Elastic-Net: Boosting Energy Efficiency and Resource Utilization in 5G C-RANs
aa r X i v : . [ c s . N I] O c t Elastic-Net: Boosting Energy Efficiency andResource Utilization in 5G C-RANs
Abolfazl Hajisami, Tuyen X. Tran, and Dario Pompili
Department of Electrical and Computer EngineeringRutgers University–New Brunswick, NJ, USAe-mail: { hajisamik, tuyen.tran, pompili } @cac.rutgers.edu Abstract
Current Distributed Radio Access Networks (D-RANs), which are characterized by a static con-figuration and deployment of Base Stations (BSs), have exposed their limitations in handling thetemporal and geographical fluctuations of capacity demands. At the same time, each BS’s spectrumand computing resources are only used by the active users in the cell range, causing idle BSs in someareas/times and overloaded BSs in other areas/times. Recently, Cloud Radio Access Network (C-RAN)has been introduced as a new centralized paradigm for wireless cellular networks in which—throughvirtualization—the BSs are physically decoupled into Virtual Base Stations (VBSs) and Remote RadioHeads (RRHs). In this paper, a novel elastic framework aimed at fully exploiting the potential of C-RAN is proposed, which is able to adapt to the fluctuation in capacity demand while at the same timemaximizing the energy efficiency and resource utilization. Simulation and testbed experiment results arepresented to illustrate the performance gains of the proposed elastic solution against the current staticdeployment.
Index Terms
Cloud Radio Access Network; Virtual Base Station; Energy Efficiency; Resource Utilization.
I. I
NTRODUCTION
Motivation:
Over the last few years, proliferation of personal mobile computing devices liketablets and smartphones along with a plethora of data-intensive mobile applications has resultedin a tremendous increase in demand for ubiquitous and high data rate wireless communications.The current practice to enhance the spectral efficiency and data rate is to increase the number of Base Stations (BSs) and go for smaller cells so as to increase the band reuse factor.
However,it is studied that increasing the BS density or the number of transmit antennas will decreasethe energy efficiency due to the dynamic traffic variation [1]. The number of active users atdifferent locations varies depending on the time of the day and week. This movement of mobilenetwork load based on time of the day/week is referred to as the “tidal effect”. In the traditionalDistributed Radio Access Network (D-RAN), each BS’s spectrum and computing resources areonly used by the active users in the cell range . Hence, deploying small cells for the peak (worstcase) traffic time leads to grossly under-utilized BSs in some areas/times and is highly energyinefficient, while deploying cell for the average traffic time leads to oversubscribed BSs in someother areas/times.At the stage of network planning, cell size and capacity are usually determined based onthe estimation of peak traffic load. However, due to the tidal effect, there are no fixed cellsize and transmission power that optimize the overall power consumption of the network . Thismeans that the use of small cells is quite efficient in terms of power consumption as well asutilization of spectrum and computing resources when the capacity demand is high and evenlydistributed in space; however, it becomes less so when the data traffic is low and/or uneven dueto the static resource provisioning and fixed BS power consumptions. On the other hand, theeconomic impact of power consumption is particularly dire in emerging markets and the FifthGeneration (5G) of wireless networks must be not only spectral efficient but also energy efficient(e.g., a 1000 × improvement in energy efficiency is expected by 2020). Although several recentefforts have been made to increase spectral efficiency [2] and to reduce the power consumptionof existing small cell networks, limited attention has been given towards optimizing the overallnetwork deployment. Hence, a novel design and architecture is necessary for the next generationof wireless network to overcome these challenges. A New Centralized Cellular Network Paradigm:
Cloud Radio Access Network (C-RAN) [3]is a new architecture for the next generation of wireless cellular networks that allows for adynamic reconfiguration of spectrum/computing resources (Fig. 1). C-RAN consists of threeparts: 1) Remote Radio Heads (RRHs) plus antennae, which are located at the remote siteand are controlled by Virtual Base Stations (VBSs) housed in a centralized processing pool,2) the Base Band Unit (BBU) (known as VBS pool) composed of high-speed programmableprocessors and real-time virtualization technology to carry out the digital processing tasks, and3) low-latency high-bandwidth optical fibers, which connect the RRHs to the VBS pool. The communication functionalities of the VBSs are implemented on Virtual Machines (VMs), whichare housed in one or more racks of a small cloud datacenter. This centralized characteristic alongwith virtualization technology and low-cost relay-like RRHs provides a higher degree of freedomto make optimized decisions, and has made C-RAN a promising candidate to be incorporatedinto the 5G networks [4]–[7].
