reported in Fig. Central DB Datacenter DB Rack DB Figure 15 presents energy consumption of network Fig.

as data updates 20 begin to propagate from Datacenter DB to multiple Rack DBs. Cluster Comput Central DB base updates for 32 Rack DBs. both core and aggregation networks carry data.6 0.6 0. It reports the downlink bandwidth consumption at Energy [Kwh] the core and aggregation layers for the scenario when each 60 accessed data item needs to be modified and updated.4 0. replica update traffic remains at a very low level.2 0.6 0. only employed in the production environments. 40 However.4 0.4 0.2 0.6 0. for data and for database updates. 15c). 16. The following forwarding. replication strategies.8 1 -1 Update rate [s ] (c) Rack DB 123 . Fig.8 1 0 0.8 1 -1 puting servers. as soon as Rack DB replication is enabled. Their perfor- the gateway and core switches are involved into update traffic mance can be measured using a set of metrics. the bandwidth usage increases to over 35 Gb/s. Update rate[s ] Fig.4 0. The access switches serve Datacenter DB 100 Rack DB both kinds of traffic. This underlines a tradeoff between energy and band- 0 width consumptions when replicating data closer to the com- 0 0.2 0. While in the case of Rack DB replication (see are the most commonly used metrics for evaluation of data Fig. The effect of data update on network bandwidth is shown 80 in Fig. 14 Energy consumption of servers 7 Performance comarsion with existing systems There are many different data replication strategies widely In the case of Datacenter DB replication (see Fig. When data is accessed from either Central DB or Data- center DB. 15 Energy consumption of 30 30 network switches 25 25 Energy [Kwh] Energy [Kwh] 20 20 15 15 10 10 5 5 0 0 0 0. 15b). which justifies their higher energy consumption.2 0.8 1 -1 Update rate [s ] Update rate [s-1] (a) Central DB (b) Datacenter DB 30 Legend 25 Access Energy [Kwh] 20 Aggr Core 15 Gateway 10 5 0 0 0.

such as availability of network bandwidth. An unexpected failure in storage infrastructure or dat. depending on the access pattern. opti- optimal number of replicas to ensure both availability and the mization of communication delays leads to improvements in QoS of cloud applications. data center network required for replica updates. we assume that 8 Conclusions and future work every data object is permanently stored at the Central DB and in addition. the proposed replication always keep the response time 15 within the boundaries required by environment [57. approach implements a dynamic replication to only maintain optimizes energy efficiency of the system. Maintaining optimal 25 number of replicated data objects minimizes network load required to keep all replicas up to date and network response 20 time.1 Availablity monitors datacenter gateway traffic load to induce repli- cation of data objects with higher access frequency. Luxembourg in the framework of available network bandwidth and properties of the imple. objects are replicated either at the Datacenter DB and/or Rack ability. The simulation results (see Fig.58]. on the other side. DBs. it is replicated This paper reviews the topic of data replication in geographi- in Datacenter DB and Rack DBs. unlike in several other methods [55. data objects that are frequently accessed are replicated Bandwidth [Gb/sec] which reduced total number of replicas. In the proposed replication approach. ing from National Research Fund. It extends a preliminary version of this work which has been published in [59]. data access and update rates. Sec- 30 ond. Therefore. The proposed replication approach takes into account the Acknowledgments The authors would like to acknowledge the fund- tradeoff between data size. ECO-CLOUD project (C12/IS/3977641). 