Energy aware planning of multiple virtual infrastructures over converged optical network and IT...

Energy aware planning of multiple virtual

infrastructures over converged optical network

and IT physical resources

Markos P. Anastasopoulos,1 Anna Tzanakaki,

1,* Konstantinos Georgakilas,

1 and

Dimitra Simeonidou2

1Athens Information Technology, 17.km Markopoulou Ave., Peania, Athens, GR - 19002, Greece 2University of Essex, Colchester, UK

*[email protected]

Abstract: This paper studies energy efficient planning of multiple

concurrent virtual infrastructures over a converged physical infrastructure

incorporating integrated optical network and IT resources. An MILP model

for virtualization of the underlying physical resources is proposed and

validated achieving significant energy savings.

©2011 Optical Society of America

OCIS codes: (060.4254) Networks, combinatorial network design; (060.4256) Networks,

network optimization.

References and links

1. M. Handley, “Why the Internet only just works,” BT Technol. J. 24(3), 119–129 (2006).

2. A. Tzanakaki, M. Anastasopoulos, K. Georgakilas, J. Buysse, M. De Leenheer, C. Develder, S. Peng, R.

Nejabati, E. Escalona, D. Simeonidou, N. Ciulli, G. Landi, M. Brogle, A. Manfredi, E. Lopez, J. F. Riera, J. A.

Garcia-Espin, P. Donadio, G. Parladori, and J. Jimenez, “Energy efficiency in integrated IT and optical network

infrastructures: the GEYSERS approach,” in IEEE INFOCOM 2011, Workshop Green Commun. and Netw.,

343–348.

3. M. Pickavet, W. Vereecken, S. Demeyer, P. Audenaert, B. Vermeulen, C. Develder, D. Colle, B. Dhoedt, and P.

Demeester, “Worldwide energy needs for ICT: The rise of power-aware networking,” in 2nd International

Symposium on Advanced Networks and Telecommunication Systems (ANTS '08) (IEEE, 2008), pp. 1–3.

4. A. Tzanakaki, K. Katrinis, T. Politi, A. Stavdas, M. Pickavet, P. Van Daele, D. Simeonidou, M. O'Mahony, S.

Aleksić, L. Wosinska, and P. Monti, “Dimensioning the future Pan-European optical network with energy

efficiency considerations,” J. Opt. Commun. Netw. 3(4), 272–280 (2011).

5. E. Kubilinskas, P. Nilsson, and M. Pioro, “Design Models for robust multi-layer next generation internet core

networks carrying elastic traffic,” in Proceedings of DRCN 2003, 61–68 (2003).

6. B. Quetier, V. Neri, and F. Cappello, “Selecting a virtualization system for grid/P2P large scale emulation,” in

Proc. EXPGRID’06 (2006).

7. Standard Performance Evaluation Corporation (SPEC) (www.spec.org).

8. X. Fan, W.-D. Weber, and L. A. Barroso, “Power provisioning for a warehouse-sized computer,” SIGARCH

Comput. Archit. News 35(2), 13–23 (2007).

9. M. P. Anastasopoulos, A. Panagopoulos, and P. Cottis, “A distributed routing protocol for QoS provisioning in

wireless mesh networks operating above 10 GHz,” Wirel. Commun. Mob. Comp. 8(10), 1233–1245 (2008).

10. P. Batchelor, B. Daino, P. Heinzmann, D. R. Hjelme, R. Inkret, H. A. Jäger, M. Joindot, A. Kuchar, E. L. Coquil,

P. Leuthold, G. D. Marchis, F. Matera, B. Mikac, H.-P. Nolting, J. Späth, F. Tillerot, B. V. Caenegem, N.

Wauters, and C. Weinert, “Study on the implementation of optical transparent transport networks in the

European environment-results of the research project COST 239,” Photonic Netw. Commun. 2(1), 15–32 (2000).

