Embed Size (px)
Citation preview
Available online at www.sciencedirect.com
H O S T E D B YDigital Communications and Networks (2015) 1, 181–190
http://dx.doi.org/12352-8648/& 2015 Carticle under the CC
E-mail address: qPeer review under
and Telecommunicat
journal homepage: www.elsevier.com/locate/dcan
Modeling and performance analysisfor composite network–compute serviceprovisioning in software-definedcloud environments
Qiang Duan
Information Sciences & Technology Department, The Pennsylvania State University Abington, PA 19001,United States
Received 18 December 2014; received in revised form 19 February 2015; accepted 16 May 2015Available online 1 June 2015
KEYWORDSCloud computing;Composite serviceprovisioning;Software-definedCloud environment;Service modeling;Performance analysis
0.1016/j.dcan.201hongqing UniversiBY-NC-ND license
[email protected] ofions.
AbstractThe crucial role of networking in Cloud computing calls for a holistic vision of both networkingand computing systems that leads to composite network–compute service provisioning.Software-Defined Network (SDN) is a fundamental advancement in networking that enablesnetwork programmability. SDN and software-defined compute/storage systems form aSoftware-Defined Cloud Environment (SDCE) that may greatly facilitate composite network–compute service provisioning to Cloud users. Therefore, networking and computing systemsneed to be modeled and analyzed as composite service provisioning systems in order to obtainthorough understanding about service performance in SDCEs. In this paper, a novel approach formodeling composite network–compute service capabilities and a technique for evaluatingcomposite network–compute service performance are developed. The analytic method pro-posed in this paper is general and agnostic to service implementation technologies; thus isapplicable to a wide variety of network–compute services in SDCEs. The results obtained in thispaper provide useful guidelines for federated control and management of networking andcomputing resources to achieve Cloud service performance guarantees.& 2015 Chongqing University of Posts and Telecommunications. Production and Hosting by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license(http://creativecommons.org/licenses/by-nc-nd/4.0/).
5.05.003ty of Posts and Telecommunicatio(http://creativecommons.org/lic
Chongqing University of Posts
1. Introduction
Cloud computing is a large scale distributed computingparadigm driven by economies of scale, in which a pool ofabstracted, virtualized, dynamically scalable computing
ns. Production and Hosting by Elsevier B.V. This is an open accessenses/by-nc-nd/4.0/).
Q. Duan182
functions are delivered on demand as services to externalcustomers over the Internet [1]. Networking plays a crucialrole in Cloud computing. From a service provisioningperspective, the services received by Cloud end userscomprise not only computing functions provided by Clouddata centers but also communications functions offered bynetworks. Results obtained from recent study on Cloudservice performance have indicated that networking has asignificant impact on quality of Cloud services, and in manycases data communications become a bottleneck that limitsClouds from supporting high-performance applications[2,3]. Therefore, networks with Quality of Service (QoS)capabilities become an indispensable ingredient for high-performance Cloud service provisioning.
For example, consider a scenario in which an applicationutilizes the Cloud infrastructure for storing and processing alarge data set and requires upper bounded response delay.This application may use the computing capability ofAmazon EC2 (Elastic Compute Cloud) and the storagecapacity of Amazon S3 (Simple Storage Service). In orderfor the user to access EC2 and S3, network services mustalso be provided for data transmissions from the user to S3virtual disk, between S3 disk and the EC2 server (if EC2 andS3 are located at different sites), and from EC2 server backto the user. Therefore, the end-to-end service offered tothe user is essentially a composition of both Cloud servicesand network services. In order to meet the service delayrequirement of the application, sufficient amount of net-working resources (e.g. transmission bandwidth and packetforwarding capacity) must be provided to guarantee net-work delay performance in addition to the computing andstorage resources offered by the Cloud infrastructure formeeting data processing and storing requirements.
The significant role that networking plays in Cloud serviceprovisioning calls for a holistic vision of the computing andnetworking systems involved a Cloud environment. Such avision requires federated management, control, and optimiza-tion of computing and networking resources for compositeservice provisioning. Software-Defined Networking (SDN) is oneof the latest revolutions in the networking field, whichdecouples network control and data forwarding functions;thus enabling network control to be programmable and under-lying network infrastructure to be abstracted for applications[4]. The logically centralized control plane in SDN allows upperlayer applications to program the underlying network platformthrough a standard API; therefore can provide better supportfor Cloud computing. Expectation of more agile Cloud servicesalso requires a high degree of programmability for computinginfrastructure, which leads to software-defined computing andstorage. Software-defined network, compute, and storagesystems, when integrated together, enable an environmentwith fully automated provisioning and orchestration of ITinfrastructures for Cloud computing, which is referred to asSoftware-Defined Cloud Environment (SDCE) [5].
A key to realize the SDCE notion is orchestration of hetero-geneous network, compute, and storage systems for compositeservice provisioning. The Service-Oriented Architecture (SOA)provides an effective mechanism for heterogeneous systemintegration through loose-coupling interactions among systemcomponents. SOA has been widely adopted in Cloud computingvia the IaaS, PaaS, and SaaS paradigms. Applying SOA in the
field of networking leads to the Network-as-a-Service (NaaS)paradigm that enables encapsulation and virtualization ofnetworking resources in the form of SOA-compliant networkservices. NaaS allows network infrastructure to be virtualized,exposed, and accessed as services that can be orchestratedwith computing services in a Cloud environment to provisioncomposite network–compute services to Cloud users [6].Recently NaaS has been proposed as a key mechanism in SDNfor achieving end-to-end QoS provisioning [7]. Therefore SOAmay form a basis for service provisioning in SDCEs.
