cost structure of the network service provider cloud platform is poorly understood. Thus, this empirical research uses a single company single case study ...
2016 IEEE International Conference on Cloud Engineering Workshops
Cost modeling of a network service provider cloud platform Jarno Lähteenmäki
Heikki Hämmäinen
Department of Communications and Networking Aalto University Espoo, Finland jarno.lahteenmaki@aalto.fi
Department of Communications and Networking Aalto University Espoo, Finland heikki.hammainen@aalto.fi
Nan Zhang
Matti Swan
Department of Communications and Networking Aalto University Espoo, Finland nan.zhang@aalto.fi
Elisa Plc. Helsinki, Finland matti.swan@elisa.fi
production has based on service dedicated platforms and tight integration with vendors [4]. This same model does not apply as such in the network service provider cloud concept due to modular software based approach. Instead, network providers need overall capability to manage the platform and wide network of vendors.
Abstract—The cloud service model is spreading into the telecommunications industry, forcing the telecommunications providers to adjust their production models. The network service provider cloud platform, also called as Telco Cloud, utilizing a cloud based production model, is a solution to respond to the changing market forces. The cloud based production model is a new concept in the context of telecommunication services and the cost structure of the network service provider cloud platform is poorly understood. Thus, this empirical research uses a single company single case study approach to examine the network service provider cloud cost structures and constructs a cost model to network service provider cloud by using semi-structured interviews. The results indicate that the network service provider cloud improves cost efficiency compared to the traditional platform model due to server virtualization and consolidation. The contributed cost model can be used for analyzing feasibility of an application on-boarding in network service providers that are utilizing cloud platforms for the communications services.
There are a few scholars who have studied the field. Krzywda et al. [5] propose a meta-model for the Telco Cloud framework. The model can be used for simulating different configurations, thus, observing behaviour of the system from performance and cost perspectives. Zhiqun et al. [6] discuss of emergence of the Telco Cloud paradigm. They observed that operators can achieve cost reduction by introducing cloud computing platforms. Scholars typically make distinction between mobile [7] and fixed [8] providers when handling network service provider cloud issues, while this study handles generally network service providers regardless which type of network is operated.
Keywords—network service provider cloud, Telco Cloud, Cost structure, Carrier Cloud, Cost simulation
I.
I NTRODUCTION
The economic problem that telecommunication operators are facing is declining per user revenues due to the fact that modern internet based services are substituting traditional telecommunications services [1]. Furthermore, growing use of cellular data [2, p.13] is increasing mobile operator costs, thus, reducing profits. At the same time, service price erosion of the telecommunications services [2, p.14] forces operators to introduce new enhanced over-the-top (OTT) and value-addedservices (VAS) to the market. This initially increases research and development (R&D) costs and eventually also revenues.
This paper is focusing on the cost analysis of the cloud service model for producing telecommunications services. The objective is to gain understanding how a cloud production model fits into the telecommunications sector. This study contributes to the network service provider cloud cost model. To get proper view of the afore mentioned issues, the qualitative and explorative research is carried out. The empirical data collection is carried out using semi-structured interviews in the case company. From these results, analysis is carried out using the qualitative data analysis techniques. The analysis is working as a base for the final contribution of this study.
Network vendors started to develop new cloud based platforms for communications services, and coined the Telco Cloud concept, in early 2010s [3]. However, small number of published scientific papers indicates that scholars have not yet found this field of research. Several areas exist that need more thorough understanding before incumbents can rely on this new concept. Traditionally the telecommunications service
The rest of this paper is organized as follows. Section II briefly summarizes how the data is collected. Section III describes the technical structure of the network service provider cloud platform. Section IV analyzes the research data and outlines the cost model. Section V concludes the study by summarizing the key findings and contributions.
/16 $31.00 © 2016 IEEE DOI 10.1109/IC2EW.2016.40
148
II.
DATA C OLLECTION
incremental investment requirement is larger. Single right level does not exist, but each infrastructure provider has to define their optimal level that meets their requirements.
Source data in this work is gathered through eight semistructured interviews with experts in the field of the telecommunication services from the case company. The interviews were carried out between May 29th and June 24th, in the year 2015. The case company is a service provider that operates globally but the main operational area is in Finland. The company is providing mainly telecommunication, over-the-top, and enterprise IT services, including infrastructure as a service (IaaS). Revenue of the case company was over 1,5 billion euros in the financial year 2014 and it had almost 4100 employees.
