Network-AwareDataPlacementStrategyinStorage ClusterSystemdownloads.hindawi.com/journals/mpe/2020/5970583.pdf · transmission time interval, the transmission waiting queue (retention

Research ArticleNetwork-Aware Data Placement Strategy in StorageCluster System

Bilin Shao1 Dan Song 1 Genqing Bian 2 and Yu Zhao1

1School of Management Xirsquoan University of Architecture and Technology Xirsquoan 710055 China2School of Information and Control Engineering Xirsquoan University of Architecture and Technology Xirsquoan 710055 China

Correspondence should be addressed to Dan Song 674355101qqcom

Received 3 September 2019 Revised 13 January 2020 Accepted 10 March 2020 Published 21 April 2020

Academic Editor Laurent Dewasme

Copyright copy 2020 Bilin Shao et al is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

e dramatic increase of storage devices in distributed storage cluster system and the inherent characteristics of distributeddeployment mode make network resources become one of the bottlenecks of data storage process By analyzing the functionalcharacteristics of data flow transmission of the components of the storage system the network topology structure of the storagesystem is constructed and the evaluation index of node flow load is put forward based on the degree centrality and the be-tweenness centrality theory of the network to explore the network topology and real-time flow characteristics According to theevaluation index of node flow load a network-aware data layout scheme is proposed By balancing the flow load of bottleneck linkcongestion and transmission delay can be reduced to further shorten the total task execution time and improve the efficiency ofdata writing

1 Introduction

In the large-scale data storage cluster system (hereinafterreferred to as storage system) the shortage of network re-sources and the sharp increase in data flow are one of themain reasons for network congestion slow data transmis-sion and service response delay in the storage system Inorder to manage large-scale data access it is undoubtedly aneffective way to improve service response efficiency to ac-curately identify network characteristics and optimize datalayout while monitoring the state of the whole network

e network is the core support of the storage systemand the bridge connecting all devices In the data storagesystem all system resources (storage devices routersswitches etc) are connected to each other through networkdevices which jointly constitute the network topologyLarge-scale data storage systems are often built on a largenumber of cheap devices and a quantity of data streams willbe generated during the cooperation between device nodesA storage device carries out data stream transmission withother devices through network link Network bandwidth is akey index to measure the transmission capacity of the

system Because the network bandwidth is very limited it isnecessary to allocate the network bandwidth reasonably toimprove the network transmission capacity

Since the different network links and storage nodescarried out different data flows the transmission charac-teristics of data flow in the network should be fully con-sidered Data flow is introduced as the weight to researchnetwork topology characteristics for accurately analyzingand identifying the storage system network characteristicsDue to the storage node location degree transmission ca-pacity between data flow and link residual capacitytransmission time interval the transmission waiting queue(retention volume) the uplink input data flow and thedownstream output data flow of the network topology allthe factors are featured to evaluate the storage node loaddegree which is easy to deploy the data and provides a goodfoundation to solve the problem of network congestion offlow engineering [1] erefore in this paper firstly thetopology is built based on the characteristics of storagesystem to extract the statistical characteristics of the networkdata flow en according to the network degree centralityand median centrality the data node flow load evaluation

HindawiMathematical Problems in EngineeringVolume 2020 Article ID 5970583 16 pageshttpsdoiorg10115520205970583

index is put forward Finally a data layout scheme is pre-sented based on network conception which treats the lessnetwork load node as the target position to reduce con-gestion waiting time lower the delay of data transmissionand enhance the efficiency of the storage

2 Related Work

21 Research Progress of Data Layout in Storage SystemIn recent years research on data layout optimization mainlyincludes three aspects computing power storage power andbandwidth optimization Part of the research focuses on theoptimization of physical resources (CPU memory andstorage) by improving the performance of storage devices toenhance data access rate and data center efficiency Anotherpart of the research is based on the performance charac-teristics of the data center network topology to optimize thedata layout and improve the efficiency of data transmission

(1) e research on the data placement solution ofstorage system in the cloud environment mainlyfocuses on node evaluation cost tradeoff linearprogramming content dependency and other as-pects Node evaluation of target node placement isbased on comprehensive consideration of the cur-rent disk space load conditions size of the availablestorage CPU ability memory processing ability diskIO communication bandwidth traffic flow re-sponse rate and its historical access record andfailure record Node evaluation usually selects theappropriate node placed data file and its replicasaccording to the key attributes such as dynamicnumber of replicas and data popularity [2ndash5] etradeoff between transport overhead storage over-head and user access performance [6ndash8] enables togain maximum performance with minimal over-head Based on this the linear programming methodis adopted to minimize the system cost [9ndash12] andreduce user response time and network load [13] byadding the consideration of network overhead Anoverhead tradeoff layout solution can result in a goodperformance of load balance but the transfer time isnot optimal due to the damage of data dependencyPlacing data blocks with high dependence in thesame data center can reduce the transmission timesacross the data center as far as possible and furtherreduce the consumption of network bandwidth andthe system delay responding to the task request in thesystem [14 15] Aiming at multilayer data center amultilayer topology structure is constructed from theperspective of network planning and a networkbandwidth model is established to localize networktraffic and reduce the communication overhead inthe upper layer network switch thus ultimately re-ducing the overall traffic across the data center andreducing the network consumption of the cloud datacenter [16 17]

(2) In other application environments such as distrib-uted storage system [18ndash21] grid computing [22]

online social network [23] peer-to-peer (P2P) net-work [24 25] software-defined network (SDN)[26ndash28] content distribution network (CDN)[29 30] and big data network storage [31 32] theresearch on data layout management is of greatvalueIn the distributed storage system the network awarerepair framework based on the dependency betweendata storage demand and available bandwidth[18ndash20] can find the data repair scheme with theminimum bandwidth cost in the dynamic networkand realize the load balance of storage and networktraffic Hedera [21] is a scalable dynamic flowscheduling system that schedules a multistageswitching structure adaptively to utilize the resourcesof the aggregate network effectivelyIn grid computing the network-aware QoS work-flow scheduling method [22] takes network char-acteristics and task dependence into account so thatcan reduce the completion time and workflow ex-ecution cost and improve the task success rate andresource throughput simultaneously According tothe characteristics of user interaction in social net-works the data placement method combining socialgraph division and data replication [23] divides usersinto a number of communities and further trans-forms the problem into a community-server sec-ondary distribution problem according to thenetwork topology of data centersIn P2P networks it is also an important direction tointroduce the concept of replica population andapply knowledge of population ecology to solve datalayout [24] e distributed topology-aware un-structured P2P file caching architecture [25] canreduce the transmission traffic on the trunk networkby caching hot data and reducing excessive cachingof nonhot dataIn SDN on the one hand analysis of network real-time large data set to predict the future demandand realize the network traffic intelligent man-agement decision [26] and on the other handevaluate the status of network real-time by cal-culating the link bandwidth delay and packet lossrate to make route decisions dynamically whichcan effectively realize load balance schedulingaccording to the estimation of traffic flow and linkutilization ratio [27 28]In CDN energy efficient delivery model (EEDM)[29] based on multicast tree can improve thescalability and uniform distribution of data storageto different degrees e learning automataadaptive probabilistic search algorithm based onfuzzy support [30] makes use of the local topologyinformation and current state of the cooperativenodes provided by the existing fog nodes and findsthe point-to-point and point-to-fog minimumjumpers by running the distributed adaptive en-hancement algorithm