As SDCE being rapidly adopted by service providers, itbecomes important to obtain thorough understanding aboutperformance of composite network–compute service provi-sioning, which is the service performance actually perceivedby Cloud users. Since networking has a strong impact onend-to-end performance for Cloud service provisioning,Cloud service performance should be evaluated with aholistic vision of both computing and networking aspects.For example, suppose a biology lab creates 100 GB of rawdata that will be processed in Amazon EC2 Cloud. Assumethat the lab obtained 10 EC2 virtual machine instances andeach instance can process 20 GB data per hour, then thetotal processing time of the Cloud service is only 30 min.However, if the lab uses a network service that offers200 Mb/s throughput for data transmission to the EC2server, then even the single-trip delay for data transmissionfrom the lab to EC2 servers will be 67 min. In this examplenetwork delay for round-trip data transmission contributesmore than 80% of the total service delay; thus demonstrat-ing the significant impact of networking on Cloud serviceperformance.
Analytical modeling and analysis provide an effectiveapproach to obtaining insights about end-to-end serviceperformance. Cloud performance analysis has attractedattention of the research community and many results havebeen reported in the literatures. However, the significantimpact of network performance on Cloud service provision-ing has not been sufficiently considered in these works. Onthe other hand, although network performance has beenextensively studied, currently available techniques typicallylack the ability to analyze composite systems that consist ofheterogeneous networking and computing functionalities.Therefore, system modeling and performance analysis forcomposite network–compute service provisioning in SDCEsare still an open problem.
Network–compute service orchestration in SDCEs bringsin new challenges to service modeling and performanceanalysis. Key features of Cloud computing and NaaS, suchas virtualization and abstraction of networking and com-puting resources, make service provisioning independentof system implementations; thus requiring modeling andanalysis methods to be general and agnostic to networkand Cloud implementations. In order to tackle this chal-lenge, a novel modeling approach is proposed in this paperfor characterizing the service capabilities of compositenetwork–compute systems. Analysis techniques are devel-oped based on this model for evaluating delay perfor-mance of composite network–compute services. Thedeveloped modeling and analysis techniques are generaland agnostic to network and Cloud implementations; thusare applicable to composite network–compute service
data transmissioninfrastructure Infrastructure
Layer
software-definednetwork
end userapplication
end userapplication
NaaS interface
data storageinfrastructure
service orchestration
ControlLayer
Application Layer
data processinfrastructure
software-definedcompute
software-definedstorage
IaaS interface
service abstractionService Layer
networking domain computing domain
Fig. 1 A framework for software-defined cloud environmentarchitecture.
network service
communication networkend user
forward datatransmit
compute service
Cloud computinginfrastructure
backward datatransmit
composite service provisioning in SDCE
data process& storage
Fig. 2 Composite network–compute service provisioningin SDCE.
183Modeling and performance analysis for composite network–compute service provisioning in software-defined
delivery systems in SDCEs.The rest of this paper is organized as follows. An overview
of network–compute service composition in SDCEs is given inSection 2. Related works on Cloud service performanceevaluation and the new challenges brought in by network–compute service composition in SDCEs are discussed inSection 3. A new service capability model is proposed inSection 4 and applied in Section 5 to characterize capabil-ities of composite network–compute service systems inSDCEs. A technique for evaluating delay performance forcomposite network–compute service is developed in Section6. Section 7 provides numerical results for illustratingapplications of the developed techniques. Section 8 drawsconclusions.
2. Composite network–compute serviceprovisioning in software-defined cloudenvironments
Virtualization is a key enabling technology for Cloud computingand the SOA forms a foundation for Cloud service provisioning.Recent research and development have been bridging thepower of SOA and virtualization in the context of Cloudcomputing ecosystem [8]. The Open Grid Forum (OGF) isworking on the Open Cloud Computing Interface (OCCI) stan-dard, which defines SOA-compliant open interfaces for inter-acting with Cloud infrastructure. The SOA principle has stronglyinfluenced the service models of most Cloud service providers.For example Amazon expose its compute and storage Cloudservices via Web service interfaces.
Virtualization will play a crucial role in the next genera-tion networks [9]. Network virtualization separates networkservice provisioning from data transport infrastructure.Applying SOA in network virtualization enables resourcesin network infrastructure to be abstracted as infrastructureservices. Network service providers can construct differentvirtual networks by utilizing network infrastructure servicesto offer end-to-end network services for meeting diverseuser requirements. Upper layer applications can utilize theunderlying networking platform by accessing the networkservices through a standard abstract interface, which isessentially a Network-as-a-Service (NaaS) paradigm.
Software-Defined Network (SDN) represents a fundamentaladvancement in the field of networking. SDN technology alsobrings new possibility for Cloud service providers. By takingadvantage of the logically centralized control of networkresources, it is possible to simplify and optimize managementof Cloud networks to achieve more flexible and efficient intra-data center networking and improved inter-datacenter com-munications. NaaS can be applied in SDN to further enhanceflexibility and performance of network service provisioning. SDNand software-defined compute and storage systems form thethree key components of a Software-Defined Cloud Environ-ment (SDCE).
A framework for service-oriented SDCE architecture is shownin Fig. 1. The infrastructure layer in this framework comprisesphysical infrastructure for data transmission, processing, andstorage. The control functions are decoupled from physicalinfrastructure and logically centralized at the software-definednetwork, compute, and storage controllers, which form thecontrol layer of the framework. The NaaS paradigm exposes
networking functionalities as SOA-compliant services in thesame way as IaaS interface exposes compute and storageresources as services. The service layer supports high-levelabstractions of both networking and computing systems andprovides a unified mechanism to orchestrate network andcompute services to form composite services, which areprovisioned to user applications on the application layer. Inthis framework, service-oriented virtualization of both net-working and computing systems enables federated control,management, and optimization of the resources in bothnetworking and computing domains.