1. Introduction

As the scale of information processing is increasing, from Petabyes of Internet data to the

projected Exabytes in networked storage at the end of this decade [1], novel network solutions

are required to support the Future Internet and its new emerging applications such as UHD

IPTV, 3D gaming, virtual worlds etc. These high-performance applications, need to be

supported by specific IT resources (e.g. computing and data repositories) that maybe remote

and geographically distributed, requiring connectivity with the end users, through a very high

capacity and increased flexibility and dynamicity network. Considering that optical

#155756 - $15.00 USD Received 30 Sep 2011; accepted 1 Nov 2011; published 28 Nov 2011(C) 2011 OSA 12 December 2011 / Vol. 19, No. 26 / OPTICS EXPRESS B503

networking satisfies these requirements, through recent technology advancements including

dynamic control planes, elasticity etc., it provides a strong candidate to support this need. In

this context, an infrastructure comprising converged optical network and IT resources that are

jointly optimized in terms of infrastructure design and operation can be envisioned as the

suitable solution to support the Future Internet.

On the other hand, in order to maximize the utilization and efficiency of infrastructures,

supporting converged network and IT resources, the concept of 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. Through the

adoption of VI solutions, optical network and IT resources can be deployed and managed as

logical services, rather than physical resources. This results into enhanced enterprise agility,

remote access to geographically distributed infrastructures and maximization of network

utilization leading to reduced capital and operational costs.

An additional consideration that needs to be taken into account, in the context of Future

Internet sustainability, is the energy efficient design and operation of the associated

infrastructure, as ICT is responsible for about 4% of all primary energy today worldwide, and

this percentage is expected to double by 2020 [3]. Specifically in VIs, energy efficiency can

be effectively addressed at the VI planning phase [2]. VI planning is responsible to generate

dynamically reconfigurable virtual networks satisfying the VI provider’s-driven requirements

and meeting any specific needs such as e.g. energy efficiency. Through this process the least

energy consuming VIs that can support the required services are identified, in terms of both

topology and resources. In the optimization process involved, joined consideration of the

energy consumption of the converged network and IT resources is performed. As IT resources

require very high levels of power for their operation and their conventional operating window

is commonly not optimized for energy efficiency, allocating IT resources in an energy-aware

manner interconnected through a relatively low energy-consuming optical network can

potentially offer significant energy savings.

In this paper we propose for the first time a Mixed Integer Linear Programming (MILP)

model, suitable for the planning of multiple VIs formed over an integrated IT and optical

network infrastructure, extending our previous work on energy efficient single VI planning

[2]. To identify the least energy consuming VIs, the detailed power consumption models and

figures of the underlying physical infrastructure, including joint consideration of optical

network and IT resources are taken into consideration [ 4]. Mapping the virtual to physical

resources and defining the energy consumption parameters of the VIs themselves is also part

of the VI planning phase.

2. Energy aware multiple Virtual Infrastructure planning

The problem is formulated using a network that is composed of one resource layer that

contains the physical infrastructure (PI) is described through an eleven-node Pan-European

topology. A similar multi-layer network optimization approach may be found in [ 5]. The

objective is to produce as an output a layer that contains a set of VIs. For each VIi, i = 1,2,..., I ,

there is a set of demands i

d ( 1 ,2 ,...,i i i i

d D= ) to be served by a set of IT servers s (s =

1,2,…,S). VIs traffic demands are carried through the PI. For simplicity, the granularity of

demands is the wavelength and the IT locations (demand destinations) at which the services

will be handled, are not specified and are of no importance to the services themselves.

Therefore, the demand destinations for each VIi will be identified through the optimization

performed by the proposed model. In order to formulate this problem, the binary variable aid s

is introduced to indicate whether demand i

d that is handled by VIi is assigned to server s or


not and it equals 1, if and only if demand i

d is processed on server s. Moreover, it is assumed

that each demand can be assigned only to one server:

1, 1 ,2 , , , 1, 2,..., ,i id s i i is

da D i I= = =∑ … (1)

Furthermore, for each demand i

d , its volume idh is realized by means of a number of

lightpaths assigned to paths of the VIi. Let 1 ,2 ,...,

i i i id s d s d s d sp P= be the candidate path list in

the VIifor the lightpaths required to support demand

id at server s and

d Sip

x the non-negative

number of lightpaths allocated to path id sp . The following demand constraints should be

satisfied in the VI:

d ss

= 1, 2,..., 1, 2,...,, ,i d Sid si

ip i ip dx h d D i Ia = =∑ ∑ (2)

