On Femto placement and decoupled access for downlink and uplink in enterprise environments

While it is easier to quench the demand for higher data rates outdoors, it is still a significant challenge when it comes to attaining similar data rates for indoor User Equipments (UEs). Femto cells were introduced for this purpose and also to minimize the traffic load on macro Base Stations (BSs) in 4G/LTE cellular networks. Indoor UEs can achieve good throughput if they get high Signal to Noise Ratio (SNR), but the inherent problem of path loss due to obstacles prevents UEs from receiving good signals. So, the efficient placement of Femtos in enterprise buildings is crucial. For the optimal placement of Femtos, we developed a Mixed Integer Linear Programming (MILP) model and solved it using the GAMS tool. Once the network planning is done, the next problem that has to be addressed is the downlink traffic imbalance which happens due to non-uniform UE traffic distribution. Traditionally load imbalance is addressed by transferring some of the UEs from the highly loaded cell to a less loaded neighboring cell but this could increase the UE uplink transmission power as it now connected to a cell which is not the closest one. To improve UE battery life and to boost the downlink throughput, we decouple the uplink and downlink (DuD) access to UEs by connecting the uplink to the shortest pathloss Femto, and the downlink to one of less loaded neighboring Femtos. Our extensive experimentation in MATLAB shows that on average, the decoupled access system achieves 70% energy savings (i.e., uplink power) when compared to coupled access system. Received on ; accepted on ; published on


Introduction
The generation of smartphones and tablets has seen a tremendous increase in the demand for higher data rates.While 3G data provided by mobile operators serves basic purposes, Long Term Evolution (LTE) is tipped to provide the best data rates for meeting ever increasing demands in outdoor environments.According to a study made by Cisco and Huawei [1], the need to address high traffic in indoor environments is high.Poor coverage and path loss due to obstacles like walls and floors inhibit the existing outdoor Base Stations (BS) providing good indoor coverage in multistorey buildings.Better data rates could be achieved in indoor for User Equipments (UEs) by deploying more macro BSs, but this would increase Capital Expenditure (CAPEX) and Operational Expenditure (OPEX) incurred by the mobile operator.In order to satisfy the demand of indoor UEs in low CAPEX, deploying Low Power Nodes (LPNs) a.k.a.Femto BSs [2], is an amenable solution.Femto (a.k.a.Home eNodeB and enterprise eNodeB) is a low power plugand-play BS designed for covering small regions like homes and offices.Bulk deployment of Femtos in enterprise environments could substantially enhance user experience of cellular systems.Femtos have the capability to serve around 20 to 30 users within the coverage range of at most 60 meters and Femtos get connected to a centralized Femto-GW (refer Fig. 1) using S1 interface for interference and handover management.
The dense deployment of Femtos in enterprise environments [3] necessitates the need for their optimal placement to guarantee good signal strength to all indoor UEs and to minimize coverage holes.Once, the placement of Femtos is done, the other varying parameters that affect the data rate in wireless systems Vanlin Sathya are load and interference.In our work, we concentrate on the former by formulating a Mixed Integer Linear Programming (MILP) model for the optimal placement of Femtos.In a typical indoor scenario with UEs, the uplink load of a cell would more or less be the same in the entire building, but the downlink load of the cell would vary from one Femto to other depending on the number of UEs being served [4,5], and their traffic demands.In traditional cellular networks, the uplink access and downlink access are coupled to the same BS as shown in Fig. 1.The user U t uses Femto 2 for both uplink and downlink communication because the signal strength from Femto 2 is very high for the user U t .Suppose a Femto is fully loaded when compared to its neighboring Femtos, the traditional offloading or load balancing algorithms [6,7] will shift some of the UEs for both uplink and downlink from the loaded cell to one of less loaded cells (target cells) provided that these UEs still could get connected to the target cell.This type of offloading is a forced handover based on the load but not based on signal strength, because the signal strength when connected to the initial serving Femto will be high compared to the target Femto.For example in Fig. 1, we observe that Femto 2 is heavily loaded so the user u t can be offloaded to the neighboring Femto i.e., Femto 1 .In a coupled access system, as shown in Fig. 2, after offloading u t uses Femto 1 for both uplink and downlink communication.Since u t and Femto 1 are separated by a wall, u t has to transmit with a higher power in Femto 1 compared to Femto 2 .By doing this offloading approach, the overall system throughput will boost but the uplink power of the shifted UEs would increase and there by drain their batteries.In order to reduce the battery drain from UEs and to boost the downlink data rate, one could use the Decoupled uplink and downlink (DuD) access method i.e., uplink connected to the closest Femto and downlink to a less loaded Femto.In this work,  the Femto which is closest (highest SNR) to an UE is called as serving Femto and the neighboring/target Femtos are called as target Femtos.For example in Fig. 3, we observe that user u t uses Femto 1 for downlink and Femto 2 for uplink communication which saves its UE battery.Before doing this, the placement of Femtos should be optimal inside the building to attain a desirable SNR (Signal-to-Noise Ratio) for indoor UEs.
In the present work, we placed Femtos optimally to obtain a desirable SNR.Two major parameters that determine the optimal Femto locations include (i) distance between Femto and the farthest point inside the building and (ii) the minimum SNR needed by each UE.Solving for optimal placement of Femtos by considering the above parameters results in a nonconvex optimization problem.We further simplify this non-convex problem to fit into MILP model and solve it using GAMS tool [8].To increase the downlink throughput and reduce battery drain, we propose an offloading algorithm in DuD access system.
The rest of the paper is organized as follows.Section 2 presents related work on Femto placement and decouped access systems.Section 3 describes the proposed work in Femto placement.Section 4 presents the proposed efficient offloading algorithm for decouped access systems.In Section 5, we show the performance results of the proposed Femto placement model and offloading algorithm in a two-storey enterprise building scenario.Finally, Section 6 presents conclusions and directions for future work.