From a service provisioning perspective, the servicesoffered to end users are composite services that compriseboth computing functions (for data process and/or storage)and communication functions (for data transmission). Atypical end-to-end provisioning system for composite net-work–compute services is shown in Fig. 2, which consists ofCloud infrastructure that offers compute service and thecommunication network that provides network services.
SDCE is an integrated service delivery environment in whichnetwork and compute services, used to be offered separatelyby different providers, converge into composite network–com-pute services. Therefore SDCE enables a new service model thatallows the roles of traditional Internet service providers andCloud service providers merge together for composite serviceprovisioning. Such a new service environment may stimulateinnovations in the development, deployment, and utilization ofCloud services; thus creating a wide variety of new businessopportunities.
Q. Duan184
3. Challenges to performance evaluation ofcomposite network–compute services
Service performance is one of the decisive factors in adoptionof the new Cloud computing paradigm for various applications.Therefore, performance evaluation of Cloud services hasattracted extensive research interest.
In [10] the authors developed an approach to evaluating theperformance of various Cloud service offerings with differentconfigurations. An experimental evaluation on Amazon ElasticCompute Cloud (EC2) was reported in the paper to verify thisapproach. Hill and Humphrey [11] have quantitatively evaluatedEC2's performance for the common MPI-based scientific comput-ing applications. The authors of [2] presented a comprehensiveperformance evaluation of Amazon EC2 for supporting high-performance computing using representative workloads of atypical supercomputing center. A portable and extensible frame-work for generating and submitting test workloads to computingClouds was designed in [12]. The authors employed this frame-work to study Cloud response time in various metrics includingoverheads for acquiring and realizing virtual computingresources. In [13] the authors quantified the presence ofmany-task-computing users in real scientific computing work-loads and performed an empirical evaluation of the performanceof four commercial Cloud services. The authors of [14] examinedthe performance and cost of using Cloud computing forsupporting scientific workflow applications, which consist of aset of loosely-coupled computing tasks connected through data-and control-flow dependencies. Li and his coauthors constructeda taxonomy of performance evaluation of commercial Cloudservices in [15], which focuses on measurement-based Cloudservice evaluation approaches.
The aforementioned research results are mainly based onmeasurement and testing experiments conducted on someparticular types of computing Clouds that have been deployedand offered to the public, such as Amazon EC2 services. Cloudcomputing is a rapidly developing field in which new services aswell as new implementations of existing services keep emer-ging. Measurement-based methods are not sufficient for deter-mining the achievable performance of new Cloud services orservices with new configuration settings. Therefore, analyticalmodeling and performance analysis techniques are necessary inorder to obtain thorough understanding about service perfor-mance for general Cloud computing scenarios. This allowsservice customers and service brokers to predict the achievableservice performance to discover and select the appropriateCloud services for various applications. Analytical methods mayalso provide guidelines for Cloud service providers to manageand configure resources for service provisioning.
Queueing theory has been applied to develop analyticalmethods for evaluating Cloud service performance. Xiongand Perros [16] modeled a Cloud computing system as anopen queue network consisting of two tandem servers withfinite buffer space. In order to study resource allocation formeeting performance requirements of clients with differentpriority levels, the authors of [17] modeled a Cloud datacenter as an M/M/C/C queueing system, which has C serverswith no buffer space. Yang et al. developed anM=M=m=mþr queueing model for Cloud data centers in[18]. In this model both arrival time and service time areassumed to be exponentially distributed and service
response time is broken into three independent parts:waiting, service, and execution periods. Due to the com-plexity of Cloud computing technologies and diversity ofuser requests, developing exact models for Cloud servicesystems is very difficult. In order to address this challenge,Khazaei and his coauthors of [19] proposed an approximateM=G=m=mþr queuing model for Cloud server farms andsolved it to obtain estimation of complete probabilitydistribution of service response time and other performanceindicators. This approximate model was extended in [20] toreflect burst arrival process of Cloud data centers, whichleads to a M½x�=G=m=mþr queuing system where arrivalprocess is a sequence of super-tasks each of which consistsof a burst of tasks.
Although many research results indicate that networkinghas a strong influence on Cloud service performance, theaforementioned research works focus on evaluating perfor-mance of Cloud data centers without sufficiently consider-ing the impact of network performance on Cloud serviceprovisioning. On the other hand, although network perfor-mance and QoS have been extensively studied, littleresearch that views networking as an integrated elementof Cloud service provisioning has been reported. Currentlyavailable methods for performance analysis basically con-sider computing and networking systems separately; thuslacking a holistic vision for evaluating performance ofcomposite network–compute service provisioning in asoftware-defined Cloud environment.
Traditional queueing analysis methods are all based onsome assumptions about certain implementation mechan-isms of the studied systems such as a Cloud data center or anetworking system. However Cloud services and theirimplementation technologies are changing rapidly. In addi-tion, resource virtualization and service orientation in Cloudcomputing and NaaS-based SDN decouple Cloud and networkservices from their implementations and make the lattertransparent to applications. End users and service brokersselect, orchestrate, and access services without knowledgeof their detailed implementations. Therefore, traditionalqueuing theory-based analysis methods are not generalenough for facing the challenges of evaluating compositenetwork-Cloud service performance.