2 Mathematical Problems in Engineering

In big data network storage system the continuousand uniform data striping layout method based onfragment label [33] and the discrete multireplicaspatial data layout scheme based on graph coloringtheory [34] can improve the scalability and uniformdistribution of data storage

(3) e optimization of virtual machine layout in thecloud environment has important inspiration for theresearch of data layout e network-aware layoutstrategy adopted in the virtual machine layout in thecloud environment focuses on the traffic demand ofthe virtual machine and takes into account thequadratic and real-time variability of traffic as wellas the network topology and routing scheme [31] Bysearching for the optimal bandwidth between av-erage throughput and peak throughput computingand network resources are allocated in a way thatbalances resource utilization efficiency and pre-dictability of performance [32 35] which solves theproblem that the general network-aware VM layoutscheme lack consideration of the optimal bandwidthallocation e two-stage virtual machine placementalgorithm of network awareness [36 37] dynamicallyperceives the stability of the physical host accordingto the node centrality and the aggregation coefficientand appropriately aggregates virtual machines by thesimilarity which improves the network communi-cation capacity and reduces the network traffic be-tween different data centers

In summary the data storage layout needs to takenetwork resources into account significantly and the net-work characteristics of the data center (topology trafficcharacteristics etc) have an important impact on the per-formance of the data layout

22 Application Scenario and Main Contributions of thePaper Different data blocks of the same file in the storagesystem are often distributed and stored in different nodes ondifferent racks ere are several storage nodes on each rackand the nodes within the same rack are connected by Top-of-Rack (ToR) switch and the nodes between different racks areconnected by core network switch as shown in Figure 1Data transfer between intrarack nodes relies on ToR switchand cross-rack data transfer depends on core network switchdeployed in storage systems

In storage systems the link from the core network switchto the rack is the main network bottleneck [38 39] Eachstorage node in the storage system network can initiate datatransmission through ToR to a storage node on the samerack or to a node on another rack with the core network

At present although cloud providers are deploying alarge number of computing and storage devices to meet thegrowing demand for computing and storage resourcesnetwork resource demand is becoming one of the key factorsfor performance bottlenecks In the storage system networkuneven flow distribution is easy to lead network congestionand especially flow load imbalance between bottleneck links

is a major cause of network delay erefore in networkstorage system according to the characteristic of data storagenetwork and complex network theory the network flowdistribution model is established and network flow concen-tration degree index and node centrality index of storagesystem are put forward to identify the characteristics of thenetwork flow implement the effective control and balancenetwork flow between multipath which have very vital sig-nificance on reducing congestion and transmission delay

In view of the current situation of insufficient bandwidthallocation research and optimization in data layout con-sidering the key role of network bandwidth in the storagesystem this paper starts with the network topology structureand flow transmission characteristics and puts forward theevaluation index of node flow load and the data layoutscheme of network awareness Firstly according to thecharacteristics of the storage system network the datatransmission between nodes is divided into cross-racktransmission and intrarack transmission and the data centernetwork topology is established Secondly by analyzing thereal-time characteristics of the network topology in thestorage system the importance and load status of the nodesin the network topology are perceived from four indexesnode strength node capacity centrality data quantitytransmitted by the nodes and concentration index of dataflow of node en the network topology characteristics areconstructed to establish the data layout Finally simulationexperiments are carried out to verify the superiority of thenew network awareness data layout strategy in the com-pletion time of transmission tasks

e contributions of this paper are as follows

(1) e evaluation index of node load considered onnetwork topology and real-time flow is proposede characteristic of cross-rack transmission andintrarack transmission is constructed Four charac-teristic indexes are proposed including nodestrength node transmission turnover node capacitycentrality and concentration index of data flow ofnode and the comprehensive evaluation index ofnode network load based on these fourcharacteristics

(2) e network awareness data layout scheme is pro-posede task is written according to the remainingnumber of data block father file e real-timecharacteristics of the network are sensed based onthe storage system network topology structure andcomprehensive evaluation index of node networkload to select target and place racks Considering thenodes network load and storage load the node isplaced in the target rack to finally complete the datalayout optimization of network awareness

3 Evaluation of Node Flow Load ofStorage System

31 Storage System Network Topology Construction and FlowStatistical Feature Extraction Network awareness is the real-time monitoring of all elements performance of the entire

Mathematical Problems in Engineering 3

network (network topology network equipment etc) and theprecaution and treatment of dynamic changes in network flowIn order to analyze the influence of network elements on theperformance of data layout and identify the nodes with heavyload and the key nodes and intervals of data flow transmissionthe attributes of key nodes and intervals should be consideredfrom two aspects network topology and the role of nodes andlinks in the process of data transmission

(1) Network topology is to map various devices of thestorage system to a node in the network e net-work architecture in the storage system determinesthe role and influence of each node and link in thedata transmission process and is an importantfactor to judge the real-time characteristics of thenetwork In general the main network devices ofstorage systems include core network switches ToRswitches and storage servers According to theconnection characteristics and transmission char-acteristics of these elements this paper constructs abrief network topology diagram as shown inFigure 2In order to facilitate modeling and simplify multi-level switch configuration it is collectively referred toas core network configuration In Figure 2 the nodein the central position represents the core networkthe dark gray node in the middle layer represents theoverhead switch and the light gray node on the edgerepresents the data storage serverAccording to the established network topology thenode abstract method is adopted to construct thedata center network topology diagram as G and G isexpressed as follows

G (V E) (1)

In formula (1) V represents the collection of all nodesin the network (routing nodes and storage nodes) andV is expressed as follows

V vi

1113868111386811138681113868 i 1 2 N1113966 1113967 (2)

E represents the collection of connecting edges betweenswitches or between switches and storage nodes V isexpressed as follows

E eij

11138681113868111386811138681113868 i j 1 2 N ine j1113882 1113883 (3)

A switch and a server correspond to node vi in Grespectively and the connection between the serverand the switch serves is seen as edge eij

(2) e role of nodes and links in the network duringdata transmission Various network devices(switches routers etc) in the network topology ofstorage systems play different roles in the datatransmission process and have different importanceAccording to the importance and ability of networkelements the strength and importance of its role indata transmission are identified and the data layoutis carried out dynamically to ensure the strongservice ability of core elements and improve theaverage utilization rate of common elements whichis crucial to improve the efficiency of the overallnetworkAccording to the constructed network topology thestatistical characteristics of network flow areextracted With each node vi as a unit all flow in-formation passing through vi is counted includingdata flow information starting from vi that ending invi and that passing through vi and current real-timetransmission rate and maximum data transmissionamount on each link A tuple is defined to reserve thedata flow information of each node (data amountinitiated by the node data amount received by thenode and data amount transferred by the node) and

Figure 2 A typical network topology of storage system

Core network

Rack 1 Rack 2 Rack R

Cross-rack linkIntra-rack link

Figure 1 Hierarchy in data center

4 Mathematical Problems in Engineering

node adjacent link information (link capacity andreal-time used capacity) e data flow informationof node vi is represented by DataFlowi

DataFlowi fsi fei fti1113858 1113859 MCijRTTij1113960 11139611113872 1113873 (4)