Related Work
In order to boost the data rate, Femtos were introduced by 3GPP.Right from its incorporation, extensive research has been carried out to improve Femto placement architectures considering various issues like frequent hand-offs, interference and physical cell ID (PCID) [3].An effective algorithm for the optimal placement of Femtos depends on the distance between the first Femto and the macro BS, as given in [9,10], but the work does not consider the Femto to Femto interference inside the building.Works on optimal relay node placement in tunnels [11], sensor placement [12,13] in terrain regions and optimal placement of Wi-Fi APs [14] exist in literature.In [15], we studied placement of the Femtos optimally inside a building while securing good SINR for UEs by considering crosstier and co-tier interference between Macro BSs and the Femtos.In [16], we studied joint Femto placement and power control for guaranteeing SINR threshold depending on user occupancy in each region inside the building.However, MILP models proposed in our previous works [15,16] target to ensure SINR threshold only at the center of the sub-regions (i.e., sub-region is of length δ x and width δ y , explained in detail latter) inside the building and hence they fail to guarantee at all the farthest points of the sub-regions.In this work, we have not considered inter-Femto interference (i.e., co-tier interference) and Macro-BS interference (i.e., cross-tier interference) in the system model.But our optimization model guarantees good SNR at the farthest points in the sub-regions inside the building.The algorithm for optimal Femto placement provided in [17,18] is of limited practical value since it ignores walls inside the buildings while determining Femto locations.Appending to the above work, we constrained the problem to be similar to realistic enterprise buildings by considering the path loss across walls inside the building.The resulting non-convex optimization model is solved by approximating it as an MILP.
After placing the Femtos optimally to guarantee good signal strength at the farthest points inside the sub-region of the building, the other problem that arises is load-imbalance across Femtos.Lot of existing literature discuss about the load balancing in LTE systems by varying the handover hysteresis margin, dynamic BS power control [15] and centralized load balance using software defined network (SDN) approaches.In [6,7,19] authors proposed a multiobjective problem with the objective of load balance by meeting QoS requirements and the network utility of other services like voice, video and online gaming, but the running time complexity of the proposed optimization problem is high.So, they proposed a practical algorithm which considers QoS guaranteed hybrid scheduling, handover of users with and without QoS requirements and call admission control.In [20], we designed a self organizing network (SON) triggered load balance/handover algorithm which takes into account the building layout to avoid ping-pong effect.
All these traditional load balancing techniques increase the battery power consumption of the UE who is being offloaded from heavily loaded Femto to lightly loaded Femto since the distance for the uplink access is high in coupled access systems.In order to save energy, decoupling is the efficient solution without degrading the performance of the UE.Along these lines, in the first work [21,22], authors proposed a way in which the uplink connects to one of less path loss small cells and the downlink connects to a macro BS, causing a reduction in the uplink power, but also decreases the downlink throughput due to path losses encountered in reaching the UE indoors.
In [23], we focused on the optimal placement of Femtos inside the building using a MILP model.In this work, we extend the work done in [23] by proposing an efficient offloading algorithm in DuD access system for addressing load imbalance in Femto cells.

