Stochastic Virtual Infrastructure Planning in Elastic Cloud Deploying ...

1 downloads 7151 Views 501KB Size Report
Jan 23, 2012 - Cloud computing services apart from IT requirements have to also support the associated network requirements. In the current formulation it is ...
OFC/NFOEC Technical Digest © 2012 OSA

Stochastic Virtual Infrastructure Planning in Elastic Cloud Deploying Optical Networking Markos P. Anastasopoulos, Anna Tzanakaki and Konstantinos Georgakilas Athens Information Technology, Peania, Greece {manast,atza,koge}@ait.gr

Abstract: This paper proposes stochastic planning suitable for elastic clouds deploying optical networking. A novel model employing stochastic traffic considerations for virtualization of the underlying physical resources is proposed and validated achieving significant cost savings. OCIS codes: (060.4254) Networks, combinatorial network design; (060.4256) Networks, network optimization.

1. Introduction As the scale of information processing is increasing, with the expectation to reach the level of Exabytes at the end of this decade [1], and new applications such as UHD IPTV, 3D gaming, virtual worlds etc. are emerging, novel network solutions are required to support the Future Internet. These new high-performance applications, need to be supported by specific IT resources, that maybe remote and geographically distributed, requiring connectivity with the end users, according to the cloud computing paradigm, through a very high capacity and increased flexibility and dynamicity network. Optical networks satisfy these requirements and it is a strong candidate to support this need. In order to maximize the utilization and efficiency of the converged optical network and IT infrastructure, virtualization of physical resources [2] can be additionally applied. The concept of virtual infrastructures (VIs) facilitates sharing of physical resources among various virtual operators introducing a new business model that suits well the nature and characteristics of the Future Internet and enables new exploitation opportunities for the underlying physical infrastructures (PIs). Through the adoption of VIs, optical network and IT resources can be deployed and managed as logical services, rather than physical resources. This effectively enables the offering of resizable and on demand computing and network capacity on a pay-as you go basis as a web service, known as elastic cloud. Supporting the paradigm of elastic cloud introduces challenges in the infrastructure design which requires accurate estimation of both network and IT resources to avoid either inability to satisfy end-users requirements or unnecessarily high operational and capital expenditures (OpEx and CapEx). In this paper, we propose for the first time, a Stochastic Integer Programming (SIP) model [3], suitable for the planning of VIs taking into account both the time variability and uncertainty of services over an integrated IT and optical network infrastructure. Existing VI planning schemes [2,4,5] do not take into account the stochastic nature of services, a parameter that is inherently imposed in elastic cloud. However, this may have a significant impact regarding the ability of the infrastructure to support the required services and the associated CapEx and OpEx 2. Stochastic Virtual Infrastructure Planning The objective of stochastic VI planning is to identify the topology and determine the virtual resources required to implement a dynamically reconfigurable VI based both on optical network and IT resources. This VI not only will meet customer’s specific needs, that may be either known in advance or predicted based on history observations, but will also satisfy the VI provider’s (VIP’s) requirements for minimum OpEx and CapEx. Since information concerning customer’s specific needs is uncertain, the VI planning problem is formulated through an SIP model that aims at minimizing jointly the CapEx and OpEX of the optical network components and IT resources. In this study, the PI is described through a topology corresponding to the Pan-European optical network in which randomly selected nodes generate demands d (d=1, 2, …,D) of service class i (i=1,2..,I) that need to be served by a set of IT servers s (s= 1,2,..,S). Each service class may require either a steady state usage of the total network and IT resources for a specific time period or may need only to run for a percentage of time over this period. An additional consideration is that the various service classes are associated with a particular delay priority, to which stricter or more relaxed delay boundaries are assigned in order to address delay sensitive (DS) or delay tolerant (DT) services, respectively. For example, critical data could be sent for processing and storage to an IT server at a specific point in time bound by a maximum tolerable delay while data of low importance could be served anytime within a much longer time period. The set of stages through which demand d can be realized by the VIP is denoted by Td , while the set of stages at which some resources reserved by demands d could be utilized at stage t is denoted by Fdt , t ϵ Td. Cloud computing services apart from IT requirements have to also support the associated network requirements. In the current formulation it is assumed that the granularity of network demands is the wavelength while the IT locations at which the services will be handled, are not specified and are of no importance to the services themselves. Therefore, identification of the suitable IT resources is part of the optimization output. To formulate this