The application layer contains virtual servers running on top of the virtualization layer. The cloud manager requests resources from a node manager for a virtual server, storage or network resource. The cloud manager also controls a virtual network function (VNF) application provisioning and unprovisioning functions. In this context, an application is any collection of software components that form one VNF entity, e.g. DNS, MME or HSS.
The interviewed persons are chosen to represent different kind of positions in the case company, e.g. lead architect, purchase manager, CTO, solution manager and service manager. The discussed topics include cloud service and network service provider cloud as paradigms, structure and financial aspects of the network services provider cloud concept, and comparison between the traditional and cloud based network service platform models. The detailed interview process and transcripts are documented in [9].
The network service provider cloud model contains also several support systems which fulfil the platform. The monitoring system watches the platform from the hardware to the application level. A configuration management database (CMDB) contains the repository of all physical and virtual components that the system contains, including all software versions of applications running in virtual servers. The application life-cycle manager controls which version of the service application is running on what node and for what customer segments. The testing module tests the system and especially the application layer so that it is compliant with the service provider and industry specific requirements.
The data categorization reveals three major view-points to the network service provider cloud area. (i) The paradigm and technical structure of it. (ii) Strategic capability for telecommunications operators. (iii) The cost structure of the network service provider cloud platform. The last view-point is chosen for further analysis as this study is focusing to economical factors of the network service provider cloud model. However, it should be noted that the technical structure of the platform can have an affect how cost structure is formed. Hence, also the technical structure is analyzed and described so that the context of the cost structure analysis would be unambiguous. III.
IV.
C OST MODELLING
Purpose of cost modelling is to reveal differences between cost components and expose relations between those. Based on that, model simulations can be carried out to test and evaluate behavior of the model. A. Activity-based Costing Modeling The Activity-based Costing (ABC) model is building on the concept of activities and it is used to model the network service provider cloud cost structure. An ABC activity is a task or operation that people or machines do. The first part of this analysis is to identify the activities. The tasks, events, incidents and processes that are defined by the interviewees are categorized into activities.
T ECHNICAL STRUCTURE OF THE PLATFORM
The analysis indicates that the platform structure can affect the cost structure. Furthermore, the network service provider cloud platform is somewhat dependent on other assets that a network service provider has. Here are summarized most relevant parts of the structure and the full technical structure is documented in [9].
The analysis reveals 13 different activities. All relevant cost sources are included in the categories. It should be noted that this cost structure purposefully exclude any other administrative costs that are not directly related to the network service provider cloud operations. Company complementary assets, such as marketing and sales are highly company specific and are challenging to generalize and are not seen relevant for the model. Structure of these costs is not analyzed in this model.
The network service provider cloud model relies on geographical distribution. There are two kinds of data centers utilized. This study proposes naming core data center and regional data center. A core data center is a site that is in a central position and a regional data center is a site that is serving customers on some geographical region. Cloud resources are located in different data centers. The platform contains logical sub-unit called a node. A node is an autonomic container that has servers, switches and storage devices. Server, storage and network virtualization create resource pools that are controlled by the node manager.
Relations between the categories are studied and linkages between activities are defined in the form of allocation bases. These activities cover several processes that an organization has to execute in order to build, maintain and develop the network service provider cloud platform.
Nodes can be distributed to several core and regionl data centers for distribution and redundancy purposes and each data center may contain several nodes. Hence, the size of a node defines also the incremental cost to introduce new capacity. When smaller nodes are used, smaller increments can be made, but in this case the relative share of the cloud layer becomes significant. However, when larger node size is used, the overhead of the cloud layer is minimized but
Not all activities require their own cost allocation base, instead several activities can share the same cost allocation base. The structure of the cost model can be seen in figure 1. A cost allocation base class steers how costs are divided from activities to different cost objects, so it functions as a
149
Configuration management is an activity to ensure that all versions and life-cycle phases of components are recorded and managed coherently. This information is part of the centralized configuration repository and it can also be used for real time cost analysis and for charging external parties using the platform. The configuration data is also used as a base for any platform simulation.
Proactive platform maintenance is to keep the platform functioning properly with proactive measures. This contains proactive defect patching planning, security threat monitoring, life-cycle management, problem management and monitoring.