Optimal Femto Placement
In this section, we present Femto placement in enterprise building environments as an optimization problem.width of each room are δ rx and δ ry , respectively.Each room is numbered ρ zxy , also be denoted as ρ z ρ x ρ y .The first digit in three digit numerical scheme signifies the floor number, second digit varies along X-axis and the third digit varies along Y -axis as shown in Fig. 5.If the room number is referred to as ρ x , it implies that the room number is varied along the X-axis only.For e.g. if ρ zxy = 122 and if ρ x + 1 operation is applied, then ρ zxy = 132.We assume that F Femtos are available to cover the entire building and they are to be placed only on the ceiling of the rooms.ρ f denotes the room number of f th Femto.We further divide each room into sub-regions as shown in Fig. 6.The sub-regions are formed for ease of calculations.The length and width of each subregion are δ x and δ y , respectively.The sub-regions are numbered with indices (i, j, k).The first index in the adopted triplet scheme (i, j, k), varies along the X-axis, the second index varies along the Y -axis and the third index designates the floor number as shown in Fig. 7.

System Model
In this work, we consider an LTE system where Femtos are deployed by network providers, each operating in a different frequency, i.e., no co-tier interference.Also the cross-tier interference between macro BS and  Femtos is ignored.We assume that the existing joint resource allocation [24,25] algorithms can be applied for avoiding co-tier and cross-tier interference.We also assume that Femtos are configured in open access where UEs are authorized to connect with any of the Femtos.All Femtos deployed in an enterprise building get connected to the centralized Femto-GW.In this work we assume the height h of all floors is same, and the length L and width W of all rooms are same.Additionally, the length δ x and width δ y of each sub-region is constant and h m is the average height of an UE.The proposed offloading algorithm in DuD access system will be running in a centralized Femto-GW to address the offloading in Femto cells.The SON feature in Femto-GW can automate the offloading algorithm efficiently.

Formulation of Femto Placement Model
Let us suppose that a given Femto f has the co-ordinates x f , y f and z f .Then, the distance from the Femto to the farthest point in a sub-region defined by (i, j, k) is given by d f ijk (from reference [18], Fig. 2).Table 1 shows the notation used in MILP formulation.
We define a binary variable λ f ρ as 1 if the f th Femto is in room ρ, and 0 otherwise.z f co-ordinate of a Femto is Farthest distance between Femto f and sub-region (i, j, k).F Set of Femtos x f , y f , z f x, y and z are the co-ordinates of Femto f .
1 if Femto f is places at sub-region (i, j, k), zero otherwise.
Represents the room number of Femto f .ρ ijk Represents the room number of sub-region i, j, k.