Our Contributions:
In this paper, we focus on optimizing the power consumption andresource utilization by leveraging the full potential of the C-RAN architecture. We proposea novel elastic resource provisioning framework, called “Elastic-Net”, to minimize the powerconsumption while addressing the fluctuations in per-user capacity demand. In our solution, wedivide the covered region of the network into clusters based on the traffic model and, within eachcluster, we dynamically adapt the active RRH density, transmission power, and size of the VM based on the traffic fluctuations so as to minimize the power consumption while maximizingthe resource utilization. To minimize the power consumption in the cell sites while ensuringa certain minimum coverage and data rate, we propose to dynamically optimize and adapt theRRH density and transmission power based on the traffic demand and user density. Likewise, tominimize the power consumption in the cloud we dynamically optimize and adapt the size ofthe VMs while ensuring that the frame-processing deadline is met. Paper Outline:
In Sect. II, we present the system model. In Sect. III, we formulate the problemand describe our demand-aware provisioning framework. In Sect. IV, we validate our statementsthrough simulation and testbed experiments. Finally, we conclude the paper in Sect. V.II. S
YSTEM M ODEL
We consider a C-RAN downlink system and assume that each user is served by the nearestactive RRH. The RRHs and users are distributed according to two independent Poisson PointProcesses (PPPs) in R , denoted as Φ r and Φ u ( t ) , respectively. Let λ r and λ u ( t ) denote theRRH density and the time-dependent user density, respectively. The set of all RRHs is denotedby L = { , . . . , L } , where A ⊆ L is the set of active RRHs and
Z ⊆ L is the set of inactiveRRHs (
A ∪ Z = L ). Let also µ a ( t ) (0 ≤ µ a ( t ) ≤ denote the RRH activity factor, whichindicates the ratio of active RRHs to all RRHs, where λ ar ( t ) = µ a ( t ) λ r is the time-dependentdensity of active RRHs. The total bandwidth is denoted by B and the bandwidth per user isgiven by B u ( t ) = B λ ar ( t ) λ u ( t ) . The size of a VM is represented in terms of its processing power, memory and storage capacity, and network interface speed. (a) Day (b) Night
Fig. 1. The use of virtualization in C-RAN allows dynamic re-provisioning of spectrum and computing resources (visualizedhere using different sizes) to Virtual Base Stations (VBSs) based on demand fluctuation; (a) and (b) illustrate the movement ofmobile network load from the downtown office area to the residential and recreational areas over the course of a day and thecorresponding changes in active RRH density and VBS size (we have used different icons for active RRH and inactive RRH).