In formula (4) fsi represents the data amountstarting from vi fei represents the data amountending to vi fti represents the data amount tran-siting vi and vj is the node connected to vi that is forvj isin V eij isin E MCij and RTTij are respectively linkcapacity and real-time used capacity of node adjacentlink

(3) Calculation of node distance In the storage nodedistance calculation method of Hadoop distributedfile system (HDFS) it is stipulated that the distancebetween the same rack node equals 2 and the dis-tance between the cross-rack node equals 4 ispaper follows this rule and the distance betweendifferent types of nodes is calculated as follows

dij 2 vi vj in same rack

4 vi vj in different racks⎧⎨

⎩ (5)

In formula (5) vi and vj are different nodes in thestorage system that is ine j and dij represent thedistance between two nodes

32 Definition and Calculation of Storage System NetworkCharacteristic Indexes Node importance indicates thepivotability of a node in the network e higher the nodeimportance is the stronger the pivotability is and theheavier the flow load is In addition nonpivotability nodesare also loaded differently due to task preferencesereforeconsidering the influence of network topology and real-timeflow on node load this paper comprehensively evaluates theload degree of nodes from the two aspects of node im-portance and real-time flow

Firstly from the perspective of network topology thedegree of nodes intuitively reflects the importance of nodesin the network and the number of node capacity centralityreflects the pivotability of nodes in the whole network dataflow transmission process

However the importance of nodes in the whole networkdoes not fully reflect the amount that data carries Generallythe higher the importance of a node is the more the datatransmission tasks it carries and the heavier the load isHowever on the one hand in the actual transmission tasksdue to task preference the actual amount of data carried bynodes of equal importance will vary On the other hand it isthe timeliness of transmission tasks that is the amount oftransmission tasks carried by each node in different timeperiods varies greatly erefore the amount of datatransmitted by the nodes directly reflects the amount of datacarried by the nodes in the whole network data transmission

concentration index of data flow of node reflects the flowbalance of the nodes in a certain period of time and thenetwork characteristic indexes are defined and explainedaccording to the network topology structure constructed inthe previous section

321 Node Strength e strength of the weighted networkcentral node is defined as the sum of the weights of all theedges associated with the node For the storage systemnetwork carrying data transmission flow the strength of thenode is the sum of the data flow of the corresponding zonecross-section e calculation method is shown in equation(6) Node strength index mainly reflects the importance ofnodes from the local network

CS(i) 1113944jisinVi

wij (6)

In formula (6) wij is the cross-sectional data flow ofconnection node vi and vj

322 Node Capacity Centrality e node capacity cen-trality is the ratio of the sum of all the cross-sectional dataflow on the shortest path passing the node and the sum of allthe cross-sectional data flow on all shortest paths in thenetwork e capability centrality reflects the node pivot-ability to the whole network flow

In the storage system network capacity centrality ofnode vi not only counts the number of path passing throughnode vi for all shortest paths in the whole network but alsoassigns different weights to each shortest path namely thesum of the cross-sectional data flow on the path so as tomore truly reflect the capacity of nodes to carry data flowe calculation method of node capacity centrality is shownin the following equation

CC(i) 1113936stisinVinest 1113936eisinRst

Fe1113872 1113873 middot φi(st)1113960 1113961

1113936stisinVinest1113936eisinRstFe

(7)

In formula (7) Rst is the shortest path between s and t eis an interval of Rst and Fe is the sum of the data streams ofthe upstream and downstream sections of interval e In thispaper Rst is calculated by the Dijkstra algorithm as follows

Rst Dijkstra(s t) (8)

e calculation method of Fe is shown as follows

Fe 1113944ijisine

wij + wji1113872 1113873 (9)

φi(st) is calculated by formula (10) which is based on therelationship between vi and Rst

φi(st) 1 i isin Rst

0 i notin Rst1113896 (10)

323 Amount of Data Transmitted by Node e amount ofdata transmitted by node vi in the storage system networkrefers to that multiplied by all data flows through node vi

Mathematical Problems in Engineering 5

with the corresponding transmission distance e calcu-lation method is shown in formula (11) e data amountindex of node transmitted mainly considers the importanceof node in topology from the two aspects of data flow sizeand data transmission distance

CT(i) 1113944iisinV

fi middot di (11)

In formula (11) fi is the data flow through node vi anddi is the transmission distance of the corresponding data fi

mainly consists of three parts fsi is the data amount with thestarting point of node vi fei is the data amount with the endpoint of vi and fti is the data amount with vi as thetransition node dsi dei and dti are the transmission distancecorresponding to the transmission process and then for-mula (11) can be further transformed into the followingequation

CT(i) 1113944iisinV

fsi middot dsi + fei middot dei + fti middot dti( 1113857 (12)

e relation between fi and fsi fei and fti is shown asfollows

fi fsi + fei + fti (13)

324 Concentration Index of Data Flow of Node HHI is acomposite index to measure industrial concentration degreeis paper uses this concept for reference puts forward CDFindex (concentration index of data flow and the CDF index)of node vi and is defined as for a period of time squared as apercentage of the data flow that was passing on a node viCalculation method is as shown in the following equation

CDF(i) fi

F1113888 1113889

2

(14)

In formula (14) fi is all data flow passing through nodevi in a certain period of time which is calculated by formula(13) and F is the total amount of network transmission in thesame period of time which is calculated as follows

F 1113944iisinV

fi (15)

When all data are transmitted by one node the data flowaggregation coefficient CDF(i) of that node is equal to 1When all nodes are carrying the same amount of datatransmission CDF 1N2 e more data amount a nodecan carry the greater the CDF

325 Node Flow Load Comprehensive Evaluation Index(CEI) e previously defined node strength CS reflects thenetwork node important degree under the different data flowstates e node capacity centrality (CC) reflects the dataflow capacity that the node loaded e node transmissiondata amount CT reflects the importance of the node in theentire network data transmission Concentration index ofdata flow (CDF) of node reflects node flow balance status fora certain period of time In order to facilitate the

comparison a comprehensive evaluation index (CEI) wasdefined and the above four indexes were integrated tocollectively judge the importance degree and flow load statusof the nodes Since the dimension of each index is differenteach index variable data are firstly standardized and con-verted into dimensionless values of CSprime CCprime CTprime and CDFprimeand then they are given weights λ1 λ2 λ3 and λ4 respec-tively e calculation method of CEIi is shown in the fol-lowing equation

CEIi λ1CSprime + λ2CCprime + λ3CTprime + λ4CDFprime (16)

Different networks focus on different needs thereforethe appropriate weight value is chosen to meet differentneeds For example to fully evaluate the significance of anode in the entire network then λ1 λ2 λ3 λ4 To evaluatethe core position of node in the transmission of data flow inthe whole network the coefficient λ3 of CT such asλ3 gt λ1 λ2 λ4 is increased to achieve the comprehensiveranking of all nodes in the whole network meeting themanagement requirements In addition there are threemethods to determine the weight subjective weightingmethod (such as expert survey method and hierarchyanalysis process) objective weighting method (such asprincipal component analysis method entropy method andmultiobject planning method) and combined weightingmethod (ldquomultiplicationrdquo integration method and ldquoaddi-tionrdquo integration method)