©Optical Society of America OW1A.3.pdf 1

1/23/2012 11:50:41 AM

OFC/NFOEC Technical Digest © 2012 OSA

requirement the binary variable αdsi is introduced to indicate whether demand d belonging to service class i is assigned to IT server s or not. This variable takes value equal to 1 if and only if demand d is processed on server s. It is also assumed that each demand can be assigned to only one IT server. Furthermore, for each demand d at time t belonging to service class i, its network demand volume hdti is supported by means of a number of lightpaths assigned to the corresponding VI paths. This demand volume has to be served at t ϵ Fdt. In practical cloud computing scenarios information concerning hdti is not precisely available to the VIP. However, for every provisioning state t it can be described by a probability distribution function (pdf). Let Ξ denote the set of all possible scenarios in every planning stage and Ξt the set of all scenarios in provisioning stage t. The set Ξ is defined as the Cartesian product of all Ξt, namely Ξ=Πt Ξt. It is also assumed that the pdf of Ξ has finite support, i.e., finite number of scenarios with corresponding probabilities p(ξ) ϵ [0,1] where ξ is a composite variable defined as ξ=(ξ1, ξ2,…, ξ|Τ|). In this paper hdti are scenarios in Ξ, namely hdti (ξ), with known pdfs. Taking into account both optical network and computing demands, let p=1,2,…,Pdti be the candidate path list in (r ) ( ) the requested the VI for the lightpaths required to support demand d in time t for service i at server s , and xdpti

non-negative number of lightpaths allocated to path p for demand d that belongs to service i at time t and scenario ξ. During the VI planning process, the requested resources will be rearranged based on service-specific constraints. In the planned VI, the wavelength conservation constraints should be satisfied (r) xdpti      tˆF xdpti (1) d  1, 2..., D , i  1, 2, ..., I , t  1, 2, ..., T , p  1, 2, ..., Pdti ˆ   , as well as the demands constraint d  1, 2..., D, i  1, 2, ..., I , t  1, 2,..., T , (2)  s  p adsi xdpti     hdti    , dt

Summing up the lightpaths through each link e (e=1,2,…,E) of the VI we can determine the required link capacity yet(ξ) for link e required at time t and scenario ξ. Using the same rationale, the capacity of each link e in the VI at time t is allocated by identifying the required lightpaths in the PI. The resulting PI lightpaths determine the load of each link g (g = 1, 2, …, G) of the PI, and hence its capacity ugt (ξ) at time t and scenario ξ. During the VI to PI mapping process the demand constraints for link e as well as the PI capacity constraint should be satisfied. So far, the SIP problem formulation guarantees that the capacities of the VI and PI links will be adequate to support the transmission of the cloud computing services over the network segment. However, once the information arrives at its destination, the IT server s should have adequate capacity φsrt of resource type r to support the traffic scenario ξ. The incoming information to an IT server is mapped from an optical network type of requirement to an computing resource.For details concerning this process refer to [6].Then, based on service-specific constraints, these demands are probabilistically scheduled at any stage t for which the IT server has adequate resources. Note that, the demands constraints for IT resources of type r as well as the capacity constraints of resource r at every server should be also satisfied. The objective of the SIP formulation is to minimize the total cost during the planned horizon of the resulting network configuration that consists of the following components: a) kg the cost of the capacity of PI link g, b) σsr the cost of the capacity resource r of IT server s. Both costs are related to the energy consumption of the infrastructure which to a large extent determines the OpEx of the infrastructures. Details of the energy consumption models may be found in [2, 5]. Finally, the stochastically planned VI is obtained by minimizing the expected cost: T min    t 1 

 k u g

g

gt

    s  r 

sr



φsrt   

(3)