RRH and Transport Network Power Consumption Model:
Since in C-RAN the BSs aredecoupled into RRHs and VBSs, we divide the network power consumption into two parts:(i) RRH and transport network power consumption and (ii) VBS pool power consumption. Forthe power consumption of a RRH, we consider a linear model as in [8], P rrh = P arrh + η P if P > P srrh if P = 0 , (1)where P arrh is the active circuit power consumption, η is the power amplifier efficiency, P isthe transmission power, and P srrh is the RRH power consumption in the sleep mode. We alsoconsider the future Passive Optical Network (PON) to provide low-cost, high-bandwidth, low-latency connections between the RRHs and VBS pool [9]. PON comprises an Optical LineTerminal (OLT) that resides in the VBS pool and connects a set of associated Optical NetworkUnits (ONUs) through a single fiber. In this paper, we consider fast/cyclic sleep mode wherethe ONU state alternates between the active state (when the RRH is in the active state) and thesleep state (when the RRH is in the sleep state). Hence, the power consumption of the transportnetwork is given as in [9], P tn = P olt + P onu , (2) where P olt is the OLT power consumption in the VBS pool and P onu is the ONU powerconsumption, given as, P onu = |A| P atl + |Z| P stl , (3)where P atl and P stl are the consumed power by each ONU in the active and sleep mode, respec-tively. Since P olt is consumed in the VBS pool, we consider it in the power consumption of theVBS pool. Therefor, the overall area RRH and transport network power consumption is givenby, P area = λ ar ( t )( P arrh + 1 η P + P atl ) + λ sr ( t )( P srrh + P stl ) . (4) VBS Pool Power Consumption Model:
As discussed in [10], compared to other systemresources, the CPU consumes the main part of the power in the VBS pool; hence, in this work,we focus on minimizing its power consumption. In order to model the power consumption ofthe VBS pool, we introduce the notion of size of a VM , which is represented in terms of itsprocessing power [CPU cycles per second]; for each VM, we consider the power model definedas, P vm = ∆( t ) P max u ( t ) + β ∆( t ) P max (1 − u ( t )) + P olt , (5)where ∆( t ) is the size of the VM in terms of CPU cycles per second, P max is the maximumpower consumed per unit VM size when the server is fully utilized, β is the fraction of powerconsumed by the idle VM, and u ( t ) is VM utilization. Note that, in our model, ∆( t ) and u ( t ) change over time due to the workload variation and hence they are functions of time.III. E LASTIC -N ET : D EMAND -A WARE P ROVISIONING
In our solution, as depicted in Fig. 1, we cluster the neighboring RRHs and their correspondingVBSs based on traffic model, and in each cluster we adapt the system parameters accordingly.We advocate demand-aware resource provisioning where in each cluster the active RRH density,transmission power, and size of the VM are dynamically changed over time to minimize thepower consumption and to meet the fluctuating traffic demand as well as network constraints.For instance, as shown in Fig. 1(a)-(b), due to the higher capacity demand during day time incluster power consumption is minimized while meeting a predefined coverage probability, per-user datarate, and frame-processing time. Our optimization problem for the i th cluster can be cast as, p : argmin µ a ,P, ∆ P iarea ( µ a ( t, i ) , P ( t, i )) + P ivm (∆( t, i )) (6a)subject to P cov ≥ εP ∞ cov , (6b) R u ≥ R , (6c) T dl ≥ T fr , (6d)where P iarea ( µ a ( t, i ) , P ( t, i )) and P ivm (∆( t, i )) are the area power consumption and VBS-Clusterpower consumption of the i th cluster, respectively, P ∞ cov is the coverage probability at no noiseregime, R is the per-user minimum data rate, T fr is the frame-processing time, T dl is framedeadline, and ε is a positive number in [0 , . µ a ( t, i ) , P ( t, i ) , and ∆( t, i ) are also the RRHactivity factor, transmission power, and size of the VM for the i th cluster, respectively. Due tothe temporal variation of traffic demand in each cluster, µ a ( t, i ) , P ( t, i ) , and ∆( t, i ) are timedependent; hence, the optimal solution [ µ ∗ a ( t, i ) , P ∗ ( t, i ) , ∆ ∗ ( t, i )] in general varies over time.The density of active and inactive RRHs in the i th are, λ ar ( t, i ) = µ a ( t, i ) λ r ( i ) , (7a) λ sr ( t, i ) = (1 − µ a ( t, i )) λ r ( i ) , (7b)where λ r ( i ) is the density of all RRHs in the i th cluster. By substituting (7a) and (7b) into (4),we can write, P i ( P ( t, i ) , µ a ( t, i )) = λ r ( i ) ( µ a ( t, i ) Q ( t, i ) + Q ) , (8)where Q ( t, i ) = P arrh + 1 η P ( t, i ) + P atl − P srrh − P stl , (9a) Q = P srrh + P stl . (9b)From (8) and (9a), we see that the objective function is non-convex because of the multipli-cation term of µ a ( t, i ) and P ( t, i ) . To minimize the objective function, as in [11], we can usethe coordinate descent algorithm and minimize µ a ( t, i ) , P ( t, i ) , and ∆( t, i ) independently. A. Optimal Active RRH Density
Lemma 1.