N
Total number of rooms in the building.
an integer indicating the Femto's residing floor number is shown by, where N is the number of rooms in the building.
Let us assume that f th Femto is residing in the room number ρ zxy .The x and y co-ordinates of the Femto are constrained to be within the bounds of the room in which it is contained as shown in Fig. 8. Equations ( 3) and ( 4) give that the x co-ordinate of the Femto should have the value greater than the left wall and less than the right wall, respectively further represented in Fig. 8.
Similarly, Equations ( 5) and (6) give that the y co-ordinate of the Femto should have the value greater than the lower wall and less than upper wall, respectively.
Assuming that a sub-region is served by only one Femto, we have, The efficiency of a Femto in serving the UEs in a subregion depends on the SNR value in that sub-region.We set a constraint that the SNR value at the farthest point of the sub-region should be higher than the threshold SNR, γ min .This would imply that every point in the sub-region would receive SNR greater than γ min , since SNR decreases with increasing distance from the Femto.The SNR inside the sub-region (i, j, k), γ ijk is given by Here P f is the Femto's transmit power, P N is the noise power, d ijk is the distance between the sub-region (i, j, k) and the serving Femto, L r ref is the loss at the reference distance r ref in linear scale and α is the pathloss exponent.SNR in dB scale is given by, Considering the attenuation factors for the SNR, the total attenuation is given by, where, L * T AF is total attenuation factor and L * FAF and L * W AF are the losses due to floor attenuation and wall attenuation respectively.SNR in dB scale considering wall and floor losses is given by, We assume two more variables for the reciprocal of SNR values.Let, γ min is the threshold value of SNR and its reciprocal is γ min .
γ min = 1 γ min (12) Now Equation ( 8) can be rewritten as [18]: Here C 0 and C 1 are constants depending on the environment.Δ is also a constant depending on the environment and ρ f ∼ ρ ijk is calculated in such a way that it gives the number of obstructions (walls or floors) between the sub-region (i, j, k) and the f th Femto.This special difference (∼) is the absolute value of the digit wise difference between ρ f and ρ ijk as shown in Fig. 5.For e.g.
The Co-efficient of β indicates the number of floors separating the f th Femto and the sub-region (i, j, k).The co-efficient of β 2 indicates the number of walls separates f th Femto and sub-region (i, j, k).So, the rooms 121 and 231 are separated by one floor and one wall.
Since the SNR received in any sub-region should be greater than the threshold SNR, we have, The above constraint (Equation ( 15)) is ensured for all the sub-regions whose occupant probability is greater than zero (i.e., p ijk > 0).If p ijk is the expected peak user density in sub-region (i, j, k), then the placement of Femtos should be in such a way that the product, p ijk γ ijk should be maximum for all the sub-regions.Alternately, the product p ijk γ ijk should be minimum.Hence, our objective is min ijk p ijk γ ijk subject to (1), ( 2), ( 3), ( 4), ( 5), ( 6), ( 7), ( 13) and (15).But Equations ( 1) and ( 13) are non-convex equations which cannot be solved by the available tools.Hence, these equations are first converted to convex equations and then to linear equations.16)

Linearization of Equation
Without loss of generality, we can convert the equality in the equations into inequalities with the help of Lemma 1 [18] as follows, Lemma 1.The constraints in Equation (1) can be equivalently replaced by, Moreover, the inequality in Equation (19) holds as equality to an optimal solution.
Equations ( 17) and ( 18) are expanded as, The above equations can be written as, (with the help of Lemma 1) Accordingly, Equation (19) becomes, Using Piece-wise Linear Approximation (PLAP), the convex constraints ( 24), ( 25) and ( 26) are transformed into linear constraints, which steers the deduction of the following equations, where, w X 1 , w Y 2 and w Z 3 are the positive weights between 0 and 1.The X S , Y S and Z S are the S pieces in their respective domains.

Linearization of Equation (13):
Let, The above equation can be written (with the help of Lemma 1 [18] ) as, Then Equation ( 13) can be written as νf ijk if and only if (from reference [18]) We linearize the convex constraint given in Equation (32) by PLAP and then obtain where, w s are the positive weights between 0 and 1. R S are S pieces in their respective domains.