Fig. 1. Activity-based Costing model for the network service provider cloud platform
Capacity management is to monitor the actual usage of the hardware platform so that also redundancy capacity is secured in the case of failure. The capacity planning, capacity monitoring development, forecasting, geographical coverage planning, measuring and road-mapping are part of this activity.
driver. An cost object can be seen as a service or VNF which the service prodiver is offering. The analysis reveals four cost allocation base classes. These are total number of applications, number of configuration items, total amount of memory and direct costs.
Server operations activity covers all costs related to continuous operations of the hardware platform. Costs such as electricity and cooling, facility costs, hardware recurring maintenance costs, networking costs and incident costs, including work hours related to this activity.
The total number of applications class forms the fixed cost part of the platform. It divides cost pool equally to all cost objects, thus to all applications running on top of the platform. These costs can be considered as a sustaining cost of the platform. Activities that are using this allocation base are the cloud development, proactive platform management, platform specification management, human resource management, and internal and external communication.
Software operations activity covers all costs related to continuous operations of the software platform, such as recurring software licenses for cloud, virtualization, automation and management systems. Platform specification management activity contains tasks such as creating and maintaining the clear specification of the platform for virtualized network function (VNF) applications.
Capacity usage driven costs are forming of the number of configuration items and total amount of memory allocation bases. The number of configuration items allocation base divides cost pools according to the number of configuration items recorded in the repository. Total amount of memory allocation base divides cost pools according to cumulative usage of the memory capacity in the platform by applications.
2) Application specific actities: Application specific activities are application on-boarding and application development. Application on-boarding activity is to introduce a new application to the platform. This includes tasks such as the software requirement engineering, software development, testing, validation, integration and project management. This is per application activity that realizes only when a new application is introduced to the platform or vendor’s activities trigger this activity. In the early phases of the platform, an operator is expected to pay most of the on-boarding costs. When the network service provider cloud paradigm matures and standardization proceeds, a software vendor is expected to take care of the most of the on-boarding and integration costs due to competitive reasons.
Application specific direct costs are allocated directly to corresponding cost objects, thus to applications. Activities such as the application on-boarding and application development are directed into the direct cost allocation base. These four allocation bases are cost targets for total of 13 activities. These activities are functioning as cost pools for all platform related expenses. These can be divided into three categories; platform activities, application specific activities and organizational activities. 1) Platform activities: Platform activities are hardware installation, asset management, cloud development, configuration management, proactive platform maintenance and capacity management activities. Hardware installation introduces new servers, network switches and other required hardware when new data centers are built or existing ones require a capacity expansion. The cost allocation unit is the number of servers installed.
Application development must be done if the platform specification is changed due to a component update. This triggers a need for the application and service development and testing activity. 3) Organizational activities: Organizational activities are human resource management and internal and external communication. Human resource management activity is to ensure that, each new person is introduced and trained for the duties. Furthermore, when new features are introduced or existing features are changed, training for existing personnel should be organized. Internal and external communication is an important activity to keep a linkage between platform development and operations, user organizations and external stakeholders such as vendors and regulators.
Asset management activity cost is mostly due to CAPEX depreciations, thus, a software or hardware asset amortizations. This involves also personnel costs if a work activity has been capitalized during the development phase. Cloud development is an activity to develop the cloud manager, automation, virtualization, provisioning interfaces and integrations to other systems. This activity contains also costs of continuous integration team or other software development activities related to the platform.
150
B. Mathematical modeling
TABLE II.
The ABC cost structure can also be presented in the mathematical form. Table I contains equations that can be used to calculate the annual costs for each activity and total annual cost of the platform. Furthermore, platform sustaining costs, capacity usage based costs and direct costs can be identified by summing up appropriate activities as described in figure 1. Table II contains definitions for the variables and values used in the simulations. Electricity and cooling cost model builds on the device average power consumption, the average data center power usage effectiveness (PUE) and size of a node in servers. Deployment cost of a node is total sum of servers, network switches and data center related costs, such as racks, cablings and power feeds. In hardware installation, time used per site is assumed constant, internal work cost is also constant and the hardware is replaced as soon as it is depreciated. Cloud development activity cost is the annual personnel cost of cloud software developers. Configuration management costs consist of configuration management database software license costs, which are expected to be licensed by the number of physical servers. TABLE I.