Summing up the lightpaths through each link i

e ( 1 ,2 ,....,i i i i

e E= ) of the VIi the

necessary link capacity ie

y for link i

e should satisfy the following constraint

e es d p

1 , 2 , ..., 1, 2, ...δ y , ,i

d Sii i ii i ii d Si

pd se E i Ix = =≤∑ ∑ ∑ (3)

eδ

i id s is a binary variable taking value equal to 1 if link e

i of VI

i belongs to path p

id s

realizing demand i

d at server s, 0 otherwise. Using the same rationale, the capacity of each

link ei in VI

i is allocated by identifying the required lightpaths in the PI. The resulting PI

lightpaths z determine the load of each link g (g = 1, 2,…, G) of the PI, and hence it capacity

ug. Assuming that 1 ,2 ,...,i i i i

q Q= is used for denoting the PI’s candidate path list realizing

linki

e , then, the following demand constraint for link i

e should be satisfied:

i ie q eq

1 , 2 , ..., 1, 2, ...z = y , ,i i i iii

e E i I= =∑ (4)

Note that the summation is taken over all paths i

q on the routing list Qi of link i

e . Finally,

in the PI the following capacity constraint should be satisfied

i ige q e q gi q

g = 1, 2,…, G,γ z u ,i ii ie

≤∑ ∑ ∑ (5)

where G is the total number of links in the PI and the summation for each link g is taken over

all lightpaths in the PI and ge qγ

i i the link-path incidence coefficients for the PI taking value

equal to 1 link g belongs to path i

q realizing ei.

Apart from link capacity constraints Eqs. (3) and (5), the total demands that are assigned

to each server should not exceed its capacity, s

p . This capacity corresponds to the underlying

physical resources, such as CPU, memory, disk storage etc. The inequality specifying servers’

capacity constraints is given by

( ) d s d sd p0s = 1, 2,…,S1 ( ) ,

d Sii i iiipid s si d

x pc a a c ≤+ ∑ ∑ ∑∑ ∑ (6)

The first term of Eq. (6) captures the additional processing requirements due to

virtualization, while the second the total demands that arrive at server s. Processing overhead

due to virtualization depends on the virtualization technology that is used. In this approach,

the User Mode Linux (UML) system has been adopted for the IT servers in which processing

overhead increases linearly with the number of VIs [ 6]. Parameter d s

c ( )d Sii p

x specifies the

computational requirements for demand i

d on server s and in practice is determined by the set


of relevant benchmarks for computer systems provided by the Standard Performance

Evaluation Corporation (SPEC) [ 7].

Fig. 1. Multiple VI architecture over a converged optical network and IT servers.

Table 1. Virtual to Physical Mapping

Virtual

Links

VI #1

S1-S4 (demands 15 wv/source)

VI #2

S1-S3 (demands 25 wv/source)

Capacity of Virtual

links (WV)

Physical Layer Paths

Realizing Virtual

Links

Capacity of

Virtual links

(WV)

Physical Layer Paths

Realizing Virtual

Links

Y1 0 - 0

Y2 15 u3-u11 25 u3-u11

Y3 15 u5-u10 55 u5-u10 (25wv)

u1-u4-u12 (5wv)

u8-u17-u13 (25wv) Y4 0 - 15 - Y5 0 - 0 - Y6 15 u22 25 u22 Y7 15 u16 25 u16

Y8 30 u12 40 u12 (15wv)

u14-u15 (25wv) Y9 0 - 15 u4

The objective of the current problem formulation is to minimize the total cost of the

resulting network configuration as this cost consists of the following components: (a) g

k that

is the cost of the capacity of link g of the PI. It consists of the energy consumed by each

lightpath due to transmission and reception of the optical signal, optical amplification at each

fiber span and switching and, (b) s

E that the energy consumption for processing us

wavelengths in the IT server s. The power consumption model adopted in this paper mainly

concentrates on the power consumption associated with the CPU load of IT resources and is

described via the following linear equation [ 8]:

+=( ) Idle busy

s ss s sE u P P u (7)

where idle

sP , busy

sP are parameters describing the energy consumption of the IT server s at idle

state and per wavelength, respectively. In addition to the power consumption due to data

processing, a 100% power overhead due to cooling has been incorporated to the energy

consumption model described above. In this context, minimum energy consuming VIs are

obtained by minimizing the following cost function


Minimize F = + ( )g g s sg sk u E u∑ ∑ (8)

The above MILP problem has been solved analytically employing the methods of

Lagrangian relaxation and dual decomposition [ 9].