Decoupled Uplink and Downlink (DuD) Access for efficient Offloading in Femtocell Networks
For uniform UE distribution, optimal Femto placement model presented in the previous section ensures that there is no coverage hole inside the building.However, the next challenge after deploying the Femtos inside the building is load balancing.In reality, all the Femtos will not be fully loaded all the times.In order to offload, some UEs which are in the cell edge region are offloaded from the heavily loaded serving Femto to one of less loaded neighboring (target) Femtos.In traditional coupled access systems, the uplink power of offloaded UEs increases because the target Femto is far from the UE than the serving Femto.But DuD access helps the offloaded UE to get the downlink access from the target Femto while the uplink access is still from the same serving Femto, thereby decreasing the battery depletion at UE.

Downlink Offloading Algorithm
The algorithm runs at the Femto-GW every n Transmission Time Interval (TTIs).We assume that the channel gain remains static for the next n TTIs.Using  the channel gain, the Femto-GW can calculate SNR value of a UE from each Femto.We use SN R uf to denote the SNR value of user u from Femto f .Using SN R uf Femto-GW can calculate the amount of data [26] that can be sent in one resource block (RB).Based on this information, the minimum number of RBs required to maintain the minimum data rate can be calculated.Let minRB uf be the minimum number of RBs required from Femto f to maintain the minimum data rate for user u.Then, Where, Data rate Demand u is the minimum data rate of user u and Data sent in one RB can be calculated using SN R uf .An RB matrix can be constructed using minRB uf value as shown in Fig. 9.
Before starting the offload algorithm, initially every UE is attached to the serving Femto.T otDemand f is the total demand in Femto f and it can be calculated as, Where, u f is the set of UEs connected to Femto f .minRB matrix is constructed using the minRB uf values.
A cell is heavily loaded if the T otDemand of the cell is more than R. Where, R is the total available RBs in each Femto.We also define another variable, TED, which is used to calculate the total excess demand in the system.
The Femtos are arranged in decreasing order based on their T otDemand f values.If T otDemand f is more than R for a Femto f then the Femto is picked for offloading.From the minRB matrix find (u * , f * ) such that u * ∈ u f , f * ∈ F − {f } and minRB u * f * is minimum.Now check if the total RB demand of Femto f * exceeds R if user u * is offloaded to f * .If it does not exceed R, then transfer the UE u * to Femto f * and update T otDemand f * and T otDemand f .If it exceeds R, then assign a large value to minRB u * f * so that the (UE, Femto) pair is not chosen again.Now choose the next best (u * , f * ) pair and proceed.After transferring a UE if the updated T otDemand f is less than R, select the next Femto.After traversing across all the Femtos, calculate the TED value.Stop the algorithm 1 if TED = 0 i.e., every Femto has sufficient RBs to satisfy the user demand.If TED 0, go to step one to repeat Algorithm 1. Stop Algorithm 1 if the TED value remains the same for two iterations.Using minRB matrix find a pair (u * , f * ) such that u * ∈ u f , f * ∈ F − {f } and minRB is the minimum

Experimental Setup and Performance Results
The system model described in Section 3 has been simulated using MATLAB.Table 2 shows the simulation parameters used in this experimental setup.The optimal Femto co-ordinates are obtained by solving MILP model with GAMS tool which uses CPLEX solver [8].The MILP algorithm is an application of a branch-and-bound search including modern algorithmic features such as cuts and heuristics.The MILP optimizer has the capability to solve large and numerically difficult MILP models with features including settable priorities on integer variables, choice of different branching, and node selection strategies.Thus the MILP solver in GAMS (CPLEX) is ideal for our purpose of solving a large MILP model with many equations and variables.

Optimal Femto Placement
To represent an enterprise scenario, we have considered a two-storey building of dimensions (120m × 80m × 6m) with walls as shown in Fig. 10.Each of its subregions are of dimensions (5m × 5m × 3m).We present