The minimum RRH activity factor for which the constraint R u ≥ R is met is givenby, µ ∗ a ( t, i ) = R λ u ( t, i ) Bλ r ( i ) h log (1 + γ ) + γ α A ( α, γ ) i , (10) where A ( α, γ ) = Z ∞ γ x − /α x dx. (11) Proof.
The spectral efficiency achievable by a randomly chosen user when it is in coverage isgiven as in [12], τ ( α, γ ) = log (1 + γ ) + γ α A ( α, γ ) . (12)Hence, the per-user data rate in the i th cluster and at time instant t can be written as, R u ( t, i ) = Bµ a ( t, i ) λ r ( i ) λ u ( t, i ) h log (1 + γ ) + γ α A ( α, γ ) i . (13)So, considering constraint (6c), we can write, µ a ( t, i ) ≥ R λ u ( t, i ) Bλ r ( i ) h log (1 + γ ) + γ α A ( α, γ ) i , (14)which establishes the minimum RRH activity factor as a function of λ u ( t, i ) to satisfy the per-userdata-rate constraint. B. Optimal Transmission Power
Given a fixed active RRH density, we can minimize the transmit power of the active RRHsso as to achieve a certain coverage and outage probability. Since in our solution the activeRRH density of different clusters changes over time based on the traffic demand, we need toalso dynamically optimize the transmit power accordingly. This can further decrease the powerconsumption of the system. For instance, when the density of active RRHs becomes higher, eachRRH has only a small coverage area and users can be in coverage even with a lower transmissionpower.
Lemma 2.
The minimum transmission power for which the constraint P cov ≥ εP ∞ cov is met isgiven by, P ∗ c ( t, i ) = L [ µ a ( t, i ) λ r ( i )] α , (15) where L = γσ Γ (cid:0) α + 1 (cid:1) π α [1 + Υ( γ, α )] α (1 − ε ) , (16a) Υ( γ, α ) = γ α Z ∞
11 + z α dz. (16b) Proof.
The coverage probability in the i th cluster is given as [13], P cov ( α, γ, µ a ) = πµ a ( t, i ) λ r ( i ) × Z ∞ e − πµ a ( t,i ) λ r ( i )(1+Υ( α,γ )) − γσ v α/ P − dv. (17)Now, by using the substitution γσ P − → s in (17) and the approximation e − sv α/ ≈ (cid:0) − sv α/ (cid:1) (in the case of low-noise regimes, i.e., σ n → ), we can write, P cov ( α, γ, µ a ) ≈ P ∞ − γσ Γ (cid:0) α + 1 (cid:1) P [ πµ a ( t, i ) λ r ( i ) (1 + Υ ( α, γ ))] α ! , (18)where P ∞ is the coverage probability without noise [13], i.e., P ∞ = (1 + Υ ( α, γ )) − , (19)and Γ ( x ) = R ∞ t x − e − t dt is the standard gamma function. The minimum transmission power P ∗ c ( t, i ) that satisfies the coverage constraints is obtained by combining (18) and (6b). C. Optimal Size of VM
We recast the power consumption of the i th VBS-Cluster, P ivm = ∆ ( t, i ) P max u ( t, i ) (1 − β ) + β ∆ ( t, i ) P max + P olt , (20)where, for a given workload, u ( t, i ) is inversely proportional to ∆( t, i ) [14]. So, to minimizethe power consumption of the VM, we need to minimize the size of the VM (CPU cores) suchthat the network requirements are met. The workload in the VBS-Cluster depends on the LTEMCS index, number of PRBs, and the channel bandwidth. Moreover, according to [15], theRound Trip Time (RTT) between RRH and VBS pool cannot exceed µ s . Since the totaldelay budget in LTE is considered as , this leaves the VBS-Cluster with only about . for signal processing. For this matter, we consider a modified model of the processing timepresented in [16], which is given by, T fr = M υ ∆( t, i ) = M υN c ( t, i ) ω , (21)where T fr is the processing time and is measured in µ s , M is the number of PRBs, υ is aMSC-dependent constant, N c ( t, i ) is the number of dedicated CPU cores to the i th VBS-Cluster,and ω is the CPU speed measured in GHz . Hence, the minimum number of required CPU coresto meet the frame deadline is given by, N ∗ c ( t, i ) = (cid:24) M υT dl ω (cid:25) , (22)where T dl is the frame deadline. TABLE IS