10 Central DB Datacenter DB 7. To overcome this problem the proposed replication technique 7. data objects are replicated in Rack DBs closer to 35 computing nodes and hence response time is reduced. of replicas should be selected carefully as excessive replica- Future work on the topic will be focused developing a tion may increase the associated costs and traffic load in the testbed implementation of the proposed solution. for this data and greatly improves overall system perfor- can reduce energy consumption. i. cloud applications. bandwidth usage and com- mance. acenter blackout could cause unavailability. quality of user experience of cloud applications.Cluster Comput 40 mented data center topology to make optimal decisions. Any failures in Datacenter cally distributed cloud computing data centers and proposes a DBs can be recovered from Central DB and vice versa. efficiency and communication processes in cloud computing Bringing and maintaining the data closer to the servers where data centers [19].. 123 . The obtained simulation results (see Fig. the simulator focusing on energy reduced data access response time for cloud applications. The obtained results confirm that replicat- applications are executed significantly decrease access time ing data closer to data consumers. 7. novel replication solution which in addition to traditional per- over. the number and location munication delays substantially. Data One of the main goals of replication is to assure data avail.3 Datacenter congestion 5 0 10 20 30 40 50 60 Network congestion can cause significant degradation of Time [sec] data center performance and one of the most sensitive to Fig. copies of data objects are maintained on redundant to indicate congestion. More. However.56]. 13) indicate that. To resist these The residual bandwidth at datacenter gateway can be used effects. In addition. probability of failures in data center components or storage infrastructure.e. 16 Downlink bandwidth with high update rate congestion points is data center gateway which is a bottle- neck handling both the incoming and outgoing traffic. confirm that replication is an effective tool to control traffic acenters. Rack DB First. the availability is usually measured by a load at the data center gateway. the proposed formance metrics.2 Response time The evaluation of the proposed replication solution is based on the developed mathematical model and simula- Another important reason for data replication is in the tions using GreenCloud. 9) infrastructures and in several geographically distributed dat.

. Shengquan. Yuan. B. Liu..M. 3(3). F. Rashmi.: Under- 13. In: Workshop on ization of a tiled architecture processor with on-chip networks. M.: Tech titans building boom.. Adamic. Liao. In: IEEE/ACS International Con. K. on Cloud and Service Computing (CSC). A. Pelley.. T. A. Li.H. (2008).. S. Sherwood. 19–22 March 1880 (2011) (2012).. replication strategy in data grids. ization layer”. Liu.: CA-DAG: communication-aware directed acyclic 35.U. CA. Abdelzaher. (2013). In: IEEE Inter.. NiciraNetworks (2009) 709–736 (2013) 22. Buyya. (2011). 13–23. P.. Weber. 34. and pareto .G. Chang. Emerg. Y.: Dynamic volt. D. Yang. In: 16th International Workshop on Quality of 8. ACM SIGCOMM 38. Koomey.6. Parallel Comput. 40–54 paper_c11-481360.S. 661–703 (2009) (2008).. 2012 IEEE International. J.. pp.. (2011).. and PaaS”. VanGilder.S. pp.: Green IP over WDM and compression techniques in cloud design. 229–238. D.. APC White paper No. Newman.: GreenCloud: a packet-level simulator of energy-aware cloud computing data centers. Wentzlaff. Chen.. 66–71..A. pp.: Market-oriented cloud com..: eStor: energy efficient ference on Measurement and modeling of computer systems. 