A. Uniform UE Distribution
For this case, the average user density on the floor #1 of the building is 1.2 UEs per sub-region and on the floor #2 is 1.9 UEs per sub-region.Proposed MILP placement model has given 20 Femtos as optimal to cover both the floors i.e., 8 Femtos to cover the entire floor #1 (F = 8 i.e., F1, F2, . . ., F8) as shown in Fig. 11 and 12 Femtos to cover the floor #2 (F = 12 i.e., F1, F2, . . ., F12) as shown in Fig. 12. Fig. 11 shows SNR heat map on the floor #1 along with Femtos locations.The darker region in SNR heat map represents the Femto locations with high SNR value with an indoor path loss constant of α = 3.5.In Fig. 11, the Femto F1 and F2 are deployed very close to the wall.For users within the boundaries of the four walls in which the Femto is present, a good SNR in the range of 15 to 35 dB can be guaranteed.But, when the signal has to cross a wall to reach the user, the range of SNR guaranteed decreases drastically to -2 to 15 dB.This is because the signal attenuates faster and degrades the signal strength in the presence of walls.As the distance increases from the Femto, most of the sub-regions get only 0 to 5 dB.Similar trend can be observed on the floor #2, refer Fig. 12. Fig. 13 shows the connectivity region of each of the Femtos on the floor #1.The same colored sub-regions are connected to the same Femto.The average number of sub-regions served by a Femto on floor #1 is more when compare to floor #2, which is due to the large deployment of Femtos in the latter.Similarly, the connectivity region of the floor #2 is shown in Fig. 14.

B. Non-uniform UE Distribution
If the density of UEs is of non-uniform (e.g., random), the proposed Femto placement model still works and provides placement, however, it may lead to coverage holes inside the building.To demonstrate this, we took non-uniform distribution of UEs as shown in Fig. 15.Here we assumed that there are no UEs in the center region of a one-floor building (represented by blue color in Fig. 15).Based on this UE density, Fig. 16 shows the total number of Femtos required (i.e., F= 8) as per the proposed Femto placement model and their connectivity regions.Fig. 17 shows SNR distribution across sub-regions inside the building.The darker region in SNR heat map represents the Femto location with high SNR value.The SNR observed in the center region of the building is < -5 dB and hence there is a coverage hole.

Performance of the Proposed System when the UE Traffic Demand is Uniform
In the following, we consider Femto deployment by assuming uniform UE distribution to avoid coverage holes inside the building and study the performance of the coupled and decoupled systems in the case of uniform UE traffic demand in the network.

A. Performance of the Proposed Offloading Algorithm
In order to analyze the performance of our proposed offloading algorithm we have considered the UE distribution as shown in Figs.18    of Femtos which is based on a uniform distribution of UEs having same data demand implies that the load across all Femtos is uniform.The assumptions of a uniform UE distribution and an equivalent data demand cannot be employed to evaluate the efficiency of our proposed offloading algorithm.To show the potential of proposed offloading algorithm, we assumed non-uniform UE distribution with fixed traffic (i.e., 400 Kbps).
(a) UE Connectivity Before and After Offloading Fig. 20 shows the user downlink connection before and after offloading on floor #1.The Femtos are represented as , while the users are shown as * , , •, •, 2, , , or + having the respective boundary color of the corresponding serving Femto which is serving them.The red lines show the offloaded UEs and their corresponding target Femtos after offloading.The Femtos F1, F3, F4, F5 and F7 are lightly loaded with lesser user count, but the other Femtos (F2, F6 and F8) are heavily loaded.In our work, the offloading algorithm will choose the efficient and closest downlink pair (u * , f * ) by checking the load with all the users from the overlapping neighboring Femtos in such a way that the throughput and load are well balanced.The reason for the offloading algorithm to choose the shortest downlink attachment is due to efficient usage of RBs.Otherwise, it requires more bandwidth from the neighboring Femtos due to less SNR.For example, the Femto F2 is heavily loaded but the neighboring Femtos (F3 and F5) are lightly loaded.Our offloading algorithm offloads 5 UEs (u1, u2, u3, u4 and u5) to Femtos F3 and F5.Similarly F1 chooses the one cell edge UE (u6) from F8 for offloading by the neighboring Femtos.Hence, the total number of offloaded downlink user-Femto pair is represented as U1, U 2, . . ., U7 respectively.Fig. 21 shows the downlink offloading on floor #2.Here, most of the Femtos (F1, F7, F9, F10, F12) are heavily loaded due to large number of UEs.Observations similar to the ones made in the floor #1 scenario can be made in this case as well.