Composite service delivery systems in a SDCE consist ofheterogeneous networking and computing systems withdiverse implementations. Therefore, the modeling andanalysis techniques for evaluating composite service per-formance must be general and applicable to the widevariety of systems coexisting in a SDCE. SOA for Cloudservice provisioning and NaaS-based SDN enable both com-puting and networking resources to be abstracted asservices and decouple service functions from their imple-mentations. This requires the modeling and analysis tech-niques to be agnostic to the implementation technologies ofnetworking and computing systems for composite serviceprovisioning. Network calculus theory [21] has been appliedin performance evaluation of Grid network services [22] andnetwork virtualization [23,24]. In previous work [25,26], theauthor explored application of network calculus to tackleperformance analysis for Cloud network services. In thispaper, such a novel idea is fully developed into a systematicapproach to modeling and analyzing composite network–
185Modeling and performance analysis for composite network–compute service provisioning in software-defined
compute service performance in order to meet the aboverequirements.
4. Modeling composite network–computeservice provisioning systems
A typical delivery system for composite network–computeservices in SDCEs is shown in Fig. 3, which consists ofnetwork services for forward and backward data transmis-sions, compute services for data process in a Cloud datacenter, and data transform functions between the datatransmissions and data process.
In order to receive performance guarantee from acomposite service, the end user must expect a certain levelQoS from both network and compute services. In general,such QoS expectation can be defined in the Service LevelAgreements (SLAs) between the user and service providers.Although SLAs may vary due to the diversity of services, ittypically includes a requirement on the minimum datatransport rate for network services and the minimum dataprocessing capacity for compute services.
In order to analyze the composite service performance,one must examine the communication capability offered bythe network service and data processing capability providedby the compute service. The methodology taken in thispaper is to first develop a general capability profile that canmodel service capabilities of both network and computeservices, and then compose the capability profiles of thetwo service components into one end-to-end profile thatmodels the service capability of the composite system. Sucha capability profile should give a lower bound of the amountof service that a user can expect from the services (includ-ing both network and compute services), should be inde-pendent of implementation technologies of the underlyingnetworking and computing infrastructures, and should alsobe easy to combine for modeling composite service cap-abilities. In order to meet these requirements, the conceptof service curve from network calculus theory [21] isemployed in this paper for developing such a general andflexible capability profile.
A capability profile for a service component can bedefined as follows. Let R(t) and E(t) respectively be theaccumulated amount of traffic that arrives at and departsfrom a service component by time t. Given a non-negative,non-decreasing function, Pð�Þ, where Pð0Þ ¼ 0, we say thatthe service component has a capability profile P(t), if forany tZ0 in the busy period of the service component
EðtÞZRðtÞ � PðtÞ ð1Þwhere � denotes the operation defined as x
�tÞ � y
�tÞ ¼
infs:0r sr t xðt�sÞþyðsÞ� �.
The capability profile of a service component, which isessentially a service curve of the component, is defined as a
networkservice 1
enduser
forwarddata
transmit
compute service networkservice 2
dataprocessin Cloud
datatransform
datatransform
backwarddata
transmit
Fig. 3 A typical delivery system for composite network–compute service in SDCE.
general function of time that specifies service capabilitythrough the relation between arrival and departure trafficat the service component. Therefore such a profile isindependent of the implementations of service components,thus is applicable to both network and compute servicecomponents with various implementations.
The capability profiles defined in (1) gives a generalapproach to modeling network and Cloud service capabil-ities. In order to obtain a more tractable profile that cancharacterize capabilities of typical network and computeservices, this paper defines a Latency-Rate (LR) profile for aservice component as follows. If a service component has acapability profile
Pðr; θÞ ¼ r½t�θ�þ ð2Þwhere ½��þ is equal to its argument if it is positive and zerootherwise; then the service component has a LR profile. Theθ and r are respectively called the latency and rateparameters of the profile.
A LR profile can serve as the capability model for typicalnetwork services. The QoS expectation of a typical networkservice includes the minimum data transmission bandwidthguaranteed to a service user, which is described by the rateparameter r in a LR profile. Data communication in anetwork also experiences a fixed delay that is independentof traffic queuing behavior; for example signal propagationdelay, link transmission delay, router/switch processingdelay, etc. The latency parameter θ of a LR profile is tocharacterize this part of fixed delay of a network service.
A LR profile can also characterize service capabilities oftypical computing systems. Cloud service providers typicallyoffer some certain service capacity units to users. Forexample each type of virtual machine (called instance) inAmazon EC2 provides a predictable amount of computingcapacity and I/O bandwidth. Each EC2 compute unit pro-vides the equivalent CPU capacity of 1.0–1.2 GHz 2007Opteron or 2007 Xeon processor. Amazon also claims aninternal I/O bandwidth of 250 Mb/s regardless of instancetype. The latency and rate parameters of the LR profile for aCloud service can be derived from the processing capacityand I/O bandwidth information specified by its provider.
In order to represent the data transformation effect ofcomputing function provided by Cloud infrastructure, theconcepts of scaling function and scaling curve, which wereoriginally developed in [27] as an extension to networkcalculus, are adopted in the model for composite network–compute service provisioning systems.
A scaling function is defined as a function Sð�Þ that assignsan amount of scaled data S(a) to an amount of data a. Ascan be seen from the definition, scaling function is a generalconcept for taking into account data transformation in asystem model. Note that it does not model any queuingeffect – a scaling function is assumed to have zero delay.Queuing related effect of Cloud computing is modeled bythe service curve-based capability profile of the computeservice component.
Given a scaling function S, the function S is called a(minimum) scaling curve of S iff 8bZ0 it applies thatSðbÞr infaZ0fSðbþaÞ�SðaÞg.
