7th Latin American Workshop On Communications - 2015 Design Of Network Infrastructure Of A Cloud Data Center For Use In Health Sector Chris Talavera Julio Santisteban Urb. Campiña Paisajista s/n Barrio Urb. Campiña Paisajista s/n Barrio de San Lázaro, Arequipa, Perú de San Lázaro, Arequipa, Perú Universidad Católica San Pablo Universidad Católica San Pablo chris.talavera@ucsp.edu.pe jsantisteban@ucsp.edu.pe Abstract—This article presents the design of the network in- Cloud, Community Cloud and Hybrid Cloud which combine frastructure of a Data Center that meets the requirements arising two or more forms of clouds (private, community or public) from Cloud Computing, for use in the Health Sector of Arequipa [2], [8], [3], [12]. city, focusing on network layer 2 and its dimensionality to meet the requirements of several health service applications. The Cloud infrastructure consists of data centers that hosts network infrastructure dimensionality calculation is a complex servers and using different levels of organization or virtu- challenge for an of the ground project , in this article we present alization techniques it offers cloud services [24]. A logical a novel approach to solve this challenge. view of a Cloud Data Center (CDC) shown in 1. This Index Terms—Data Center, Cloud Computing, Network Design. model represents the basic components or building blocks of any CDC. This view introduces encapsulation and insulation I. I NTRODUCTION layers and impose support system modularity. There are differ- ent layers: infrastructure, databases, middleware, applications, We live in a connected world. Almost two billion people management, monitoring and security layer, which one have connect to the Internet and to address this need the community specific roles and consolidated once formed the Data Center of information technology has created a new service deliv- in Cloud. ery mechanism called "Cloud Computing". In the healthcare industry, Cloud Computing might be a paradigm shift in the use of information technology, among others: transparent man- III. S TATE OF THE A RT agement and access to electronic health records of patients, There are many benefits by incorporating Cloud Computing secure and reliable data storage and transmission, automation in the healthcare industry, but to implement that, the design processes, streamlining workflow and consolidate assets of of a Data Center of next generation is necessary, thereby, information technologies for providers of healthcare services; some services providers have developed a reference archi- thus leading to obtain a higher quality of service. tectures, for example Cisco [20], proposes an architecture Cloud computing especially facilitate the provision of which consists of three blocks: the first block is composed healthcare products and services to patients in remote areas by network, computing and storage, this layer houses all the and those who have limited access to quality medical services. services provided to consumer. The second block is security For that reason, comunication infraestructure has to be power- layer, the key point is that security should be end-to-end full and it needs a hardy data center. Having a data center is not architecture. The third layer is about infrastructure and services a new idea, but they need to make some changes to support managment. This architecture just shows goals to take account the specific characteristics of Cloud Computing in the most on the creation of Cloud Data Center but does not deliver a optimal way. Therefore, this article shows how to design a clear methodology. network infrastructure using as a stege the MINSA (Ministerio Concerning the design of the data center network on [6] de Salud) namely system of Healthcare in Arequipa, Peru. can be found the more used topologies types, as a the Fat Tree topology, consisting of two sets of elements, the core and II. T HEORETICAL F RAMEWORK Pods; the Bcube topology that was proposed for Modular Data The National Institute of Standards and Technology (NIST) Center, building to allow installation and procedures simpler define Cloud Computing as a technology model that enables physical migration compared with regular Data Centers and ubiquitous, adapted and demand access network to share a DCELL topology defined recursively and uses servers for set of configurable computing resources that can be quick packet forwarding [25]. provisioned and released with management efforts reduced or Another important issue of Cloud Data Center are the minimal interaction of the service provider [2], [8]. The main virtualization techniques, respecto that [14] shows evidence features of Cloud Computing are self-demand, comprehensive that the latest network technologies have not been developed network access, resource pooling, scalability, it is based on keeping in mind the needs of virtualization, and as a result, the supply of services mainly Software as a Service (SaaS), the network can become a bottleneck for these implementa- Platform as Service (PaaS) and Infrastructure as a Service tions. This article, also expose that static topologies require (IaaS) and there are 04 types of Cloud: Public Cloud, Private manual intervention to deploy and migrate virtual machines, Copyright © 2015 for the individual papers by the papers’ authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors. Latin American Workshop On Communications' 2015 Arequipa, Peru Published on CEUR-WS: http://ceur-ws.org/Vol-1538/ Figure 2. IT Parameters Year Attentions 2011 2 770 054 2012 2 920 191 2013 3 078 465 2014 3 245 318 2015 3 421 214 2016 3 606 644 2017 3 802 124 2018 4 008 199 Table I P ROJECTION OF USERS - ANUAL ATTENTIONS A. Current Situation of MINSA Overall, the potential beneficiaries in healthcare industry is the staff working in MINSA: health professionals, admin- istrative staff and patients. On [Guias MINSA] it is shown that in Peru there are various categories of establishments which respond to different social and health realities and they are designed to meet demands equivalent. Thus, the level of complexity of the care services is directly related to health service development, specialization and modernization of its resources. There are 111 health facilities located in the province, which are distributed as I-1, I-2, I-3, I-4, II-1, II-2, II-E,III-1, III-2, III-E. Figure 1. Reference Architecture Cloud Data Center which adds cost and hinders the ability of the organization B. Design Parameters to respond quickly to changes in the environment, for that For proper planning process of infrastructure cloud data reason OpenFlow is presented as an open source standard center, three fundamental IT parameters has to be considered: designed to address these shortcomings. Based on Ethernet criticality, capacity and growth or expansion plan. It is shown technology, OpenFlow separates the data path and control in the 2 that only criticality and growth plan directly affects path by an independent controller. This introduces a new the design of the network infrastructure [16]. network abstraction layer, analogous to server virtualization, According to [5]] to choose this parameters there are several in consequence allows the network to act as a single structure. methods for example the TIER UPTIME which gives 4 levels The benefits are simplicity, being open, scalable and fast. of availability. A second method is tied to TIA 942 [1], [17] where the division of 4 levels or Tiers is standard: TIER IV. D ESIGN OF C LOUD DATA C ENTER I for basic infrastructure without redundancies, TIER II for Infrastructure components with redundant capacity, TIER III In this section is proposed the solution of Cloud Data for redundancy N+1 and TIER IV infrastructure for fault- Center, the first step is identified the current stage of the tolerant 2(N+1). For a healthcare cloud data center it is healthcare industry specially the main beneficiaries; in the considered TIER IV. second stage design parameters are defined. The third process The first step of design of Cloud Data Center is understand- is develop the analysis of network traffic; in the fourth step ing the needs of the healthcare industry, therefore, the number different network topologies are identified and compared with of network users considering the use of statistical data was each other in order to choose the best performance. The final projected, as Perú has a constant growth, the average annual step is to perform the dimensioning of links and finally the growth rates is 5.42% . The projection per each year is shown Data Center interconnect with each of the health centers. in I and II. Year Medical Staff Administrative Staff 2011 3 173 851 because Cloud Computing, as part of scalability, automatic re- 2012 3 419 973 sources allocation is performed using mechanisms autoscaling 2013 3 684 1 113 where alarms are configured appropriately to respond in the 2014 3 969 1 273 best way to a requirement, precisely the most used algorithms 2015 4 276 1 456 2016 4 607 1 665 keep on queuing theory [11], [13], [21], [10], [18]. 