33 IndexApplicationandResultAnalysis For CEI proposedabove the topology structure containing 64 nodes is taken asan example for testing and the corresponding topologystructure is shown in Figure 3

e data transmission task quantity was set as 500 filesand the data flow through each node was countedAccording to the corresponding formula the node strengthcapacity centrality data amount transmitted and concen-tration index of data flow are calculated Finally the com-prehensive evaluation index (CEI) is figured out and theresult graph is drawn

For the topology structure mentioned above differentamounts of data transmission task (DF 500) are producedIn four times the flow load on each link is extracted and at acertain moment each node data amount is detected efour indexes of each node are calculated as CS CC CT andCDF After normalization of data λ1 λ2 λ3 λ4 1 andCEI is figured out According to the load index value of eachnode the load of each link and the corresponding node isplotted as shown in Figure 4 e darker the node color isthe larger the size is indicating the heavier the load of thenode is Correspondingly the larger the link width is theheavier the load of the link at this moment is

As shown in Figure 4 the load of each node and linkvaries at different times e link load with dark color andlarge width is large and the color and size of correspondingnode is large that is the CEI value is large e CEI value ofthe node in the central position is always large indicatingthat the node plays a pivotal role in the network and carries aheavy load of data e CEI value of the node at the edge is

6 Mathematical Problems in Engineering

generally small since they are not responsible for forwardingdata flow and other tasks the importance of the node is lowand its value is mainly determined by the flow size of theassociated link erefore CEI proposed in this paper cancomprehensively reflect the node importance and flow loadcharacteristics in the storage system network topology

4 Network Awareness Data Layout Scheme

41 Design Target Generally most of the cross-rack linkload in a short period is more than two-thirds of the totalload of links which has increased the impact of a congestionlink If there is a data block of a file that needs to

(a) (b)

(c) (d)

Figure 4 Link load and node loads at different times in the 64-node topology of storage system (a) time t1 (b) time t2 (c) time t3 (d) timet4

Figure 3 A network topology of storage system with 64 nodes

Mathematical Problems in Engineering 7

communicate through congestion bottleneck link the datablock transmission progress will directly affect the entire filedata transmission completion time namely the duration ofthe file transmission is completed by the slowest subfile datablock transmission time

During data writing bottleneck links are almost alwaysthe hot spots Considering the load condition of the networklink in the storage system the location selection and writingof different data blocks cut from a file are independent andeach data block is determined separately erefore themain objectives of the data layout scheme design in thispaper are as follows

(1) Minimize the completion time of a single file eoptimal data block writing request sorting algorithmshould consider the number of remaining blocks inthe parent file of the data block e data blocks witha small number of remaining blocks should bewritten first to speed up the completion of thetransmission task of a single file

(2) Minimize the unbalanced load of the bottleneck linke optimal link selection algorithm should firsteliminate the load imbalance on the bottleneck linkand avoid too many transmission tasks concentratedon a small number of links that is the data to bewritten through the appropriate cross-rack link so asto minimize the transmission delay

(3) Minimize the unbalanced load of storage nodesAccording to the flow load and space load of thestorage node the optimal layout algorithm shouldselect the best target storage node for the arrivedwriting task so that the network load and space loadbalance effect of the storage node of the intrarack isoptimal

emathematical description of the data layout problemdiscussed in this paper is as follows

In the distributed cluster storage system suppose thereare a node set V that contains m data nodes V v1 v2 vm and a file set F to be written as F f1 f2 fk All k fileswill be stored in node set V and data layout strategy is toassign these k files to m data nodes that achieve the optimaltarget function

ree target functions are discussed in the networkaware data placement strategy

(1) Suppose the writing completion time of a single fileas Tsingle and Tsingle tfinish minus tstart where tstart is thestart time of the file writing process and tfinish is theend time of writing to the file It takes the least timefor completing a single file writing task with the leastvalue of Tsingle

(2) Bottleneck link load balancing can be measured bynetwork load changes in rack nodes Standard de-viation is appropriate for measuring the dispersiondegree of data it is consistent with the dimension ofdata so the load balance of the rack node can beexpressed by the standard deviation of load and usedas the standard to measure the load balance of the

system e smaller the standard deviation of theload is the better the load balancing ability is

e load balancing law of bottleneck link LV is defined as

LV

1113936mj1 (CEI(j) minus CEI)2

m minus 1

1113971

(17)

In formula (17) CEI is the average of system loadCEI (1m) times 1113936

mj1 CEI(j) and CEI(j) is the traffic

load of node vj(3) e load balancing of the storage node is denoted as

L e storage load of the data node Dj can be cal-culated by the sizes of files that are stored in itandL(Dj) is calculated with the following equation

L Dj1113872 1113873 1113944n

i1Sk (18)

In formula (18) Sk is the size of all files on Dj

Similarly the standard deviation of the storage node loadin each rack L(R) is used to represent the rack load balancee better performance of rack load balance is interrelatedwith the smaller L(R) e calculation of L(R) is shown asfollows

L(R)

1113936mj1 L Dj1113872 1113873 minus L1113872 1113873

2

m minus 1

1113971

(19)

In formula (19) L is the average of system load andL (1m) times 1113936

mj1 L(Dj)

erefore the objective optimization problem of datalayout can be represented by the mathematical model of thefollowing equation

minTsingle

min LV

minL(R)

⎧⎪⎪⎨

⎪⎪⎩(20)

42 Network Awareness Data Layout Strategy When thestorage system is writing data it first divides the data intoseveral data blocks of the same size and then the writing jobof a file is divided into the writing task of several data blocksTo get the best file writing efficiency it needs to optimize thecompletion time of each task e main goal of data blockwriting in the storage system is to increase the writing rate ofa single file by balancing the load on the bottleneck link tominimize the writing time of the data block e optimallayout algorithmmust allocate the best target location for theblock writing request to let it pass through the appropriatebottleneck link

In order to simplify the model the following assump-tions are made for the above analysis

(1) e size of the data block to be written is fixedAssuming all blocks are the same size the impact of

8 Mathematical Problems in Engineering

the data block size difference on writing time isignored

(2) During the writing of a single data block the linkstate is fixed Assuming that the link utilization re-mains stable for a short period of time it is easy to getthe bottleneck link utilization very clearly during theentire data block writing process

(3) e bottleneck link is easy to identify In the storagesystem the link between the rack and the corenetwork is often the easiest and is most likely tobecome the bottleneck link erefore this paperbelieves that the network bottleneck link is the link ofin and out rack that is the dark link in Figure 2

(4) Decision-making process of different data blocklayouts is independent ere is no impact be-tween the writing decision processes of the lastdata block and the next data block and they areindependent

On the one hand the network awareness copy placementscheme needs to be sorted according to the arrival of datablock requests on the other hand it needs to select ap-propriate links and target nodes so the scheme contains thefollowing three stages

(1) Sorting of data block writing requestse interval time between the two data block layoutsis set as the decision time of the writing requestsorting denoted as s e data block writing requestarriving in the s decision time is sorted according tothe number of remaining blocks in the parent file Toensure the speed at which a single file transmissiontask can be completed the smaller the number ofremaining blocks is the higher the ranking is Whens is equal to 0 it means that the layout scheme is anonline decision-making process without the sortingprocess which is processed directly according to thearrival order of data block writing requests

e s value of the decision duration time determineswhether there is the sorting process of data blocks to bewritten that is the data blocks to perform link selectionand allocation will affect the layout decision of datablock e larger the s value is the better the sortingresult will be obtained by the algorithm but at the sametime it will increase the writing time of the data blockerefore the value of s is a compromise process