IMULATION P ARAMETERS . Parameters Mode/Value
Cellular Layout Homogeneous Poisson Point ProcessChannel Model Path Loss and ShadowingChannel Bandwidth
20 MHz
Number of Antennas ( N TX , N RX ) (1 , OLT power consumption ( P olt )
20 W
ONU Power Consumption in Active Mode ( P atl ) ONU Power Consumption in Sleep Mode ( P stl ) . RRH Circuit Power Consumption in Active Mode ( P arrh ) . RRH power consumption in sleep mode ( P srrh ) . Maximum power consumed per each CPU core ( P max )
72 W
Power Amplifier Efficiency ( η ) . MSC Dependent Constant ( α ) . Fraction of Power Consumed by Idle VBS ( β ) . Fraction of Minimum Coverage Probability ( ε ) . Minimum Data Rate ( R )
200 Kbps
IV. P
ERFORMANCE E VALUATION
In this section, we provide a range of simulations and real-time emulations to evaluate theperformance of our solution. In the simulations, we consider a cellular network where the RRHsand the users are distributed according to two independent homogeneous PPPs. Table I lists thestimulation parameters used during our experiments. In order to show the performance of ourdynamic solution, we simulate the traffic fluctuation on a typical operational day and show how
Elastic-Net dynamically adapts the RRH density, transmission power, and size of VBS-Clustersto minimize the power consumption while at the same time meeting the network constraints. As Hours of Day U s e r D en s i t y ( m - ) Cluster
Hours of Day A c t i v e RRH D en s i t y ( m - ) × -4 Cluster
Hours of Day M i n i m u m T r an s m i ss i on P o w e r ( d B m ) Cluster
Hours of Day P o w e r C on s u m p t i on o f VBS - C l u s t e r ( W ) × Cluster (a) (b) (c) (d)
Fig. 2. (a) Traffic fluctuation on a typical operational day for three different clusters; (b) Fluctuation of active RRH density tomeet the user fluctuation for different times of the day; (c) Fluctuation of minimum transmission power to reduce the powerconsumption while guaranteeing a predefined Quality of Service (QoS); (d) Power consumption of different VBS-Clusters (CPUspeed = . ). shown in Fig. 1, we consider a network with clusters where each cluster covers an area of km and has its own traffic and user density fluctuation.As shown in Fig. 2(a), clusters . decrease in power consumption by Elastic-Net. However,for the peak traffic times (8 AM–7 PM), we have only . decrease in power consumption.This confirms our statements in Sect. I that a small-cell deployment is efficient when the capacitydemand or user density is high, while it becomes less so when the traffic demand is low. Hours of Day T o t a l P o w e r C on s u m p t i on ( W ) × Elastic-NetStatic-Net
Hours of Day T o t a l P o w e r C on s u m p t i on ( W ) × Elastic-NetStatic-Net
Hours of Day T o t a l P o w e r C on s u m p t i on ( W ) × Elastic-NetStatic-Net (a) (b) (c)
Fig. 3. Comparison of power consumption between Elastic-Net (in C-RAN) and Static-Net (in D-RAN) for (a) Cluster
V. C
ONCLUSION
In the context of C-RAN—a new centralized paradigm for wireless cellular networks inwhich the Base Stations (BSs) are physically decoupled into Virtual Base Stations (VBSs) andRemote Radio Heads (RRHs)—we proposed a novel demand-aware reconfigurable solution,named
Elastic-Net , to minimize the network power consumption and to adapt to the fluctuationsin per-user capacity demand. We divided the covered region of the cellular network into clustersbased on active traffic, and dynamically adjusted RRH density, transmission power, and size ofthe Virtual Machine (VM) holding the VBS so that the network power consumption is minimizedand the network constraints are met. Simulation results corroborated our analysis and confirmedthe benefits of our solution.