2017 4 964 1 904 The queuing model used is denoted as M/M/c/c, Where M 2018 5 348 2 178 is a system of arrivals that occurs according to Poison process Table II P ROJECTED NUMBER OF WORKERS IN MINSA ratio of λ, where the arrival times are exponentially distributed with mean µ, c represents the number of servers and the maximum number of customers system´s allowed (when c + 1 Implicated Fc (%) ρhora (Erl.) ρhorapico (Erl.) 80 122 140 requests coming into the system, the service is denied for the Patient 65 99 114 latter). 40 6 7 In addition, as a parameter of quality of service has decided 80 3 4 Medical Staff 65 2 2 to consider the total response time of the service(s) for a Cloud 40 2 2 Data Center should not be over 450ms [18]. It may have been Administrative Staff 100 2 2 chosen as a quality parameter the CPU utilization of the server, Table III T RAFFIC INTENSITY BY TYPE OF HEALTHCARE USERS which according to[13] should be at least 85%. Thus, following Little relations and queuing theory, the following relationship was obtained (1) [18], [7], [19]. C. Traffic Analysis µ s= λ (1) ∗µ In this section the calculation of minimum, maximum and 1 + cn margin for the network throughput is found, in this way the goal is comply with the parameters of future growth. Using the where: concurrency factor (CF) which determines the ratio between s : Average service time total simultaneous users and users who use the network in the λ : Arrival rate day, not having accurate statistics, an analysis is made for each µ : Service time involved in the healthcare industry. c : Number of servers Criteria or considerations for the calculation of traffic: n : Number of cores server 1) The peak time is 10 to 20% of daily traffic, so they will The first required parameter is the arrival rate to the system take 15% to make the calculations. (λ), Number that can be taken as the maximum network 2) FC for each involved in the health industry was found, throughput, 141 Erlangs plus margin of 15 Erlangs, ie 161 IE, patients, health and administrative staff. Erlangs. Regarding the length of service, is necesary to know how long a server take to process a request, no doubt this The III summarizes the data obtained and thereby the require- parameter is random, but an approach can be arrive with ment for the network is known. some tests such as those in [18], where different instances Respecto to growth parameters defined, the IV shows the are analyzed in Amazon, so the average value is 178ms (µ). minimum, maximum and margin capacity. It has also considered a single core server (n = 1) And the average service time is 450ms. The number of required servers D. Definition of Network Servers is 48, which should form clusters or it have to be virtualized. Although the determination of servers of CDC is essential, In order to test these results, real cases have been investi- there is no standard way to find the exact number of these gated, in this way it is possible to have a more realistic idea devices [18], this reality is that any service provider that offers of how many servers would be required in an environment of Cloud Computing had to start him infrastructure from zero, is Healthcare. So, first a survey was conducted to people involved actually found in the process of adapting their traditional data of Information Technology area with goal to know the used center to the new trend. way that they use to perform sizing of servers and the most In this paper, the number of servers was calculated based on used applications; the survey and its results can be seen in [15]. the modeling of the process to entry to themselves, using the On the other hand, statistics of the use of networks, servers queuing theory and prefixing a parameter of quality of service and applications that run on public institutions was obtained, as: time of service or the CPU usage threshold. This idea borns as well as the number of concurrent users that it houses. The important thing is to know how many cores of 1GHz each Parameter Quantity (Erl) Total (Erl) institution uses on their network and how many concurrent Max. Throughput 146 161 users are allowed. In theV, the information is shown. Min. Throughput 127 142 The data presented show a ratio factor equal to 0.049, Margin 15 through it the necessary number of cores is calculated to allow Total 288 Table IV 421 users, which is the number of jobs per peak hour at this G ROWTH PLAN PARAMETERS stage. The total number of servers to use is 27, according to real statistics, this result shows that the formula previously Institution Cores 1GHz % Use Real Cores Concurrent Optimizes All Uplink Extesión