(2) Evaluation and sorting of rack loadsIn Δt time interval the current load data of all cross-rack links are obtained Based on the evaluationindex in Section 32 calculate the comprehensiveevaluation index (CEI) of rack nodes and sorted racknodes by CEI e CEI is the basis for selecting thetarget rack Rack with the least CEI having low trafficload will be the preferred target rack

(3) Rack selection and storage node determinatione sorting result of load CEI of rack nodes calcu-lated in the previous stage is read to take the rack

with low CEI value as the target rack of data blockwriting request In the target rack according to theremaining space and flow load of the storage servernode two reachable server nodes with low load areselected as the target storage location

e process of network awareness data layout is shownin Figure 5 Each dotted box in the figure represents thespecific operation of each stage

e process of network-aware data layout strategy is asfollows

Step 1 determine the order of block to be writtenWhen the block write request arrives the decisioninterval s is firstly determined If sgt 0 the ordering ofwritten blocks is completed within the decision time sIn order to minimize the completion time of a singlefile written block needs to sort in line with the numberof remaining blocks in the parent file of the blockBlocks in the top with the least number of remainingblocks in the parent file which may shorten thecompletion progress of writing a single file If s 0block writing queen is sort by the ldquoearly come earlyservicerdquo principle to execute write operationStep 2 evaluate the rack node load Cluster manageraccording to the received link transports informationfrom each server node during Δt and updates the CEIvalue of rack node to maintenance load queue of racknode in timeStep 3 select the target racke cluster manager allocatesthe target rack for the block to be written e rack withthe least network load is evaluated as the minimum CEIvalue so the cluster manager chooses the rack with theleast CEI value present as the target rack During the Δttime interval rack node with a lower CEI value is chosenfor writing blocks and then the selected rack temporarilymoved to the tail of the load queue until workload queueis updated at the next Δt time updateStep 4 select the appropriate data node in target racke data nodes with less load are selected to place thedata block in accordance with the load degree of thedata nodes in the target rack Network load LL andstorage space load SL of data nodes in each rack arerequired e load of each data nodes in the rack FF (n)is calculated to choose the data node with the minimumload as the target node for block placement

43 Data Layout Algorithm of Network AwarenessAccording to the content and layout process of the threestages of the network awareness data layout strategy thecorresponding algorithms of the three stages are given be-low as shown in Algorithm 1ndash3 respectively

Algorithm 1 implements the sorting process of datablock writing task When s is equal to 0 the link selectionoperation is performed directly according to the arrivalorder of data block requests or the sequence is sortedaccording to the number of remaining data blocks in theparent file of the data block and the target rack and data

Mathematical Problems in Engineering 9

node are selected firstly for the data blocks with a smallnumber of remaining data blocks in the parent file

Algorithm 2 firstly obtains the CEI value of the nodeaccording to the above calculation method and selects therack with the smallest CEI Link utilization assessment usesthe information collected by the cluster manager (cluster

topology link load on the topology and machine failureconditions) to make decisions

e bottleneck link set Rr is composed of the links con-necting the rack and the core network in the topology CEIr isused to express the current congestion degree of the link ecalculation method of the CEI is described in Section 32

Begin

Data writingrequest arrives

Yes No

Calculate thenumber of

remaining blocksin the blockparent file

Sort by thenumber of

remaining blocksin the block

parent file fromsmall to large

Block writingqueue

Data block transmissionand writing

End

Choose target node withmaximum capacity factor

Calculate the capacityfactor of nodes

Calculate remainingbandwidth ratio of

nodes in chosen rack

Calculate remainingstorage ratio of nodes

in chosen rack

Choose target rackwith the lowest CEI

Calculate load of rack (CEI)

Get the cross-sectionaldata flow of cross-racklink at the current time

Node selectionrequest arrives

Begin

s gt 0

(1) (2)

(3)In order of

arrival

Figure 5 Network-aware data placement process

Input n nodes in rack Rr link load storage loadOutput data schedule queue Q

(1) Initialization D d1 d2 dm(2) if s 0 then(3) return LinkSelection(L)(4) end if(5) QaddToQue(D) add data block to queue(6) Qsort() Order by policy(7) for all data block d in Q do(8) return LinkSelection(L)(9) end for(10) end

ALGORITHM 1 Request schedule algorithm

10 Mathematical Problems in Engineering

e cluster manager receives link information from eachserver at regular intervals including the load status of eachlink in the bottleneck link set After receiving a single updatethe utilization rate of each potential bottleneck link is cal-culated by the cluster manager If the link information ismissing it is considered that the link is 100 utilized and hasno available capacity that is the load factor is 1 At this timetransmission tasks are no longer assigned to the link

Link updated time Δt decides the precision of the linkinformation Δt is smaller the link updates at the higherfrequency and the result is closer to the current actual loadBut if Δt is too small it increased the load of clustermanagement server transmitting links is paper uses thetypical values of storage system Δt 1 s [40]

Algorithm 3 calculates the load factor value of the nodebased on the storage load of each node in the selected rackand the link load from ToR to the node and selects the nodewith the minimum load factor value as the final placementlocation of the data block

e network awareness data layout strategy has a certaindelay Once the writing request of a data block is accom-plished the evaluation value of the current utilization rate ofall links involved in the transmission of the data block mustbe adjusted and updated in time to ensure the accuracy ofsubsequent layout decisions and avoid repeated decisionresults

Expansibility description of network awareness datalayout strategy this scheme can be used in combination withsome layout optimization strategies in the aspect of ensuringfault tolerance partition fault tolerance storage balance anddata reconstruction so as to achieve better performance Forexample since the scheme in this paper focuses on thebalance of flow load if this scheme is combined with thelayout scheme of storage balance it can theoretically achievebetter network balance performance while optimizingstorage load balance

44 Time Complexity of Network Awareness Data LayoutAlgorithm For a given data node set V with size m V= v1v2 vm file is set F to be written with size k F= f1 f2 fk and each file is divided into a number of data blocks towrite Suppose the number of individual racks is n and thenumber of data nodes in each rack is mn Firstly timecomplexity of sorting blocks to be written is the decision-making time s e time complexity of calculating and findingthe rack with the minimum load is O (n) and the timecomplexity of calculating load of data nodes in rack and findingthe data nodes with minimum load isO (mn) Maintenance ofblocks writing queue and selection of rack and data node isexecuted concurrently so take the worst time complexity of thetwo as the time complexity of the layout algorithm

Input L link loadOutput link utilization selected rack

(1) Initialization requireNR nr1 nr2 nrjW w1 w2 wj Fe F1 F2 Fj F f1 f2 fj d d1 d2 di Ftotalλ1 λ2 λ3 λ4

(2) for nr in NR do(3) CSnr 1113936jisinVi

wnrj

CCnr (1113936stisinVnrnest[(1113936eisinRstFe) middot φi(st)])(1113936stisinVnrnest1113936eisinRst

Fe)

CTnr 1113936nrisinVfnr middot dnr

CDFnr (fnrFtotal)2

CSnrprime CCnrprime CTnrprime CDFnr

prime uniformization (CSnr CCnr CTnr CDFnr)CEInr λ1CSprime+ λ2CCprime+ λ3CTprime+ λ4CDFprime