414–421. H. Xue: Powersleep: ity data center network architecture. Electricity Information 2012. Circuits Syst. Santa Clara. Al-Fares. International Conference on Network Applications.J.R. 1–7.: Energy character. Wenisch. pp. Peh. D. International Symposium on Low Power Electronics and Design. A.: Zipf. Parulkary. Horvath. Shang. and reality for delivering IT services as com. J. Yeo. C. 91–102.. Ivaturi. tecture (HPCA).: Determining total cost of ownership for data center 5. Rasmussen.H.: Growth in Data center electricity uses 2005 to 2010.. Commun. S. and network room infrastructure.: A dynamic weighted data Service. Kuang. no.F.. 20. Bouvry.: A novel cost-effective dynamic data on Hot Topics in Storage and File Systems.. Khan. K. N. R&D Lab. Qin. 39(11). 17. and resilient data center storage. In: ACM SIGOPS European Workshop: 12. Katz. (2012). 37.Y.A. May. S.: Worldwide electricity used in data centers. N. J. 444–458 (2007) 33. IaaS.. M. R. Cheng. Liu. Glottometrics 3. A. 123 ... Li.. Gibby. Systems Conference networks with data centers.: A scalable. age scaling in multitier web servers with end-to-end delay control. tions in empirical data. In: Conference on puting (DASC). R. (2009). 26. 31. J.. (2003). In: Ninth 30. Dale. S. (2008).-K. (2008). costs in hosting centers. X. G. Shalizi. Meisner. In: Proceedings of 5th USENIX conf. L. Liu. Realtimepublishers.-D. Nov (2011). Szabo. 63–74 a smart power-saving scheme with sleep for servers under response (2008) time constraint. Gu.: Managing server energy and operational USA.: Trends in social National Laboratory. G. Bellosa. S.: Statistics and social network of Pike Research. (2010).. 56(4). pp. Dalian.. 18.. “Cloud Computing Energy Efficiency. Wu. S.A. mance distributed computing systems. 143–150 (2012) ter energy efficiency: public law 109–431.. Yapy.... 36. Katz. Hassan. Li.. Lett. Yang. In: IEEE International Conference on High Per. A.K.a 9. et al. Oakland (2011) 25. W. (2000). Analytics Press. Vahdat. In: Inter. Appenzellery. 1861– (SysCon). Cisco. S. Chernicoff. Kliazovich. Hwang. 424–427.5 Design Guide International Symposium on High-Performance Computer Archi. C. Wang. L. (2005)..: A Solution to the network challenges of data recov- tional Conference on Distributed Computing Systems Workshops ery in erasure-coded distributed storage systems: a study on the (ICDCSW). Casado.. D.S.. Loukissas.. Venugopal. J. Bouvry.U. In: Energy Efficient Design (WEED)..org/power_ssj2008/ national Conference on Cloud Computing (CLOUD). Opportunities for Data Centers Utilizing SaaS. YouTube videos. S...E.T. Ganesh. In: 5th International AAAI Confer- 10. 46(2). 51.W. J. A. Autonomic and Secure Com. R.. X. H.. (2003). K. Huberman. China. B. El-Gorashi. New York (2009) formance Computing and Communications (HPCC). T. Cisco Visual Networking Index: Forecast and Method- 3. A. X. March Res. Kim. and Services. 289– 32. Elmirghani. In: ACM International Symposium on 14. Server Power and Performance characteristics. Barroso. Shah. 2011–2016.. 28. IEEE Spectr. A. Environ.: Flowvisor: a network virtual- 2. S.. Cisco Data Center Infrastructure 2.: A survey on resource allocation in high perfor... Cluster Comput References 19. P. F. J. pp. 29(12). 27.spec. Y. 37–42. 1(3).: Greening the switch. J. (2013). R. Zomaya..P.: Dynamic replication algorithm in data grid: 7. Brown. Das. pp. Hussain. Lin. Chen. J. H. X.. M. Hayes.H. T.. Sel.: The shortcut guide to data center energy efficiency. 