Applying the above defined capability profile and scalingcurve, a composite network–compute service provisioningsystem can be modeled with capability profiles of network
Q. Duan186
and compute service components and the scaling curvesthat represent data transform between networking andcomputing, as shown in Fig. 4. In this system the networkservices for forward data transmission (from user to datacenter) and backward data transmission (from data centerto user) are respectively modeled by the profiles Pn1ðtÞ andPn2ðtÞ. The compute service offered by the data center ismodeled by the profile PC(t). The scaling curve Sn2c modelsdata transform from forward transmission to computingserver while the scaling curve S c2n models data transformfrom computing server to backward data transmission.
5. End-to-end capability profile for compositenetwork–compute services
This section presents a technique for integrating the cap-ability profiles of all service components and scaling curvesin a composite network–compute service system into anend-to-end profile. Such an end-to-end profile models theservice capability guaranteed by the service delivery systemto its end user, which forms the basis for analyzing delayperformance of composite network–compute services.
It is known from network calculus theory that the servicecurve of a system consisting of a series of tandem serverscan be obtained from convolution of the service curves ofall these servers. The capability profile defined in (1) isessentially the service curve of a service component.However, due to the scaling curves for data transformbetween networking and computing, the convolution opera-tion cannot be directly applied to this model. In order tosolve this problem, the alternative scaled servers (Theorem3.1 in [27]) are used to modify the system model so that anend-to-end profile can be obtained through network calcu-lus convolution.
Considering the two systems shown in Fig. 5, system(a) consists of a server with service curve β whose output isscaled with a scaling function S; and system (b) consists of ascaling function S whose output is input to a server with aservice curve βS. It is proved in [27] that given system (a),the lower bound of system (b) output function SðRÞ � βS is avalid lower bound for the output function of system (a), ifβS ¼ SðβÞ. Given system (b), the lower bound of system(a) output function SðR � βÞ is a valid lower bound for theoutput function of system (b), if β¼ S �1ðβSÞ where S �1
stands for the inverse function of S.
Fig. 4 A profile-based capability model for composite net-work–compute service provisioning.
Fig. 5 Alternative service systems with a scaling function.
This means in effect that performance bounds for systems(b)/(a) under the respective assumption are also validbounds for system (a)/(b). Therefore, a scaling functionand a service component in a model can be switchedwithout changing performance bounds, as long as thecapability profile of the service component is transformedusing the scaling curve. In addition it is also shown in [27]that performance bounds obtained in the alternative sys-tems after the above switch operation remain tight.
Applying the above alternative server method in themodel shown in Fig. 4, network service 1 for forward datatransmission and the scaling function for network/computedata transform can be switched without impacting theperformance bound guaranteed by the system, if thecapability profile of network service 1 is transformed toPSn1 ¼ Sn2cðPn1Þ. Similarly, the scaling function for compute/
network data transform and network service 2 for backwarddata transmission can be switched, if the capability profileof network service 2 is transformed to PS
n2 ¼ S �1c2n ðPn2Þ. Then
the composite service system has the alternative model asshown in Fig. 6.
Since the capability profile defined in (1) is essentially theservice curve of a service component, the capability profilesfor both network service components (for forward andbackward data transmissions) and the compute servicecomponent can be integrated into one profile using theconvolution operation defined in network calculus. There-fore, the end-to-end capability profile for the compositesystem, denoted by Pe2eðtÞ, can be determined as
Pe2eðtÞ ¼ Sn2cðPn1ðtÞÞ � PCðtÞ � S �1c2n ðPn2ðtÞÞ: ð3Þ
Suppose each service component in a composite network–compute system has a LR profile; that is, Pn1 ¼P½r1; θ1�,PC ¼P½rC; θC�, and Pn2 ¼P½r2; θ2�. Then the transformedprofile for network service 1 is
PSn1 ¼ Sn2cðPn1Þ ¼P½Sn2cðr1Þ; θ1� ð4Þ
and the transformed profile for network service 2 will be
PSn2 ¼ S �1
c2n ðPn2Þ ¼P½S �1c2n ðr2Þ; θ2�: ð5Þ
Then it can be proved that the end-to-end capabilityprofile of the composite service provisioning system is
Pe2e ¼ Sn2cðP½r1; θ1�Þ � P½rC; θC� � S �1½Pðr2; θ2Þ�¼P½Sn2cðr1Þ; θ1� � P½rC; θC� � P½S �1
c2n ðr2Þ; θ2�¼P½re; θe� ð6Þ
where
re ¼min Sn2cðr1Þ; rC;S �1c2n ðr2Þ
n o; θe ¼ θ1þθCþθ2: ð7Þ
Eqs. (6) and (7) imply that for a composite network–compute service provisioning system, if all service compo-nents, including both forward and backward network
Fig. 6 Alternative model for composite network–computeservice provisioning.
Fig. 7 End-to-end delay model for composite network-Cloudservice provisioning.
187Modeling and performance analysis for composite network–compute service provisioning in software-defined
services and the compute service, can be modeled by LRprofiles, then the end-to-end service capability of thecomposite provisioning system can also be modeled by aLR profile with the network/compute scaling curve in frontof it and the compute/network scaling curve behind it. Thelatency parameter of the end-to-end LR profile is thesummation of latency parameters of all service componentsin the system. The service rate parameter of the end-to-endLR profile is determined by the minimum value of the Cloudservice rate and the transformed service rates of forwardand backward network services.