Scaling users the density in active VLAN between link Provincial Municipality of 11.90 80 9.52 200 of Access state support Switch Arequipa Switch District Municipality of 34.36 60 20.62 350 Trinagle NO YES NO NO Cerro Colorado Loop Arequipa Judiciary 59.20 70 41.44 800 Square Loop YES YES YES YES Catholic San Pablo 96.00 60 57.60 1300 U free Loop YES NO YES NO University Inverted U YES YES YES YES Table V free loop S TATISTICAL INFORMATION FROM SERVERS AT DIFFERENT INSTITUTIONS Table VII C OMPARING DESIGNS L AYER 2 NETWORK Fat - Tree DCell BCube Scalability Good Excelent Good Incremental Good Poor Poor Scalabily The hierarchical model divides networks into modular blocks: Agility Yes Yes Yes access layer, distribution, and core, the next step in design to Cabling Easy Very Dificult Dificult CDC consist of to select features for each layer in order to Switch fault Poor Good Good tolerance improve network performance. Thus, the core layer should be Link fault tolerance Good Very Poor Poor work on Layer 3 of the OSI model to enable the core links Server fault God Very Poor Poor to achieve scalability, rapid convergence and to avoid risk of tolerance uncontrollable broadcast. Throughput Constant Incremental Incremental degradation degradation The aggregation layer is very important as this determines Cost Regular Low Low the stability and scalability of the entire data center network, Tráffic balance Yes No Yes Table VI as recommended in [22], it is best to model the aggregation C OMPARATION BETWEEN TOPOLOGIES OF C LOUD DATA C ENTER layer switches with pairs of interconnected modules that provide services such as content switching, firewall, intrusion detection, and network analysis. Redundancy is important to consider, in this sense, integrated services will be defined in used to calculate the number of servers, allows us to have a the "active/active" mode. reliability of about 57%, which can be improved if we use The access layer works in layer 2 and the model with square another queue. loop was chosen, because its resistance to failure is greater compared to model-free loop in addition, the comparison made E. Network Topology in the VII shows that this topology provides benefits such as: Each topology network has several advantages regarding extension of VLAN, virtual machine mobility, service module performance, remember that there are many dimensions to redundancy. characterize this parameter, such as: latency, bandwidth, cost, In this CDC network design a subnet storage must be con- resistance to failure, etc. sidered, specifically a SAN (Storage Area Network) because it In VI a summary of the comparison of technologies is is a subnet with high speed storage devices. It is an important presented, considering the above data and some others taken part of design therefore it allows a high throughput and lowest from [6] and its translation to the different dimensions of latency which creates a high performance across the network. performance. To find the size of the links in the network, calculate the By the above comparison, it can be stated that the hierar- current and future demand for traffic per user is needed, chical Fat-Tree topology is the best suited for network design therefore, an estimated analysis of the various applications Cloud Data Center. Even though Fat-Tree topology is not per- and services that use each involved in the industry is made fect in fact its biggest problem is the emergence of bottlenecks health. But this analysis of traffic must not specifically take in the root of the tree, but its advantages and differences with each application else must make a distinction made by type of other network topologies make to take the decision to use this traffic. It is important to note that various services of Cloud design topology of the network architecture. Computing (SaaS, PaaS or IaaS), does not introduce a new Considering the traffic analysis and the procedures per- traffic pattern themselves instead, they should be seen as a formed to find the number of network servers, the number new way of consuming different resources [20]. of ports required for each server can be calculated, because For each applications or services more important the traffic a fat-tree topology is constructed by k-ports and can support ua calculated, considering in each case the concurrency factor, 3 a 100% throughput performance between k4 servers, using k2 VIII shows the results. border switches and k2 aggregation [6], [23]. Then, the analysis establishes that the peak bandwidth Therefore, theoretically it has: required