(4) end for(5) find the minimum CEInr(6) return rack nr corresponding to CEInr(7) end

ALGORITHM 2 Node load evaluation and selection algorithm

Input n nodes in rack Rr link load storage loadOutput the optimal node for placing one chunk

(1) Initialization Rr N1 N2 Nn FFn FF1 FF2 FFn(2) for each node in Rr do(3) SL(n) storage size of Ntotal storage capacity(4) LL(n) link capacity from N to TORtotal link capacity from N to TOR(5) FF(n) SL(n) + LL(n)(6) find the minimum FF(n)(7) return data node with minimum FF(n)(8) end

ALGORITHM 3 Node selection algorithm for link and storage load balancing

Mathematical Problems in Engineering 11

erefore the time complexity of the network-awaredata placement algorithm is expressed as follows

T Max ks Okm

n + kn1113896 1113897 (21)

45 Functional Characteristics of Network Awareness DataLayout Algorithm e core of the network-aware datalayout strategy is to combine the load of the network linkwith the evaluation of the importance of the node to obtainmore accurate node network load performance and thenoptimize the choice of target racks to balance network re-sources and reduce latency of data writing to reduce taskcompletion time

(1) e node load evaluation and selection algorithmcan fully consider the importance of the node in thenetwork topology and the real-time transmission ofadjacent links to calculate the load of the nodeSelecting a node with a smaller load as the targetstorage location can avoid assigning new transmis-sion tasks to congested cross-rack links therebyeliminating load imbalances of bottleneck link Inlarge-scale cluster storage systems some links areprone to congestion in the network e network-aware data layout strategy will select nodes with lesslink load to place data based on the node loadevaluation results avoiding selecting link with heavytransmission tasks to reduce task latency

(2) e data block write request processing algorithm canminimize the completion time of a single fileAccording to the value of the decision duration sdifferent sorting strategies for writing data blocks areflexibly adopted When sgt 0 tasks can be sorted basedon the number of remaining blocks in the parent file ofthe block Files with a small number of remainingblocks are processed preferentially which can shortenthe writing completion time of a single file

(3) e node selection algorithm for link load andstorage load balancing can minimize the load im-balance of storage nodes When selecting a storagenode the algorithm not only considers the load ofstorage space but also considers the network trafficload of the internal link of the rack e target datanode can be selected based on the network trafficload of the internal link of the rack and the load ofthe data node storage space can obtain a better loadbalance of the storage nodes inside the rack

(4) Maintaining the ordering of write task queues andnode loads will increase task completion time Firstlysorting the write queue in time s will increase the taskexecution time e larger the value of s the bettersorting result can be achieved but at the same time itwill increase the data block write timeerefore take asuitable value s as an important process Furthermoreupdating link information takes Δt time and calcu-lating the node value and selecting a node with a small

CEI value also takes a little time but the results have animportant effect on balancing the link load and re-ducing task waiting time Furthermore the selection ofstorage nodes inside the rack consumes some timeeinternal link load of the rack is lower than bottlenecklinks so the time it takes to calculate and sort the loadvalue of the storage node has negligible effect on thedata writing time

5 Experimental Evaluation

51 Experimental Setup In the simulation experiment thenumber of nodes was set as (1) 3000 including 150 racksand each rack had 20 server nodes and (2) 300 including 15racks and each rack had 20 server nodes e networktopological structure of the storage system in the experi-mental test is shown in Figure 6 in which only 15 racks witha total of 300 nodes are drawn Data transmission tasknumber increases from 500 1000 1500 2000 2500 and5000 respectively and the experiment tests the datatransmission completion time of the layout scheme in thispaper at two states of normal link transmission congestionand link congestion In this experiment the size of datablock is set to be the same

e network was the only bottleneck set in the experi-ment e cross-rack link is isomorphic with a maximumcapacity of 10240MB and so is the intra-rack link with amaximum transmission capacity of 256MB e transmis-sion rate for the cross-rack link is 1024Mbs and thetransfer rate for the inner link of rack is 64Mbs e initialload of the link is generated randomly as well as the usedspace size of each storage node e arrival rate of the datatransmission task is 10 per second the size of each data blockis the same as fixed at 64MB and the transmission task isexecuted in the order

In the test on the cluster storage system HDFS clusterwas built based on Hadoop 274 in the Linux environmentand three different cluster sizes were configured (1)1Master + 3DataNodes (2) 1Master + 7DataNodes and (3)1Master + 11DataNodes Firstly the task completion timeunder different file writing tasks was tested by changing thenumber of file writing tasks so as to analyze the performancewhen file writing load increaseden through changing thenumber of cluster nodes the completion time of writingtasks for the same number of files under the three clustersizes of 4 nodes 8 nodes and 12 nodes is tested which is toanalyze the impact on the performance of the layout strategyof cluster size

52 Performance Effects of Network Status and NetworkSize First of all the experiment tested the layout of 15 rackswith a total of 300 nodes and the data block transmissiontask with different numbers under noncongestion state ofthe link counted the transmission task completion timeunder the network awareness data layout scheme andmeasured the total transmission task completion time underthe layout scheme without considering network loadcharacteristics e specific results are shown in Figure 7

12 Mathematical Problems in Engineering

Secondly on the basis of the previous test the congestionof the link is increased to test the total task completion timeof the above two schemes in the case of different tasktransmission task quantities e specific results are shownin Figure 8

As shown in Figures 7 and 8 when the number of nodesis 300 the task completion time increases with the increasein congestion In the condition of congestion the datatransmission task needs to wait for the link to be free beforeperforming the transmission operation so the waiting timeis correspondingly increased resulting in the increase in thetotal task completion time e time for the networkawareness layout scheme to complete the same number oftransmission tasks is less than the execution time of thescheme without considering the network characteristics Onthe one hand the network awareness layout scheme avoidsrelatively more congested links and reduces the task waitingtime On the other hand although the network awarenessscheme costs sometime in the process of searching for high-quality nodes it is found in the experiment that the time

spent in searching for nodes in the topology with a smallnumber of nodes is negligible

en the topological network with a total of 3000 nodesof 150 racks was tested Under the condition of relativelycongested links the total time for data transmission task ofthe above two layout schemes was measured and the specificresults are shown in Figure 9

By comparing the results in Figures 8 and 9 the more thenumber of nodes increases the more the total completiontime of transmission tasks with the same number is On theone hand as the number of nodes increases the time re-quired for the node searching process adds which results inan increase in the total time On the other hand as thenumber of nodes increases for the scheme with no sensingthe possibility of repeatedly selecting the same node to storedata is reduced so the congestion is improved and thechange range of the total time to complete the task is smallerthan that of the scheme with fewer nodes

53 Performance Effects of the Number of Sort Policy esimulation experiment tested the effect of different sortingstrategies on the task completion time under 300 nodes thatcontain 15 racks Under the noncongested network changethe value of s respectively as 0 1 2 and 5 to create fourdifferent blocks writing queue e four different blockswriting queue include queue sorted by time of arrival queuesorted by the remaining father file during s= 1 second in-terval queue sorted by the remaining father file during s= 2seconds interval and queue sorted by the remaining fatherfile during s= 5 seconds intervalen the task finish time of500-block data transmission under the network-aware datalayout method is tested and recordede specific results areas shown in Figure 10