39. national Conference on Dependable. pp. Comput. available on http:// graphs for modeling cloud computing applications.E.: Cloud computing. McKeowny. Proto- Assessment of Energy Savings and Carbon Emissions Reduction cols. C. Asur. J. puting utilities. N. San Diego. G. C. SIAM Rev. Upadhyay. Jian-Jia. Ram- Towards optimal data replication across data centers. USA replication strategy for reliability in cloud data centres. Borthakur D. Lightwave Technol. power-laws.K. X. ACM 51(7).. Cisco Systems. Kliazovich. Madi M. Technical report...: The benefits of event driven energy accounting in 298 (2011) power-sensitive systems. links for power optimization of interconnection networks.. W.com/en/ (2008) US/solutions/collateral/ns341/ns525/ns537/ns705/ns827/white_ 4..V. Tchernykh...: Power-law distribu- ference on Computer Systems and Applications. Berkeley (2008) media: persistence and decay. Balihalli. Vabbalareddy. pp. D. Jun. hype. In: Interna. warehouse-sized computer. L. Stanford University. D. IEEE J. Chang. D.R. N. pp.B. Top. 7.: Deduplication 16. Clauset.pdf (2012) (2009) 23. McConnell.. B. 21.. 1–6. Wang. ranking tutorial. R. Koomey. chandran K. White paper http://www. 5–13. 9–11 ology. Y. 034008 (2008) (2005). R. Rao. Taylor.: Report to congress on server and data cen. W. Pecero. (2008). www. C. Miller. Inc. Deutsche Telekom Inc.. facebook warehouse cluster. International Energy Agency.. Jha.A. Strategic and Tactical survey. L. comput. Berkely. X. 496–502. Lawrence Berkeley 29. Ping.. M.. In: 2011 International Conference 303–314. pp.H. pp. Skadron. et al. R. In: ACM SIGMETRICS international con- 15. IEEE Trans.: 38. (2007). beyond the PC: new challenges for the operating system. T. G. Huberman. vol.: Power provisioning for a pp. 6. super- 1.. standing and abstracting total data center power. Khan. Power aware computing and systems. Wang. 24. Fan. 62(3). S. Dong. Q. commod- 11.. P.S. 366–371.. (2011).. pp. Mag.. 1263–1283 (2012) puting: vision. J. B.B.cisco.. Computer Architecture.: Dynamic voltage scaling with ence on Weblogs and Social Media (2011).

Sanchez-Macian. Vishwanath. Boru. pp. Rabanillo.. J. Luxembourg.: Data networks are lightly utilized. nologies from the University of tial evaluation of energy efficient ethernet. (2012). pp. Int. Ranganathan. Chansler. S. Computing Systems. In: 8th International Dejene Boru received his M. In: IFIP International Networking Workshops (SUN. Berlin (2011). zovich is the author of more than 90 research papers and Editorial Board USA. Kanemasa. Santos-Boada. Member of the IEEE Communications Surveys and Tutorials. He is a 57. Chandra.. Farrington... (2009). pp. Dr. (2010). 1–10.isi.. A.. The Network Simulator Ns2 http://www. He served as general and TPC chair in a number 56. European Research aware network operations.. J. Netw. Math..: On the selfsimilar winning Ph... from University of Trento (Italy).. A. istry of Education. Banerjee. Y.Sc IFIP-TC 6 Networking Conference. In: 6th USENIX Symposium on Networked Systems Design cations. Li. School of Information Technolo- ture in the age of merchant silicon.A. Wilson. In: IEEE INFOCOM workshops. S. D.. 378. F. 578–580 (2011) ing scholar at the Computer Sci- 51. K.T. Con- G. Maestro. D. Technology..: Energy mainly from the IEEE Communications Society. 98–104. Scott. and will stay that Science.. K.: The Google file system. Bouvry.: An energy consumption from Addis Ababa University model for energy efficient ethernet switches..: Characterizing 2008 to Aug. and Com- way”.. D. and Applications (GC13 WS - 42. Cisco Corpora. (1999).. CCSNA). Sohan. Syst. R. Kliazovich is a holder of a large number of scientific awards.: Profiling per. 29–43 (2003) (GreenCom) in 2010.... A. 7(6). Zhang. P. In: Reliable Dis.. D. M. IEEE Trans. Reviriego.. J.W. Dr. research interests are energy effi- 46. Trento (Italy). Aachen. Gupta.. V.. 37(5). NET(Italy) since Jan. Sole-Pareta.: cient cloud computing. Klia- Mass Storage Systems and Technologies. 1–5.. P. many (2009)..com/p/en_US/products/server/processor/ xeon5000 sity of California at Los Ange- (2010) les (UCLA) in 2005 and at the 52. In: IEEE Conference on Computer Communications (INFO.Y. Moore. 