by the network user is 3.16 Gbps. To avoid saturation 1) Number of ports: 6 on network ports, these should be at least twice the calculated 2) Number of pods: 6 capacity, ie. about 6.31 Gbps. Therefore the network ports of 3) Number of core switches: 6 access switches must be 10 Gbps. 4) Number of aggregation switches: 3 To calculate the speed of the backbone links distribution 5) Number of access switches: 3 Poisson formula is used to find the probability of arrivals to Individual Capacity Traffic Type Individual Capacity (Kbps) Applications Total TRraffic (Mbps) (Kbps) Telephony over IP 22.58 Telephony over IP 88.8 37.69 Video over IP 884.74 Vídeo over IP 2 530.0 1 073.86 Messaging 0.088 Mail 2.58 11.97 Data Bases 94.38 Data bases 94.38 893.02 File Sharing 11.38 Share files 11.38 59.95 Internet Download 11.38 Internet Download 11.38 68.52 Access Web Pages 56.89 Acces to Web Page 56.89 342.50 Table IX Complementary 669.8 691.80 I NDIVIDUAL CAPACITY BY TYPE OF TRAFFIC services Total (Mbps) 3.39 3 156.46 Table VIII T OTAL CAPACITY FOR NETWORK SERVICES Figure 4. General view of Cloud Data Center distribution of MINSA patients and staff by level of care and health establishment category, so individual capacity traffic Figure 3. Network Diagram type is shown at IX. On the other hand, because the information handled in the healthcare industry is very delicate, it is important to consider the up-link ports, based on 2. a backup to the whole network, but for Cloud Computing the e−λ (λ)r current traditional model of active Data Center and passive P (r) = (2) Data Center, has to be replaced by a new model of extended r! single data center, in which the different locations DC look as Where: if they were a single seat and the service is actively provided P (r) : Probability of arrivals to up-link ports from different physical locations. Therefore the network in r : Number of arrivals to up-link port general, will be seen as shown in 4. λ : Average rate of arrivals to up-link port To calculate, we need the number of ports of each switch, at this case 6 but adding redundancies will take as approx. V. C ONCLUSIONS AND F UTURE W ORK 12 ports. Thus, assuming that switches of 12 ports is used, 1) It has been identified the technical mechanisms required the number of simultaneous arrivals is at least 12, the average for the design of network infrastructure Cloud Data speed is 12 arrivals per unit time and probability of arrival in Center, these are: Criticality through which we can the up-link will be 0.11437. The result is used to calculate the choose according to the characteristics of applications speed links up-link Access Switch, by3, proposed by[9], [4]. available network; Capacity and Growth, these design factors set out to find the maximum and minimum network load and an expansion margin considering it V el.ptosup−link ≥ (N úm.ptos)∗(V el.ptoshalf −duplex )∗P (r) should be a short time because it is active equipment (3) and technology in general. Therefore, the above result is determined the speed uplink 2) This work has completed an estimate of network traffic, ports it must be greater than 13.7244Gbps, so the ports should based on an analysis of the reality of health facilities and be 40 Gbps or 100 Gbps for the smooth operation of the in general of the MINSA (Ministry of Health), it is also switch and the entire network is ensured. The network design thought of short growth of the number of beneficiaries. is shown at 3. Thus, it is estimated that the network requires links 10 and 40 GbE. On the other hand, via a mathematical F. WAN Interconnection formula validated through statistical defined design that To find the speed of the WAN links that reach health facil- requires about 48 servers of 1 core. ities traffic demand of each one must be calculate. To achieve 3) A data center is a centralized area for storage, handling this, the first step is to calculate the individual requirements of and distribution of data and information, which consists each person according to the type of traffic and then make a of several components such as network infrastructure, services infrastructure, infrastructure management, mon- [14] Riso Mehra. Design and building a datacenter network: An alternative itoring, including other. Each has a specific work to be approach with openflow. Technical report, Corporación NEC, 2012. [15] Chris Talavera Ormeño. Diseño de la Infraestructura de Red bajo el performed optimally allows the entire system to function modelo de Computación en la Nube para su uso en el Sector Salud de properly. Indeed, this work has a significant contribution Arequipa. 