I NITIAL VALUES
Variable Nt Nc Ns Nsw Nspare
Value 20 pcs 3 pcs 17 pcs 5 pcs 10 pcs
Nappl
8 pcs
Npers Nnewpers
2 persons 10 nodes
Measuring values Description Servers per node Servers reserved for the cloud layer per node Servers reserved for the application capacity per node Network switches per node Servers reserved for a spare, testing and redundancy use on total level Average resource requirement per application converted to physical servers Number of initial personnel in operational activities New operational person recruited every Nth new node
Variable Ps Psw Epue
Value 400 W 60 W 2
Technical specification values Description Power consumption per server Power consumption per network switch Average data centre PUE value
Variable Cm Cs Csw Cdc Ds Cw Ccm Ccl
Value 120 C 10 000 C 10 000 C 20 000 C 5 years 900 C 120 C 1300 C
Variable Thw
Value 5 workdays 320 workdays 1 hour
Financial values Description Cost of MWh electricity power Average purchase price of a server Average purchase price of a network switch Average purchase price of a rack and cabling Depreciation time of a server and a network switch hardware Internal day price for specialist work Configuration management annual per unit software cost Cloud layer software annual license and support costs per server
T HE MATHEMATICAL COST MODEL
Description Total power consumption per node Total electricity cost per node per year Total node deployment cost (CAPEX) Asset management activity annual cost, hardware depreciations Hardware installation activity annual cost Cloud Development activity annual cost Configuration management annual cost Proactive management and Capacity management annual activity costs Server operations annual activity costs Software operations annual activity costs Platform specification management annual activity costs Human resource management annual activity costs
Equation Pn = Nt Ps + Nsw Psw Ce = Pn Cm Epue (365 ∗ 24) Cn = Cs Nt + Csw Nsw + Cdc N Cn Cam = nodecount Ds
Internal and external communication and Vendor Management activity costs Total annual cost
Ccomm = Tvm Cw
Cinst =
Tcd Tca
Nnodecount Cw Thw Ds
Tpl Thr
Ccd = Cw Tcd Cconf = Nnodecount Nt Ccm Ccapa = Nnodecount Nt Cca
Tvm
5 workdays 5 workdays 4 workdays
Time estimates for activities Description Hardware installation: Time required to install a hardware for a new node Cloud development: Annual time required for the activity Proactive management and Capacity management: Annual time required per physical server for the activities Platform specification management: Annual time required for the activity HR activities: Required annual work time per employee Vendor management and Internal communication: required annual work time per application
Chwops = Nnodecount Ce Cswops = Nnodecount Nt Ccl Cplatf orm = Tpl Cw Chr
=
Nnodecount Nnewpers
Thr Cw (Npers
relative share of the Cloud development activity cost is high. When the number of nodes increases, the Asset management activity cost becomes the biggest. Another major cost is the Software operations, e.g. commercial cloud license costs; and the Server operations, e.g. electricity costs.
+
) (Ns −Nspare ) Nappl
Ctotal = Cam + Cinst + Ccd + Cconf + Ccapa + Chwops + Cswops + Cplatf orm + Chr + Ccomm
In the second simulation, the variable is the number of applications in the platform. In the network service provider cloud model, the average cost per application decreases very quickly when the number of applications increases in the platform. The illustrative graph can be seen in figure 3. This indicates that the economics of scale is realized in this model. The x-axis in the figure represents the number of applications in the platform. The y-axis represents the average cost per application.
C. Cost Structure Simulation During the analysis, the cost for each activity is estimated to evaluate the cost allocation behavior. In this simulation, the scenario variable is the number of nodes. Several assumptions are made for the estimation based on experts’ judgement. Table II contains values for estimated parameters.
In a traditional application dedicated platform model, no costs exist related to the Cloud development and Cloud licensing. Furthermore, each application requires dedicated hardware infrastructure. On the other hand, the required number of servers is higher due to the redundancy requirement and lack of statistical multiplexing leverage. Hence, there are no synergies between different applications. Applying the same assumptions as in the network service provider cloud model, the average annual cost per application for the traditional model is 113 units per application. In figure 3, the red line is illustrating the
Technical specification values are collected from literature sources and from vendor specifications. The average server and network switch electricity consumptions are estimated using the tool released by the hardware vendor [10]. The average DC PUE of 2.0 is used as a compromise [11]. Figure 2 illustrates estimated sharing of the cost between different activities. When the number of nodes is small, the
151
sufficient to justify the costs of the cloud related activities.