As shown in Figure 10 with the increase in file numbers(FNs) task completion time is in an upward trend Firstlyby comparing the task completion time under s 0 ands gt 0 it is shown that the sorting algorithm did not sig-nificantly increase the time of data writing task at sgt 0

Figure 6 e network topology of the storage system under ex-perimental test

Network-aware placementNon-network-aware placement

0

20

40

60

80

100

Task

com

plet

ion

time

250020001500 3000 3500500 10000Number of transmission tasks

Figure 7 Task completion time of different schemes undernoncongested network of 300 nodes

20

40

60

80

100

120

140

160

Task

com

plet

ion

time

500 10000 2000 2500 3000 35001500Number of transmission tasks

Network-aware placementNon-network-aware placement

Figure 8 Task completion time of different schemes under con-gested condition of 300 nodes

Mathematical Problems in Engineering 13

which indicates that the sorting decision had little impacton the completion time of data writing task en wecompared the completion time of written tasks under thedecision times of s 1 s 2 and s 5 it is shown that thetask completion time when s 5 was significantly higherthan that s 1 and s 2 As stated in Section 42 the largerthe value of s the better the sorting results will be obtainedby the algorithm but the data block writing time will beincreased at the same time erefore the value of s is acompromise process In this experiment s 1 and s 2 aretwo suitable values

54 Performance Effects of Cluster Size and Workloade test results on cluster storage system of the network-aware data layout algorithm are shown in Figures 11 and12 e performance of data layout algorithm under

different scales was tested in HDFS By increasing thenumber of cluster nodes to expand the cluster sizerecord the task completion time of the algorithm whenFN 100

As shown in Figure 11 with the increase in thenumber of cluster nodes the effect of the networkawareness algorithm becomes better and better As thenumber of files increases the number of tasks to betransmitted increases and the network load of the clusterstorage system increases so that some link congestion islikely to occur Network-aware data layout algorithm canavoid links with heavy transmission tasks and selectnodes with less load to place data thus reducing the taskwaiting time However as the number of tasks continuesto increase almost every link is saturated and the newwriting task is added to waiting queue and the perfor-mance of the network-aware data layout algorithmdropped because no matter which link is selected blocksneed to wait in this case

0

50

100

150

200

250

300

350

Task

com

plet

ion

time

500 1000 1500 2000 2500 3000 35000Number of transmission tasks

Network-aware placementNon-network-aware placement

Figure 9 Task completion time of different schemes under con-gested condition of 3000 nodes

0

500

1000

1500

2000

2500

Task

com

plet

ion

time

200 300 400 500100FN

1 + 31 + 71 + 11

Figure 11 Task completion time under different cluster sizes

5

0s = 0 s = 1 s = 2 s = 5

S

10

15

20

25

30

35

40

45

50

Task

com

plet

ion

time

FN = 100

FN = 1000FN = 500

Figure 10 Task completion time of under different decision timess

0

10

20

30

40

50

60

70

80St

orag

e use

d

73 6 8 942 5 10 111Data node

Figure 12 Storage load of each data nodes while FN 100

14 Mathematical Problems in Engineering

In the test the storage space load of 11 data nodes in the1master + 11data nodes cluster was obtained under FN 100write task was completed as shown in Figure 12

As shown in Figure 12 the storage load of 11 nodesfluctuates between 65 and 80 which indicates that thealgorithm has a good storage balancing effect ere arestill some nodes with large load differences such as nodes2 and 6 because the load of storage space is not onlyconsidered in the selection of nodes but also the networktraffic load of the internal link of the racke storage loadof the cluster indicates that Algorithm 3 has a good loadbalancing effect in selecting the target data node based onthe internal link network traffic load and the data storagespace load

6 Conclusion

Storage system network features will change significantlyafter carrying the data flow Based on the complex net-work theory this paper analyzes centrality index ofstorage nodes under the influence of storage systemnetwork data flow to identify node important degreebearing capacity and the equilibrium condition in theprocess of the storage system data transmission andfurther data layout performance optimization Firstlyconsidering the local characteristics of data transmissionthe path selection of data transmission the distance ofdata transmission and the carrying capacity of the nodesfour indexes of node strength ability betweenness datatransmission amount and concentration index of dataflow are proposed for the identification of the node flowload in the data flow network en according to thearrival time of the task and the data amount of theremaining blocks in the parent file of the data block aflexible sorting method of the data block writing task isproposed Finally according to the result of node flowload identification the target rack and storage nodeare selected according to the principle of leastload and a network awareness data layout scheme isproposed

Experimental results show that the proposed data layoutscheme of network awareness in this paper is better than thatwithout considering the network characteristics of the layoutplan in the aspect of transmission task completion time toimprove the efficiency of data transmission task executionreduce task execution time effectively enhance the efficiencyof data storage and achieve the effect of network flowequilibrium In the future research work the networkawareness data layout scheme based on future flow pre-diction will be further studied

Data Availability

e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

e authors declare that there are no conflicts of interest

Authorsrsquo Contributions

D S designed the algorithms and wrote the paper B S GB and Y Z made a careful revision of the article andproposed amendments

Acknowledgments

is work was supported by the National Natural ScienceFoundation of China under Grant no 61672416 the Na-tional Natural Science Foundation of China under Grant no61872284 and the Shaanxi Natural Science Foundation ofChina (2018JM6105)

References

[1] M Shojafar N Cordeschi D Amendola et al ldquoEnergy-savingadaptive computing and traffic engineering for real-time-service data centersrdquo in Proceedings of the 2015 IEEE Inter-national Conference on Communication Workshop (ICCW)pp 1800ndash1806 IEEE London UK June 2015

[2] D Jianguang Z Yuelong and Y Huaqiang ldquoDynamic datareplication management strategy in cloud computing envi-ronmentrdquo Journal of Huazhong University of Science andTechnology (Natural Science Edition) vol 43 no 10 pp 53ndash57 2015

[3] Y Lina ldquoImprovement of HDFS balanced placement strat-egyrdquo Computer Science vol 44 no S2 pp 397ndash399+4312017

[4] H Dongmei D Yanling H Qi et al ldquoMarine monitoringdata replica layout strategy based on multiple attribute op-timizationrdquo Computer Science vol 45 no 6 pp 72ndash75 2018

[5] T Yongcai B Yang S Lei et al ldquoManagement mechanism ofdynamic cloud data replica based on availabilityrdquo Journal ofChinese Computer Systems vol 39 no 3 pp 490ndash495 2018

[6] W A Xiuguo ldquoReplica strategy considering cost and storagespace in cloud environmentrdquo Computer Engineering vol 44no 3 pp 19ndash26 2018

[7] L Jun and H Mengshu ldquoReplica placement strategy based onglowworm swarm optimizationrdquo Application Research ofComputers vol 36 no 2 pp 584ndash587 2019

[8] Z Bang W Xingwei and H Min ldquoIntelligent multiple datareplica placement scheme for cloud storagerdquo Journal ofFrontiers of Computer Science and Technology vol 8 no 10pp 1177ndash1186 2014

[9] M Barshan H Moens S Latre B Volckaert and F D TurckldquoAlgorithms for network-aware application componentplacement for cloud resource allocationrdquo Journal of Com-munications and Networks vol 19 no 5 pp 493ndash508 2017