Suppose the data transform between network and com-pute services can be characterized by a piece-wise linearscaling curve SðaÞ ¼mini ¼ 1;⋯Nfpiaþqig, then S �1 bð Þ ¼maxi ¼ 1;⋯N
1pi½b�qi�þ
n o. Therefore, the transformed cap-
ability profile for network services 1 and 2 will be
PSn1 ¼ S
�Pn1Þ ¼ min
i ¼ 1;⋯Npi½r1ðt�θ1Þ�þqi
� � ð8Þ
and
PSn2 ¼ S �1 Pn2ð Þ ¼ max
i ¼ 1;⋯N
1pi½r2ðt�θ2Þ�qi�þ
� �ð9Þ
6. Delay performance analysis for compositenetwork–compute service provisioning
Based on the service model presented in the precedingsection, this section will develop a delay analysis techniquefor composite network–compute service provisioning. Theanalysis focuses on the maximum service response delay(between the time instants when a user sends out a requestto the Cloud data center and receives the correspondingresponse), which is a significant performance parameter forhigh-performance Cloud applications.
Delay analysis for a service provisioning system needs anapproach to characterizing the load on the system since asystem with a certain service capacity achieves differentlevels of delay performance under different loads. Due tothe diversity in traffic generated by various Cloud applica-tions, a general profile is needed to specify the load forcomposite service systems. The concept of arrival curvefrom network calculus is employed here to define a generalload profile as follows.
Let R(t) denote the accumulated amount of traffic thatarrives at the entry of a composite network–computeservice provisioning system by any time instant t. Given anon-negative, non-decreasing function, Lð�Þ, the servicesystem is said to have a load profile LðtÞ if for all timeinstants s and t such that 0osot
RðtÞ�RðsÞrLðt�sÞ: ð10ÞA load profile gives an upper bound for the amount of trafficthat the end user can load on a service provisioning system.Since the profile is defined as a general function of time, itcan be used to describe traffic load generated by anyapplication.
Currently most QoS-capable networking systems applytraffic regulation mechanisms at network boundaries toshape arrival traffic from end users. The traffic regulators
that are most commonly used in practice are leaky buckets.A networking session constrained by a leaky bucket con-troller has a traffic load profile
L½p; ρ; σ� ¼min pt; σþρt� �
; ð11Þwhere p, ρ, and σ are respectively called the peak rate,sustained rate, and maximal burst size for the traffic.
The maximum service delay guaranteed by a compositenetwork–compute system to an end user is determined bytwo factors: (i) service capacity offered by the system tothe user, which is modeled by the end-to-end capabilityprofile; and (ii) characteristic of the traffic that the userloads the system, which is described by a load profile. Theentry of the composite network–compute service systemwhere user applications load the system is the boundary ofthe forward networking system; therefore the traffic loadfor forward data transmission is the load of the compositeservice system. Based on the alternative model given inFig. 6, service delay performance is determined by the end-to-end profile Pe2eðtÞ because scaling functions do notintroduce any delay. Due to the network/compute scalingcurve Sn2c in front of the end-to-end profile, the actual loadthat determines service delay performance should becharacterized by a transformed load profile LSðtÞ ¼Sn2cðLðtÞÞ, as shown in Fig. 7. Therefore, given the end-to-end capability profile Pe2eðtÞ of a composite network–compute service system, the maximum service delay dmax
guaranteed by the system to the user can be determined as
dmax ¼ maxt:tZ0
min δ : δZ0 Sn2cðLðtÞÞrPe2eðtþδÞn on o
: ð12Þ
Suppose a composite network–compute system has a LRprofile for each service component and a leaky-bucket loadprofile L½p; ρ; σ�, then the transformed load profile for thesystem is Sn2cðL½p; ρ; σ�Þ. Following (6) and (11), the maximumservice delay guaranteed by this system can be determined as
dmax ¼ θΣþSn2cðpÞ
re�1
� � Sn2cðσÞSn2cðpÞ�Sn2cðρÞ
ð13Þ
where θΣ ¼ θn1þθn2þθC is the total service latency includinground-trip network latency and latency of the compute service.
The latency parameter of a LR profile for a network servicereflects a system property of the network that may be seen asthe worst-case delay experienced by the first traffic bit in abusy period of a networking session. Therefore, this parametercan be estimated based on link transmission delay and packetprocessing delay, if signal propagation delay is assumed to beignorable; that is, θffiL=rþL=R, where L and R are respec-tively the maximum packet length and maximum link rate ofthe network service.
Suppose the forward and backward networks services ofthe composite network–compute service provisioning systemhave identical link rate R and packet length L, then the
Q. Duan188
maximum response delay performance of the compositeservice becomes
dmax ¼ LXi ¼ 1;2
1riþ 1
Ri
� �þθCþ
pre
�1� �
σ
p�ρ: ð14Þ
7. Numerical results
This section gives numerical examples for illustrating appli-cations of the developed modeling and analysis techniques.Considering the service provisioning system as shown inFig. 2, in which an end user transmits data to a Cloud datacenter for processing and then receives the processed databack from the Cloud. Based on the measurement resultsreported in [2,3], traffic parameters of the load profile forthis testing case are assumed to be 320 Mb/s, 120 Mb/s, and200 kbits for the peak rate, sustained rate, and burst sizerespectively. For simplicity, forward and backward datatransmissions are assumed to be provided by the samenetwork service with a 10 Gb/s link capacity and a packetlength of 1500 bytes.