When the application count reaches 40, the average cost per application is 71 units in the network service provider cloud model. Thus, when compared to the traditional model, the calculated model indicates that the saving is more than 30 percent.
"
In a traditional platform model the short run incremental cost (SRIC) and long run incremental cost (LRIC) stay quite constant due to the static nature of the platform, while in the network service provider cloud the dynamic nature is actually quite strong. The network effects in the platform enable lowered production cost for all services when a new service is introduced to the same platform. Furthermore, while the application layer is detached from the hardware layer, the life-cycle management process can optimize the hardware layer independently. Eventually, it is expected that even the cloud layer might be replaced if interfaces remain as specified by the standard. This modular and standardized model brings several benefits for network service providers.
!
Fig. 2.
The cost structure as such can be seen as a non-operational tool for analyzing the behavior of the platform. Some implications can be derived to the operational level from this cost structure. As seen earlier, in a large scale operation, the biggest cost source is hardware and software asset depreciations. Hence, to ensure a productive efficient market, an operator has to make sure that the number of servers is kept at low level. This can be achieved by choosing the right combination of CPU and memory intensive applications to the platform. Thus, high level of statistical multiplexing is pursued. However, this is a dynamic optimization problem that might be very hard to solve. The cloud scheduling optimization is a NP-hard problem, that might be solvable in a polynomial time but there is no guarantee of it [12]. Hence, the simulation approach taken by Krzywda et al. [5] is actually a sound solution for saving costs. The large scale simulation and scenario analysis of workloads may reveal ways to optimize the platform, thus, to avoid unnecessary hardware purchases.
Calculated cost share between activities
D. Limitations of the Study
The ABC cost model is known to lack an ability to reveal the long run incremental costs [13]. On the other hand, used ABC and mathematical models do not reveal well dynamic nature of the system. For example, electricity pricing is highly location and time dependent and might have significant impact of simulation results.
!
Fig. 3.
The simulation of the study contains assumptions which limits applicability of the results. Initial values used in the simulation are not fully accurate due to the high level of averaging and single case study approach. Thus, these estimates should not be used as a base for any quantitative research and these simulation results should be considered as qualitative indications of the phenomena.
Calculated average cost per application
constant cost of traditional model applications. The blue line illustrates decreasing costs per application when the number of applications increases.
This study concentrates on direct financial implications. This selection is done mostly due to the problem statements in the case company. Thus, technological and social aspects that may have some financial consequences are mostly excluded. Furthermore, the telecommunication provider view point is selected. The regulative and legislative region chosen for this
The interesting value is the break-even point between the traditional and network service provider cloud cost models. With these estimations, the break-even point is when the number of applications reaches seven. This corresponds to 66 servers. Thus, relatively small number of applications is
152
research is Finland due to main operational location of the case company. V.
ACKNOWLEDGEMENT The authors would like to thank all interviewed persons in the case company.
C ONCLUSIONS
R EFERENCES
The objective of this study is to analyze cost structure of the network service provider cloud platform from the telecommunication operator point of view. Finnish operator planning to implement a network service provider cloud platform is selected as a case company and several managers are interviewed for acquiring information on the subject.
[1]
[2]
The research shows that the network service provider cloud paradigm is heavily based on the general cloud computing paradigm. However, the telecommunication market has industry specific requirements, such as high availability and redundancy that are increasing the service level of requirements.
[3]
Based on interviews, 13 different activities are defined for the cost structure. The network service provider cloud cost model has four cost allocation bases. First, application onboarding and development costs are directly allocated for a product as a direct cost. Second, configuration management costs are allocated using number of configuration items. Third, a cloud development, proactive platform management, platform specification, HR management and communication costs are allocated using number of applications as a driver. Finally, asset management, server and software operations, capacity management and hardware installation costs are allocated using amount of memory assigned for an application.
[4] [5]
[6]
The cost structure shows that both fixed and capacity usage driven cost activities exist. The economy of scope can be seen in the cost structure. The cost per application decreases radically when the number of applications and nodes increases. In the traditional model, all activities are direct costs for a product, thus, the economy of scope is not realized as such. The main difference between a traditional telco platform and the network service provider cloud is the efficient use of hardware assets in the network service provider cloud model.