[10] J Xiao B Wu X Jiang A Pattavina H Wen and L ZhangldquoScalable data center network architecture with distributedplacement of optical switches and racksrdquo Journal of OpticalCommunications and Networking vol 6 no 3 pp 270ndash2812014

[11] W Xiuguo ldquoResearch on minimum cost data replica distri-bution based on dynamic planning in cloud storage systemrdquoComputer Engineering vol 43 no 7 pp 29ndash37 2017

[12] M Alicherry and T V Lakshman ldquoNetwork aware resourceallocation in distributed cloudsrdquo in Proceedings of the IEEEINFOCOM (2012) pp 963ndash971 IEEE Orlando FL USAMarch 2012

Mathematical Problems in Engineering 15

[13] W Xiuguo ldquoMinimum-cost based data replication strategy incloud computing environmentrdquo Computer Science vol 41no 10 pp 154ndash159+190 2014

[14] L Xuejun W Yang L Xiao et al ldquoDatacenter-oriented dataplacement strategy of workflows in hybrid cloudrdquo Journal ofSoftware vol 27 no 7 pp 1861ndash1875 2016

[15] W Yan and W Jinkuan ldquoA dynamic replication placementmechanism in cloud storagerdquo Computer Engineering andScience vol 39 no 9 pp 1581ndash1587 2017

[16] M H Ferdaus M Murshed R N Calheiros and R BuyyaldquoAn algorithm for network and data-aware placement ofmulti-tier applications in cloud data centersrdquo Journal ofNetwork and Computer Applications vol 98 pp 65ndash83 2017

[17] L Qingying X Lin and L Xicong ldquoEnergy efficient clouddata replica layout algorithm considering network band-widthrdquo Science Technology and Engineering vol 19 no 5pp 172ndash178 2019

[18] A Uta O Danner C van der Weegen et al ldquoMemEFS anetwork-aware elastic in-memory runtime distributed filesystemrdquo Future Generation Computer Systems vol 82pp 631ndash646 2018

[19] M Sipos J Gahm N Venkat and D Oran ldquoNetwork-awarefeasible repairs for erasure-coded storagerdquo IEEEACMTransactions on Networking vol 26 no 3 pp 1404ndash14172018

[20] A Epstein E K Kolodner and D Sotnikov ldquoNetwork awarereliability analysis for distributed storage systemsrdquo in Pro-ceedings of the 2016 IEEE 35th Symposium on Reliable Dis-tributed Systems (SRDS) pp 249ndash258 IEEE BudapestHungary September 2016

[21] M Al-Fares S Radhakrishnan B Raghavan et al ldquoHederadynamic flow scheduling for data center networksrdquo NSDIvol 10 no 8 pp 89ndash92 2010

[22] S John and M Mohamed ldquoA network performance awareQoS based workflow scheduling for grid servicesrdquo Ce In-ternational Arab Journal of Information Technology vol 5no 15 pp 894ndash903 2018

[23] Z Jingya F Jianxi and W Jin ldquoData placement approach forscalable online social networks (in Chinese)rdquo SCIENTIASINICA Informationis vol 48 no 3 pp 329ndash348 2018

[24] X Meng Y Wang and Y Gong ldquoPerspective of space andtime based replica population organizing strategy in un-structured peer-to-peer networksrdquo Journal of Network andComputer Applications vol 49 pp 1ndash14 2015

[25] G Gao R Li H He and Z Xu ldquoDistributed caching inunstructured peer-to-peer file sharing networksrdquo Computersamp Electrical Engineering vol 40 no 2 pp 688ndash703 2014

[26] S K Bhatti M I U Lali B Shahzad F Javid F U Manglaand M Ramzan ldquoLeveraging the big data produced by thenetwork to take intelligent decisions on flow managementrdquoIEEE Access vol 6 pp 12197ndash12205 2018

[27] L Qi W Lu Y Xiao et al ldquoPath selection algorithm based onopen daylight network awareness and user requirementsrdquoJournal of Chinese Computer Systems vol 39 no 8pp 1737ndash1743 2018

[28] R Wang S Mangiante A Davy et al ldquoQoS-aware multi-pathing in datacenters using effective bandwidth estimationand SDNrdquo in Proceedings of the 2016 12th InternationalConference on Network and Service Management (CNSM)pp 342ndash347 IEEE Montreal Canada November 2016

[29] L Yujie L Dianjie and Z Guijuan ldquoCloud content deliverynetwork based on energy optimizationrdquo Journal of ChineseComputer Systems vol 39 no 10 pp 2216ndash2221 2018

[30] M Shojafar Z Pooranian and P G V Baccarelli ldquoFLAPSbandwidth and delay-efficient distributed data searching infog-supported P2P content delivery networksrdquoCe Journal ofSupercomputing vol 73 no 12 pp 5239ndash5260 2017

[31] O Biran A Corradi M Fanelli et al ldquoA stable network-aware vm placement for cloud systemsrdquo in Proceedings of the12th IEEEACM International Symposium on Cluster Cloudand Grid Computing (ccgrid 2012) pp 498ndash506 IEEE Ot-tawa Canada May 2012

[32] R Wang J A Wickboldt R P Esteves L Shi B Jenningsand L Z Granville ldquoUsing empirical estimates of effectivebandwidth in network-aware placement of virtual machinesin datacentersrdquo IEEE Transactions on Network and ServiceManagement vol 13 no 2 pp 267ndash280 2016

[33] F Dongyu Z Ligu X Zida et al ldquoApproach for optimizingdata placement on mongo DB clusterrdquo Computer Engineeringand Applications vol 53 no 17 pp 77ndash84 2017

[34] L Shengbin T Xiaoming L Zhiqing et al ldquoDiscrete layoutstrategy for multiple replica of spatial data based on parallelcomputingrdquo Journal of Jilin University (Science Edition)vol 54 no 3 pp 524ndash528 2016

[35] R Wang R Esteves L Shi et al ldquoNetwork-aware placementof virtual machine ensembles using effective bandwidth es-timationrdquo in Proceedings of the 10th International Conferenceon Network and Service Management (CNSM) and Workshoppp 100ndash108 IEEE Rio de Janeiro Brazil November 2014

[36] W Xiaojie X Mingwei and W Sixiu ldquoTwo-phase virtualmachine placement algorithm based on network awarenessrdquoComputer Engineering vol 43 no 8 pp 32ndash37 2017

[37] C Lei Z Jing and C Lijun ldquoA network-aware two-phasevirtual machine allocation algorithmrdquo Journal of HunanUniversity (Natural Sciences) vol 43 no 4 pp 120ndash132 2016

[38] F Ahmad S T Chakradhar A Raghunathan et al ldquoShuf-fleWatcher shuffle-aware scheduling in multitenant Map-Reduce clustersrdquo in Proceedings of the 2014 USENIX AnnualTechnical Conference (USENIX ATC 14) pp 1ndash13 Phila-delphia PA USA June 2014

[39] J Li S Yang X Wang et al ldquoTree-structured data regen-eration in distributed storage systems with regeneratingcodesrdquo in Proceedings IEEE INFOCOM 2010 pp 1ndash9 IEEESan Diego CA USA March 2010

[40] M Chowdhury S Kandula and I Stoica ldquoLeveraging end-point flexibility in data-intensive clustersrdquo ACM SIGCOMMComputer Communication Review vol 43 no 4 pp 231ndash2422013

16 Mathematical Problems in Engineering