2015. on this point, because although data center is not a new [16] Neil Rasmussen and Suzanne Niles. Data center projects: System planning. Technical report, American Power Conversion, 2007. issue, Cloud Data Center is it and to take in account [17] Reichle Y De-Massari AG (RYM). RYM Data Center. 2011. issues performance to take right decisions of design is [18] Huber Flores Satish Srirama and Michele Mazzucco. Performance necessary. testing of cloud applications, interim release. REMICS Consortium 2010-2013, 2012. 4) This article focused on network infrastructure, because [19] JM Sidi and Asad Khamisy. Single server queueing models for this is the main part of a Cloud Data Center because communication systems, 2011. it acts as the heart of communication. Thus, a thorough [20] Cisco Systems. Cisco computación en la nube - data center strategy, architecture and solutions. Technical report, Cisco Systems, 2009. investigation of the features and functionality changes, [21] Wenhong Tian. Adaptive dimensioning of cloud data centers. In new considerations and approaches that should be taken Dependable, Autonomic and Secure Computing, 2009. DASC’09. Eighth into account in order to design a Cloud Data Center was IEEE International Conference on, pages 5–10. IEEE, 2009. [22] Jhon Tiso. Designing Cisco Network Service Architecture. Cisco Press, performed. 2012. 5) Taking into account the above considerations, the design [23] Amin Vahdat, Mohammad Al-Fares, Nathan Farrington, Radhika Niran- of the network infrastructure of a Cloud Data Center was jan Mysore, George Porter, and Sivasankar Radhakrishnan. Scale-out networking in the data center. IEEE micro, 30(4):29–41, 2010. proposed. Such design has important features are listed [24] Fabio Luciano Verdi, Christian Esteve Rothenberg, Rafael Pasquini, and below: M Magalhaes. Novas arquiteturas de data center para cloud computing. Modular design, with good scalability. XXVIII Simpósio Brasileiro de Redes de Computadores e Sistemas Distribuídos - Gramado RS, 2010. • Allows easily detect network failures and it is a net- [25] Kaishun Wu, Jiang Xiao, and Lionel M Ni. Rethinking the architec- work with redundancy that allows combat failures ture design of data center networks. Frontiers of Computer Science, 6(5):596–603, 2012. • Access quickly to storage devices via the SAN subnet. • Work with virtualization allowing the use of physi- cal resources effectively.. VI. B IBLIOGRAPHY R EFERENCES [1] Norma ansi/tia94. [2] "cloud computing synopsis and recommendations", 2012. [3] Introduction to cloud computing architecture. White paper, Sum Mi- crosystems, 2012. [4] Maroa Aguirre Patiño, Rut Ester España, Ivï¿œn Solí Granda, and Alfonso Aranda Segovia. Diseño y simulación de un data center cloud computing que cumpla con la norma pci-dss. 2011. [5] Victor Avelar. Guidelines for specifying data center criticality/tier levels. American Power Conversion (APC), pages 2007–0, 2007. [6] Rodrigo S Couto, Miguel Elias M Campista, and Luis Henrique MK Costa. A reliability analysis of datacenter topologies. In Global Communications Conference (GLOBECOM), 2012 IEEE, pages 1890– 1895. IEEE, 2012. [7] David de la Fuente García and Raúl Pino Díez. Teoría de líneas de espera: modelos de colas. Universidad de Oviedo, 2001. [8] "Observatorio Nacional de las Telecomunicaciones y de las TI". Com- putación en la nube retos y opotunidades, 2012. [9] Vanessa Garay Olivo. Estudio y diseño de un centro de asistencia remota para una empresa de soporte de equipos oftalmológicos utilizando voz e imágenes fijas y móviles sobre ip. Master’s thesis, Escuela Politécnica Nacional, 2013. [10] Daniel Gmach, Jerry Rolia, Ludmila Cherkasova, and Alfons Kemper. Capacity management and demand prediction for next generation data centers. In Web Services, 2007. ICWS 2007. IEEE International Conference on, pages 43–50. IEEE, 2007. [11] V Goswami, SS Patra, and GB Mund. Performance analysis of cloud with queue-dependent virtual machines. In Recent Advances in Information Technology (RAIT), 2012 1st International Conference on, pages 357–362. IEEE, 2012. [12] Yashpalsing Jadeja and Kirit Modi. Cloud computing-concepts, ar- chitecture and challenges. In Computing, Electronics and Electrical Technologies (ICCEET), 2012 International Conference on, pages 877– 880. IEEE, 2012. [13] Hamzeh Khazaei, Jelena Misic, and Vojislav B Misic. Performance analysis of cloud computing centers using m/g/m/m+ r queuing systems. Parallel and Distributed Systems, IEEE Transactions on, 23(5):936–943, 2012.