2409– SET).. Mahadevan.: The Hadoop of high-ranked international conferences. Jie. V.: An ini. Careglio. Hairong. Graduate Assistant at Dire Dawa pp. Jinli. In: Proceedings the 26th Symposium on tional Conference on Cloud Networking (CLOUDNET 2014).. S.. He was a ference on High Performance Computing and Simulation (HPCS). S.edu/nsnam/ns/ computing environments received Best Paper Award at the IEEE/ACM 55. 53. Jayasinghe. Christensen. 2013. K. Networks. He holds an award- 49. Sivaraman.. Intel Xeon Processor 5000 Sequence. Q. You... Zhao. Rev. pp. 2012 IEEE 31st Symposium on. S. He is work- 10 Gbps network interface energy consumption. P. Mahadevan. May 3–7. A. Jian: An energy-effective adaptive replication strategy networking. Appl.. Rice. Kharitonov. Z. pp. San Jose (2006) arch Fellow at the Faculty of 48. Palmieri. 43. and software defined and Implementation. Granelli.. J. Reviriego. pp. IEEE Commun. energy efficient communications. (2011). Vishwanath.. networking. 268–271. 365–380. 59. in cloud storage system. Agarwal. Technical Report Center for Discrete Mathematics & Theo. U.: A power benchmarking framework for network devices.383. Willinger. Wan... P. Ghemawat.: Improving the Energy Efficiency of Networks: A Focus on Ethernet and End Devices. S. Radia. 2010... Z. C. M. 331–336. He was a visit- 5(15). available at: http:// ence Department of the Univer- www... University (Ethiopia) from Nov. Russell. pp. A. Wang. Kawaba. Dzmitry Kliazovich is a Rese- tion.A. D. performance network element design. J. Odlyzko. Huang. J.. In: 17th IEEE symposium on gies of the University of Sydney high performance interconnects (HOTI ’09). Berkeley. Vahdat. cal and Computer Engineering A. Inf. ters. D. USA (2013).: coordinator and principle investigator of Energy-Efficient Cloud Com- Response time reliability in cloud environments: an empirical study puting and Communications initiative funded by the National Research of n-tier applications at high resource utilization. P. Fund of Luxembourg. Bahl. W. Consortium for Informatics and Mathematics (ERCIM) and Italian Min- 1–6. Shvachko. (2009). Fiore. in 2013 and B. International Conference on Green Computing and Communications ACM SIGOPS Oper. E. including the IEEE Interna- distributed file system. P. 291–300.. H. Mansley. Sivaraman.. Ranganathan.. P. in Information nature of ethernet traffic. 45. A. In: IEEE 35th ing as a researcher at CREATE- Conference on Local Computer Networks (LCN). Lett. 93–102. P. (2012).. vol. Zhao. GA. Jiang. Incline Village. 47. Atlanta. C. 58. W. in 2014. R. 8–11 Oct. Leland. NV. Russell. F. J. Ricciardi. K. pp. K. USENIX Association networking. Nordman. 2419 (2013) 41.: Data center switch architec. Sci. P. wireless Somniloquy: augmenting network interfaces to reduce PC energy indoor localizations and appli- usage. Banerjee.. Taqqu.. In: IEEE Globecom 2013 International Workshop on Cloud COM) Workshops. 795–808. Xindong: Zhou. pp.D. cloud computing. Leung. Hodges.. A. pp.. Ger. R.: packet and per-byte energy consumption in the NetFPGA Gigabit Energy-efficient data replication in cloud computing datacen- router. R. S. (2009). (2009). 123 . Sharma. Li.. Christensen. (Ethiopia) in 2008.: Time-domain approach to energy efficiency: high. Y. His work on energy-efficient scheduling in cloud 54. Gobioff. 2(1). and next-generation no. His main research activities are in the field of tributed Systems (SRDS). Kliazovich.Sc in Electri- 44. In: International Con. 1–15 (1994) and Telecommunication Tech- 50. Springer.. His (2010). P. IEEE GLOBECOM Work- shops.: Analyzing local strategies for energy-efficient gfeng. B. Sharma. Zomaya.Cluster Comput 40. N. Rubow. Intel Inc.intel. munication of the University of retical Computer Science ACM. Maestro..