Fig. 2. (a) Comparison of the Energy Aware with Closest IT scheme (3 VIs) (b) Impact of VIs

on power consumption.

Fig. 3. Utilization of optical network and IT Resources.

3. Numerical results

To investigate the energy efficiency of the proposed VI design scheme, the architecture

illustrated in Fig. 1 is considered: the lower layer depicts the PI and the layer above depicts

the VIs. For the PI, the COST239 [ 10] European topology has been used in which four

randomly selected nodes generate demands to be served by three IT servers. Furthermore, we

assume a single fiber per link, 40 wavelengths per fiber, and wavelength channels of 10Gb/s

each. It is also assumed that each IT server can process up to 2Tb/s and its power

consumption ranges from 6.6 to 13.2KW, under idle and full load, respectively. Finally, the

virtualization processing overhead is 3% per VI.

An example of the optimal VI topology design with two VIs is depicted in Fig. 1. In this

scenario, VI #1 includes four source nodes that are located in London, Vienna, Copenhagen

and Paris generating demands equal to 15 wavelengths each, while VI #2 incorporates three

source nodes that are located in London, Vienna and Copenhagen generating demands equal

to 25 wavelengths. The generated VI #1 topology consists of 5 virtual links and 6 virtual

nodes, while all demands are routed to the IT server in Luxemburg. The capacity of each

virtual link along with its mapping to the PI is given in Table 1 where e.g. it is observed that


virtual link Y3 connecting Copenhagen and Luxemburg is realized via the physical layer path

u5-u10, with capacity 25 wavelengths.

Figure 2a, illustrates the total power consumption of the infrastructure (optical network

and IT resources) when applying the proposed MILP approach optimizing for energy or

distance between sources and IT servers (closest It scheme). Comparing these two schemes, it

is observed that the energy aware VI design consumes significantly lower energy to serve the

same amount of demands compared to the closest IT scheme providing an overall saving of

the order of 40%. This is due to that, 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 infrastructures, switching-off the unused

IT resources achieves significant reduction of energy consumption. Furthermore, it is

observed that in both schemes the average power consumption increases almost linearly with

the number of demands. However, the relative benefit of the energy aware design decreases

slightly with the number of demands, as we get closer to full system load. Figure 2b depicts

the impact of the number of VIs on power consumption. It is observed that the virtualization

cost has a more severe impact in terms of power consumption when applying the closest IT

scheme than that observed in the case of energy aware VI planning, as in the case of the

closest IT scheme more virtual machines per IT server are activated.

In addition, it is observed that there is a clear trade-off between the utilization of optical

network resources and the number of active IT servers. Specifically, as the energy cost for

activating an IT server predominately affects the overall energy consumption of the

converged infrastructure, the energy aware VI planning scheme forwards traffic to the

minimum possible number of IT servers. However, following this approach more optical

network resources are employed as data travel larger distances to arrive to their destination

(IT server). On the other hand when the VIs are planned using the Closest IT scheme, all

demands are routed to their closest IT servers, thus minimizing the utilization of the optical

network resources at the expense of the total number of active IT servers required. This trade-

off can is illustrated in Fig. 3a where for example, in the closest IT scheme it is observed that

less than 30% of the total optical network resources are employed to transfer demands from

the source nodes to three active IT servers. In contrast, as depicted in Fig. 3b the energy aware

scheme (supporting the same demands) routes all demands to only a single IT server at the

expense of increased utilization of the optical network resources.

4. Conclusions

The paper studied the problem of energy efficient service provisioning in converged optical

network and IT infrastructures. An MILP model for virtualization of the underlying physical

resources was proposed and validated achieving significantly lower energy consumption for

serving the same amount of demands (with an overall saving of the order of 40%).

Furthermore, switching-off unused IT resources achieves significant reduction of energy

consumption. It was also proven that the virtualization cost has a more severe impact in terms

of power consumption when applying the closest IT scheme compared to the proposed model

since more virtual machines per IT server are activated. Finally, numerical results indicate

that there is trade-off between the utilization of optical network resources and the number of

active IT servers determining the level of the overall power consumption.

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.


Energy aware planning of multiple virtual infrastructures over converged optical network and IT...

Documents

Transcript of Energy aware planning of multiple virtual infrastructures over converged optical network and IT...