[7]
[8]
The used assumptions set the break-even point to seven applications, thus, the network service provider cloud platform is more cost-effective than the traditional one if there are more applications than seven. More important than the exact number is how the associated costs develop in the model. It can be seen that when the total cloud capacity is small, share of cloud development activity is a dominating cost. When the capacity grows and the number of nodes increases, the share of the cloud development activity decreases and the asset management activity becomes the dominant. Hence, firstly, when the platform grows, the costs are moving from fixed activity costs to capacity usage dependent activities. Secondly, the economics of scope is greatly reducing per application costs by sharing fixed costs among all applications in the platform. These make the network service provider cloud platform more cost-effective than the traditional platform.
[9]
[10]
[11]
[12]
This study uses a static model of the platform and does not reveal any dynamic behaviour of it. In particular, a sensitivity analysis should be carried out to find out effect of different variables in the model. Furthermore, A multi-case study of the same area shall be carried out to ensure that results are more generalizable.
[13]
153
Finnish Communications Regulatory Authority, Arpu (average revenue per unit), 2014. [Online]. Available: https : / / www . viestintavirasto . fi / en / aboutthesector / statistics/communicationsmarket/arpu.html (visited on 03/09/2015). ——, Toimialakatsaus 2013, 2013. [Online]. Available: https : / / www . viestintavirasto . fi / attachments / toimialatieto / Toimialakatsaus _ 2013 _ fi _ webb _ .pdf (visited on 01/20/2015). P. Bosch, A. Duminuco, F. Pianese, and T. Wood, “Telco clouds and virtual telco: consolidation, convergence, and beyond,” in Integrated Network Management (IM), 2011 IFIP/IEEE International Symposium on, May 2011, pp. 982–988. S. Pocuca, “Service delivery platforms (sdp) for telco applications,” in MIPRO, 2011 Proceedings of the 34th International Convention, 2011, pp. 478–482. J. Krzywda, W. Tärneberg, P.-O. Östberg, M. Kihl, and E. Elmroth, “Telco clouds: modelling and simulation,” in Closer 2015, Lisbon, Protugal, 2015. [Online]. Available: http : / / lup . lub. lu . se / record / 5431120 (visited on 08/05/2015). X. Zhiqun, C. Duan, H. Zhiyuan, and S. Qunying, “Emerging of telco cloud,” Communications, China, vol. 10, no. 6, pp. 79–85, Jun. 2013, ISSN: 1673-5447. V. Yazıcı, U. Kozat, and M. Oguz Sunay, “A new control plane for 5g network architecture with a case study on unified handoff, mobility, and routing management,” Communications Magazine, IEEE, vol. 52, no. 11, pp. 76–85, 2014, ISSN: 0163-6804. A. Blenk, A. Basta, M. Reisslein, and W. Kellerer, “Survey on network virtualization hypervisors for software defined networking,” arXiv preprint arXiv:1506.07275, 2015. J. Lähteenmäki, Emergence of a telco cloud platform: paradigm, cost structure and strategic implications, master’s thesis, 2015. [Online]. Available: http : / / urn . fi/URN:NBN:fi:aalto-201511205185. Dell Inc., Dell energy smart solution advisor, 2015. [Online]. Available: http://essa.us.dell.com/DellStarOnline/ DCCP.aspx?c=us&l=en&s=biz&Template=6945c07e3be7-47aa-b318-18f9052df893 (visited on 07/30/2015). R. Miller, “Uptime Institute: The Average PUE is 1.8,” Tech. Rep., 2011. [Online]. Available: http : / / www . datacenterknowledge . com / archives / 2011 / 05 / 10 / uptime- institute- the- average- pue- is- 1 - 8/ (visited on 07/26/2015). M. Tawfeek, A. El-Sisi, A. E. Keshk, F. Torkey, et al., “Cloud task scheduling based on ant colony optimization,” in Computer Engineering & Systems (ICCES), 2013 8th International Conference on, IEEE, 2013, pp. 64–69. C. Courcoubetis and R. Weber, Pricing communication networks: Economy, technology and modelling. Wiley, 2003, ISBN: 978-0-470-85130-2.