He is also the he was visiting professor at the State University of Campinas (Brasil). ACM Transactions on 500 scientific papers and articles Modeling and Computer Simulation. Professor Bouvry co- ciate Professor at the Dept. elected head of the Interdiscipli- nary Laboratory for Intelligent and Adaptive Systems (http:// ilias.Sc. Professor Zomaya’s research interests are in the areas of parallel and distributed computing and complex systems. the IEEE 2008.lu). In August 2004. Director of the Centre for Dis- He is author or co-author of more than 140 papers with topics related to tributed and High Performance networking. He is Founder and General tor of more than 20 books. He received the «Laurea» (M. Achievement Award (2014). co-author or edi- Systems. the ACM tion Links and Networks (CAMAD’06 and IEEE CAMAD’10). Italy. Chair 2009–2010) of the IEEE ComSoc Technical Com. Pascal Bouvry has academic and industry experience on Soc Distinguished Lecturer for 3 continents working as manager and executive in major companies the period 2012–15. working in the School of Infor- versity of Genoa. wireless communi. Professor and smart grid communications. August 2010 and April 2013. Granelli was guest-editor of ACM Zomaya published more than Journal on Mobile Networks and Applications. Computing which was estab- cations and networks.lu). in 1997 mation Technologies. degree (’94) in Computer Sci- ence at the University of Greno- ble (INPG). 2009 and 2012. uni. and Hindawi Journal of Computer and is author. cognitive radios and networks. with focus on performance modeling. Zomaya is currently in Electronic Engineering and the Chair Professor of High Per- the Ph. He obtained his Ph. green networking lished in late 2009. His research at the IMAG laboratory focussed on Mapping and scheduling task graphs onto Distributed Memory Parallel 123 . Analysis. such as. He is Vice-Chair of the First International Conference on Wireless Internet the Editor in Chief of the IEEE (WICON’05) and General Chair of the 11th and 15th IEEE Workshop Transactions on Computers and Springer’s Scalable Computing and on Computer-Aided Modeling. Computers.D. He is Computing Surveys and Journal of Parallel and Distributed Computing. authored over 200 publications in international peer-reviewed venues. Spacebel. of Information Engineering and He also organized numerous top conferences worldwide and is member Computer Science (DISI) of of several editorial boards. uni. Metasolv. He is a Chartered Engineer. Scalable Computing (2011). zine. Cluster Comput Fabrizio Granelli is IEEE Com. he is deputy head computing. on Parallel Processing Outstanding Service Award (2011). Vice-Chair Technical Committee on Scalable Computing Medal for Excellence in 2007–2008.D.uni. The Uni- and 2001. and Design of Communica. and Asso. Reliability and Performance Modeling” in the years 2007. Current interests of Professor Bouvry encompass cloud and grid From 2008. Professor Zomaya is the recipient of the IEEE Technical Committee tions QoS. serves as an associate editor for 22 leading journals. He was officer (Secretary 2005–2006. Pascal Bouvry is Professor at the University of Luxembourg. a Fellow of and Associate Editor of IEEE Communications Letters (2007–2011). including IEEE Cloud Computing Maga- the University of Trento (Italy). versity of Sydney. optimization issues. Networks and Communications. Lat45. His group (http://pcog. Dr. director of studies for the certificate “Smart ICT for business innovation” and Faculty member of the SnT (http://snt.) degree Albert Y. France. formance Computing & Net- tions Engineering from the Uni. such as FICS. IEEE. respectively. AAAS. TPC Co-Chair of IEEE GLOBECOM Symposium on “Communica. and the IEEE Computer Society Technical mittee on Communication Systems Integration and Modeling (CSIM).lu) is also managing the HPC facility of the University of Lux- embourg. SDC. IET (UK). and related complex problems. in Telecommunica. of the academic council in Infor- mation Engineering.