A communication intensive service scenario was firstexamined. In this scenario data transform from forwardtransmission to the computing server decreases the loadwith a scaling factor Sn2c ¼ 1=4 and data transform fromcomputing server to backward transmission increases theload with a scaling factor Sc2n ¼ 2. The impacts of networkservice rate and computing server capacity on the maximumservice delay were analyzed and the obtained results areshown in Fig. 8. In this scenario we considered two cases inwhich latency parameters of the computing server are150 μs and 300 μs. The obtained end-to-end delay upperbounds for both cases are denoted respectively as de
c1 anddec2 in Fig. 8. From this figure we can see that both curves
drop with increasing network service rate, which indicatesthat leasing more bandwidth from the network serviceproviders may significantly improve end-to-end delay per-formance for composite service provisioning. Comparing thetwo curves of de
c1 and dec2 shows that given the same network
service rate, smaller server latency may give a tighter end-
Fig. 8 End-to-end service delay vs. network service rate for acommunication intensive application.
to-end delay bound but its impact is not as significant asthat of increasing network bandwidth.
The relationship between the maximum service delay andavailable computing capacity in this scenario was alsoanalyzed with a given network service rate. The obtainedresults are plotted in Fig. 9, in which de
n1, den2, and de
n3respectively denote the service delay bounds when networkservice rate is 150, 200, and 250 Mbps. The three curves inFig. 9 are all basically flat, which implies that given acertain network service rate, increasing computing servercapacity has almost no impact on the maximum end-to-enddelay of this service case. Comparing the three delay curvesin Fig. 9 shows that higher network service rate gives lowerservice delay. This is because data transmission betweenuser and the computing server forms a service bottleneck inthis scenario that determines the worst-case delay perfor-mance; therefore allocating more server capacity in datacenter does not help improving end-to-end delay perfor-mance. Notice that Fig. 8 indicates that less latency at thecomputing server does help reducing service delay to acertain degree. However, this latency parameter is mainlydetermined by data center implementation such as hard-ware and operating system speed, which cannot be easilyreduced. Therefore, the above results shown in Figs. 8 and 9tell us that for such communication intensive services,leasing sufficient bandwidth from network service providersinstead of purchasing more computing service capacity isthe most effective strategy for achieving an end-to-endservice delay guarantee.
Then we examined a computing intensive service scenarioin which data transforms between network and computeservices increase the load for data processing with a scalingfactor Sn2c ¼ 2 and decrease the load for data transmissionwith a scaling factor Sc2n ¼ 1=4. For this scenario the end-to-end delay upper bounds achieved with different networkservice rates were analyzed and the results are plotted inFig. 10. The delay bounds were obtained with threecomputing capacity values, rc ¼ 100, 125, and 150 Mbps,and their delay bounds are denoted as de
c1, dec2, and de
c3respectively in Fig. 10. All three curves in this figure droponly slightly with increasing network service rate, whichimplies that leasing more bandwidth from network service
Fig. 9 End-to-end service delay vs. compute server capacityfor a communication intensive application.
Fig. 10 End-to-end service delay vs. network service rate fora computing intensive application.
Fig. 11 End-to-end service delay vs. compute server capacityfor a computing intensive application.
189Modeling and performance analysis for composite network–compute service provisioning in software-defined
makes little contribution to reducing end-to-end servicedelay in this scenario. Comparing the three delay curvesshows that for a given network service rate, increasingcompute service capacity may significantly improve end-to-end service delay performance in this scenario.
The end-to-end service delay bounds with various com-pute service capacities are given in Fig. 11, in which de
n1,den2, de
n3 denote the delay bounds obtained with 100, 150,and 200 Mbps network service rates. This figure shows thatmaximum delay of the composite service decreases signifi-cantly with increasing compute service capacity, whichconfirms the observation we obtained from Fig. 10. Fig. 11also shows that the three delay curves are very close toeach other although not completely overlap, which impliesthat given the same compute service capacity, differentamounts of network bandwidth do not change the delaybound much. The results shown in Figs. 10 and 11 indicatethat in this scenario computing server capacity is thedecisive factor for the maximum service delay and increas-ing network service rate only has minor contribution todelay performance improvement. This is because the com-pute service forms a bottleneck for this computing intensivescenario; therefore obtaining sufficient computing capacityin the data center is the key to achieving end-to-end delayguarantee for the composite service.
8. Conclusions
The crucial role of networking in Cloud computing requires aholistic vision of both networking and computing resourcesin a software-defined Cloud environment, which leads tocomposition of network and compute service provisioning.The research presented in this paper studies the problem ofmodeling and performance analysis for composite network–compute service provisioning systems. The main contribu-tions made in this paper include a new approach tomodeling service capabilities of composite network–com-pute service systems and analysis techniques for evaluatingdelay performance of composite network–compute services.Application of the recent development in network calculuswith data scaling makes the modeling and analysis techni-ques developed in this paper general and agnostic to
network and Cloud implementations; thus are applicableto various heterogeneous networking and computing sys-tems coexisting in a Cloud environment for compositenetwork–compute service provisioning. Both analytical andnumerical results obtained in this paper indicate thatcapacities of network and compute services as well as thedata transform factors between these two types of serviceshave direct impact on delay performance of compositeservices. The developed model and analysis techniquesprovide Cloud service customers, service brokers, andservice providers with an effective tool for evaluatingachievable service performance in software-defined Cloudenvironments.
References
[1] I. Foster, Y. Zhao, I. Raicu, S. Lu, Cloud computing and Gridcomputing 360-degrees compared, in: Proceedings of the 2008Grid Computing Environment Workshop, November 2008,pp. 1–10.
[2] K.R. Jackson, K. Muriki, S. Canon, S. Cholia, J. Shalf,Performance analysis of high performance computing applica-tions on the Amazon Web services Cloud, in: Proceedings of the2010 IEEE International Conference on Cloud Computing Tech-nology and Science, November 2010, pp. 159–168.