Acknowledgment:
This work was supported in part by the National Science Foundation GrantNo. CNS-1319945. R
EFERENCES [1] C. Li, J. Zhang, and K. Letaief, “Throughput and energy efficiency analysis of small cell networks with multi-antennabase stations,”
IEEE Trans. Wireless Commun. , vol. 13, no. 5, pp. 2505–2517, 2013.[2] M. Hajimirsadeghi, G. Sridharan, W. Saad, and N. B. Mandayam, “Inter-network dynamic spectrum allocation via a colonelblotto game,” in
Proc. IEEE CISS , 2016, pp. 252–257.[3] C. M. R. Institute, “C-RAN: The Road Towards Green RAN,” in
C-RAN International Workshop , Oct. 2011.[4] T. X. Tran, A. Hajisami, and D. Pompili, “QuaRo: A queue-aware robust coordinated transmission strategy for downlinkC-RANs,” in
Proc. IEEE SECON , 2016, pp. 1–9.[5] A. Hajisami and D. Pompili, “Dynamic joint processing: Achieving high spectral efficiency in uplink 5g cellular networks,”
Computer Networks , vol. 126, pp. 44–56, 2017. [6] A. Hajisami, H. Viswanathan, and D. Pompili, “Cocktail Party in the Cloud: Blind Source Separation for Co-operativeCellular Communication in Cloud RAN,” IEEE International Conference on Mobile Ad hoc and Sensor Systems (MASS) ,pp. 37–45, 2014.[7] M. Hajimirsadeghi, N. B. Mandayam, and A. Reznik, “Joint caching and pricing strategies for popular content in informationcentric networks,”
IEEE Journal on Selected Areas in Communications , vol. 35, no. 3, pp. 654–667, 2017.[8] G. Auer, V. Giannini, C. Desset et al. , “How much energy is needed to run a wireless network?”
IEEE Wireless Commun. ,vol. 18, no. 5, pp. 40–49, 2011.[9] A. R. Dhaini, P.-H. Ho, G. Shen, and B. Shihada, “Energy efficiency in tdma-based next-generation passive optical accessnetworks,”
IEEE/ACM Trans. Netw. (TON) , vol. 22, no. 3, pp. 850–863, 2014.[10] A. Beloglazov, J. Abawajy, and R. Buyya, “Energy-aware resource allocation heuristics for efficient management of datacenters for cloud computing,”
Future Generation Computer Systems , vol. 28, no. 5, pp. 755–768, 2012.[11] K. Lin, J. Xu, I. M. Baytas, S. Ji, and J. Zhou, “Multi-task feature interaction learning,” in
Proc. ACM Int. Conf. KnowledgeDiscovery and Data Mining , 2016, pp. 1735–1744.[12] H. S. Dhillon, R. K. Ganti, F. Baccelli, and J. G. Andrews, “Modeling and analysis of k-tier downlink heterogeneouscellular networks,”
IEEE J. Sel. Areas Commun. (JSAC) , vol. 30, no. 3, pp. 550–560, 2012.[13] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to coverage and rate in cellular networks,”
IEEE Trans.Commun. , vol. 59, no. 11, pp. 3122–3134, 2011.[14] K. S. Trivedi, R. A. Wagner, and T. M. Sigmon, “Optimal selection of cpu speed, device capacities, and file assignments,”
Journal of the ACM (JACM) , vol. 27, no. 3, pp. 457–473, Jul. 1980.[15] P. Chanclou, A. Pizzinat, F. Le Clech, T. Reedeker, Y. Lagadec, F. Saliou, B. Le Guyader, L. Guillo, Q. Deniel, andS. Gosselin, “Optical fiber solution for mobile fronthaul to achieve cloud radio access network,” in
Proc. IEEE FutureNetwork and Mobile Summit , 2013, pp. 1–11.[16] I. Alyafawi, E. Schiller, T. Braun, D. Dimitrova, A. Gomes, and N. Nikaein, “Critical issues of centralized and cloudifiedlte-fdd radio access networks,”