This problem has been solved employing the Sample Average Approximation (SAA) approach [3]. 3. Numerical results and Conclusions To investigate the performance of the proposed VI design scheme, the multilayer network architecture illustrated in Fig.1 is considered. The lower layer depicts the PI and the layer above depicts the VI. For the PI the COST239 PanEuropean reference topology has been used in which stochastically generated traffic demands following a specific pdf need to be served by 3 IT servers. For simplicity, the granularity of service duration for the generated services is the time period. Furthermore, we assume a single fiber per link, 40 wavelengths (wv) per fiber, wavelength channels of 10Gb/s each and that each IT server can process up to 2Tb/s. An example of the stochastic VI topology design for a planning horizon of 4 stages is depicted in Fig.1 for a traffic scenario with the following characteristics: i)4 source nodes generate demands normally distributed with a mean of 60wv and a standard deviation of 20wv, ii) the number of arrivals in any given time interval [0,t] follows the Poisson distribution with λ=1 time period, iii) service times follows the exponential distribution with mean value μ=1 time period, iv) DT services can be scheduled with a maximum delay of 2 time periods while DS require instant access to the services, v) each wavelength requires 10Tb/s of processing power. As depicted in Fig.1a, the associated VI topology for the first two time periods consists of 3 virtual links and 4 virtual nodes, while all demands corresponding to DS services are routed to one IT server.

OW1A.3.pdf 2

1/23/2012 11:50:41 AM

OFC/NFOEC Technical Digest © 2012 OSA

Fig. 1 Virtualization of a PI: a,b) Design with stochastc traffic considerations (DT, DS arrive equiprobably) and c) without stochastic traffic

Fig.2. a)Overall Energy consumption of the planned VI with and without stochastic traffic considerations b)Utillation of network resources

For the following time periods the planned VI topology consists of 4 virtual links and 5 virtual nodes (Fig.1b). Again all demands corresponding to DT services are forwarded to the same one IT server. Note that the dashed lines in the PI correspond to all unused physical links. In Fig.1c it is seen that without stochastic optimization additional IT and network resources are activated for a longer time period e.g., 2 IT servers are activated for 4 time periods. In Fig.2a, the performance of the proposed energy aware VI design with stochastic considerations is compared to the scheme presented in [2] where the VI is planned for its entire operating horizon assuming a predetermined set of permanent services that the VI needs to support. Comparing these two schemes it is seen that the stochastic model achieves significant overall energy savings compared to the deterministic. Also, the average power consumption for the stochastic scheme increases almost linearly with the number of demands (for various DT and DS services combinations), in contrast to the deterministic where a stepwise increase of the power consumption is observed. For traffic demands above 50 wavelengths per source, the stochastic VI scheme consumes significantly lower energy for serving the same amount of demands compared to the deterministic VI design scheme of the ordnger of 35%, as in the former approach fewer IT servers are activated to serve the same amount of demands. Given that the power consumption required for the operation of the IT servers is dominant in this type of converged infrastructures, by appropriately scheduling demands and switching-off unused IT resources significant reduction of the energy consumption is provided. Moreover, for large traffic demands the stochastic scheme achieves lower utilization of the optical network resources leaving a higher level of network resources available for future use (Fig.2b). This is due to that, the proposed model allows for the optimization of the allocation of both network and IT resources, by accurately estimating and decoupling the optical network and IT resource requirements associated with each individual service request. Note that as the number of requests for DS services increases, less optical network resources are employed by the stochastic model to cover the same amount of demands. Statistically more IT servers per time period and per geographic region are activated leading to on average shorter end-to-end lighpaths and effectively smaller network utilization. Acknowledgments This work was carried out with the support of the GEYSERS (FP7-ICT-248657) project funded by the European Commission through the 7th ICT Framework Program. 4. References [1] M. Handley, “ Why the Internet only just works”, BT Technology Journal, Vol. 24, No 3, 2006 [2] A.Tzanakaki et al., "Energy Efficiency in integrated IT and Optical Network Infrastructures: The GEYSERS approach", in proc. of IEEE INFOCOM 2011, Workshop on Green Communications and Networking (2011) [3] P.Kall, J.Mayer, “Stochastic Linear Programming: Models, Theory, and Computation”, Kluwer Academic Publishers, 2005. [4] A.Tzanakaki et al., “Energy Aware Planning of Multiple VIs over Converged Optical Network and IT Physical Resources”, in Proc. ECOC 11 [5] A.Tzanakaki et.al, "Dimensioning the future Pan-European optical network with energy efficiency considerations," JOCN 3, 272-280 (2011) [6] Standard Performance Evaluation Corporation (SPEC) (www.spec.org)

OW1A.3.pdf 3

1/23/2012 11:50:41 AM

Suggest Documents