[3] G. Wang, T.S.E. Ng, The impact of virtualization on networkperformance of Amazon EC2 data center, in: Proceedings ofthe IEEE INFOCOM2010, March 2010, pp. 1–9.
[4] ONF, Open Networking Foundation Software-Defined Network-ing (SDN) Definition, ⟨https://www.opennetworking.org/sdn-resources/sdn-definition⟩, 2013.
[5] C. Li, B. Brech, S. Crowder, D. Dias, H. Franke, M. Hogstrom,D. Lindquist, G. Pacifici, S. Pappe, B. Rajaraman, J. Rao,R. Ratnaparkhi, R. Smith, M. Williams, Software-definedenvironments: an introduction, IBM J. Res. Dev. 58 (March)(2014) 1–11.
[6] Q. Duan, Y. Yan, T.V. Vaslikos, A survey on service-orientednetwork virtualization toward a convergence of networkingand Cloud computing, IEEE Trans. Netw. Serv. Manag. 9 (4)(2012) 373–392.
[7] Q. Duan, Network-as-a-Service in Software-Defined Networksfor end-to-end QoS provisioning, in: Proceedings of the 2014IEEE Wireless and Optical Communications Conference, May2014.
Q. Duan190
[8] L.-J. Zhang, Q. Zhou, CCOA: Cloud computing open architec-ture, in: Proceedings of the First Symposium on NetworkSystem Design and Implementation, April 2009.
[9] N. Feamster, L. Gao, J. Rexford, How to lease the Internet inyour spare time, ACM SIGCOMM Comput. Commun. Rev. 37 (1)(2007) 61–64.
[10] V. Stantchev, Performance evaluation of Cloud computingofferings, in: Proceedings of the Third International Confer-ence on Advanced Engineering Computing and Applications inSciences, October 2009, pp. 187–192.
[11] Z. Hill, M. Humphrey, A quantitative analysis of high perfor-mance computing with Amazon's EC2 infrastructure: The deathof the local cluster?, in: Proceedings of the 10th IEEE/ACMInternational Conference on Grid Computing, October 2009,pp. 26–33.
[12] N. Yigitbasi, A. Iosup, D. Epema, S. Ostermann, C-meter: aframework for performance analysis of computing Clouds, in:Proceedings of the Ninth IEEE/ACM International Symposiumon Cluster Computing and the Grid, 2009, pp. 472–477.
[13] A. Iosup, S. Ostemann, M. Yigitbasi, R. Prodan, T. Fahringer,D. Epema, Performance analysis of Cloud computing servicesfor many-tasks scientific computing, IEEE Trans. Parallel Dis-trib. Syst. 22 (6) (2011) 931–945.
[14] G. Juve, E. Deelman, K. Vahi, G. Mehta, B. Berriman, B.P.Berman, P. Maechling, Scientific workflow applications onAmazon EC2, in Proceedings of the Fifth IEEE InternationalConference on e-Science Workshop, December 2009.
[15] Z. Li, L. O'Brien, R. Cai, H. Zhang, Towards a taxonomy ofperformance evaluation of commercial Cloud services, in:Proceedings of the Fifth IEEE International Conference onCloud Computing, June 2012, pp. 344–351.
[16] K. Xiong, H. Perros, Service performance and analysis in cloudcomputing, in: Proceedings of the IEEE 2009 World Conferenceon Services, 2009, pp. 693–700.
[17] W. Ellens, M. Zivkovic, J. Akkerboom, R. Litijens, H. Berg,Performance of Cloud computing centers with multiple priority
classes, in: Proceedings of the Fifth IEEE International Con-ference on Cloud Computing, June 2012, pp. 245–252.
[18] B. Yang, F. Tan, Y. Dai, S. Guo, Performance evaluation ofCloud service considering faulty recovery, in: Proceedings ofthe First International Conference on Cloud Computing,December 2009, pp. 571–576.
[19] H. Khazaei, J. Misic, V.B. Misic, Performance analysis of Cloudcomputing centers using M/G/m/m+r queuing systems, IEEETrans. Parallel Distrib. Syst. 23 (5) (2012) 936–943.
[20] H. Khazaei, J. Misic, V.B. Misic, Performance analysis of Cloudcenters under burst arrivals and total rejection policy, in:Proceedings of the IEEE Globecom2011, December 2011, pp. 1–6.
[21] J.-Y.L. Boudec, P. Thiran, Network calculus: a theory ofdeterministic queueing systems for the Internet, in: SpringerLecture Notes in Computer Science, 2001.
[22] Q. Duan, Network service description and discovery for high-performance ubiquitous and pervasive Grids, ACM Trans.Auton. Adapt. Syst. (TAAS) 6 (1) (2011) 3.
[23] Q. Duan, Analysis on quality of service provisioning forcommunication services in network virtualization, J. Commun.7 (2) (2012) 143–154.
[24] Q. Duan, End-to-end modelling and performance analysis fornetwork virtualisation in the next generation internet, Int. J.Commun. Netw. Distrib. Syst. 8 (1) (2012) 53–69.
[25] Q. Duan, Modeling and performance analysis on networkvirtualization for composite network-Cloud service provision-ing, in: Proceeding of the 2011 IEEE World Congress onServices, July 2011, pp. 548–555.
[26] Q. Duan, Modeling and delay analysis for converged network-Cloud service provisioning systems, in: Proceedings of the 2013IEEE International Conference on Computing, Networking &Communications, February 2013, pp. 66–70.
[27] M. Fidler, J.B. Schmitt, On the way to a distributed systemscalculus: an end-to-end network calculus with data scaling, in:Proceedings of SIGMetrics/Performance'06, 2006, pp. 287–298.