(b) Required RBs in each Femto
Fig. 22 shows the required RBs for each Femto (T otDemand f ) on floor #1 before and after offloading.Before offloading, Femtos F2, F6 and F8 require excess number of RBs (i.e., 3400000 RBS which is greater than the limit 1    has two overlapping target Femtos (i.e., F3 and F5) as shown in Fig. 20.But, the Femtos F3 and F5 are lightly loaded and require only 400000 RBs and 200000 RBs respectively, which is lesser than the limit, 2500000 RBs.So, the target Femtos can share the RBs with the heavily loaded serving Femto F2 which brings down F2's requirement to 2200000 RBs.Similarly, F6 also needs excess RBs due to heavy load (refer Fig. 20).The overlapping Femto F7 can balance the load by sharing their RBs.Hence, minimum data rate (i.e., 400 Kbps) is maintained for all users in each Femto.Further, to guarantee more than the minimum data rate and to ensure fairness among all UEs.Operator should incorporate existing scheduling algorithm like proportional fair during offloading (But it is beyond the scope of our work).Similar pattern can be observed on floor #2 for Femtos F12, F7 and F8 (Fig. 23).

(c) Downlink User Count
Fig. 24 shows the user count on floor #1 for each Femto before and after offloading.If we observe Fig. 24, the number of users in Femto F2 is very high before offloading.After offloading with target Femtos F3 and F5, five users have been offloaded.This reduces the load in the overloaded Femto F2 and the QoS demand for the remaining connected users in Femto F2 is guaranteed.
Similarly, Fig. 25 shows the user count on floor #2 for each Femto before and after offloading.The load in Femtos F1, F7, F9, F10 and F12 was high due to higher number of users.Since F2 is closer, more number of users from F12 got offloaded to F2.We show the SNR value of the offloaded UEs when they transmit with full power in coupled and decoupled access systems.For the fixed uplink SNR, we also show the advantages in terms of the transmission power when UEs are in coupled and decoupled access systems.

B. Comparison between Coupled and Decoupled Access Systems
(a) Maximum Achieved Uplink SNR Fig. 26 shows the maximum SNR value that the UEs can achieve in full power transmission when it connect to the serving Femto or the target Femto on floor #1.Due to heavy load in F8, the user (u6) got offloaded to F1.The user u6 receives roughly +1 dB SNR during uplink from serving Femto F8 because it has to cross 2 walls.Thus for the same uplink transmission power, when it gets connected to the target Femto, it receives very less i.e., roughly -1 dB SNR because it is located in the cell edge of Femto F1.Fig. 27 shows the maximum SNR value achieved on floor #2.As the Femtos are densely deployed on floor #2, the inter distance between the Femtos is very less (refer Fig. 21).This assists most of the users to achieve better SNR value when compared to floor #1.Even after offloading 9 users, all of them are able to achieve a decent SNR value.Compared to decoupled access system in floor #1 and floor #2, uplink SNR has decreased by 52% in coupled access system.
(b) Uplink Power to maintain SN R = 0 dB Fig. 28 shows the uplink power emitted by the downlink offloaded UEs to maintain SN R = 0 dB for floor #1.In this scenario, the serving Femto will always allow the UE to transmit with less power due to shorter distance.For example, to maintain SN R = 0 dB, the UE u6 transmits 0.09 W to the serving Femto.If the same UE wants to connect with the target Femto, it       has to transmit > 0.1 W.However, according to 3GPP standard, the user cannot maintain communication with the target Femto.To maintain communication with the target Femto, the user has to tune its SN R threshold to less than 0 dB.Fig. 29 shows the uplink power emitted by the downlink offloaded UEs to maintain SN R = 0 dB for floor #2.Due to dense deployment of Femtos and less inter distance between them, the uplink power is reduced drastically when compare to floor #1.For example most of the UEs among the 9 are allowed to transmit with less power (i.e., 0.025 W) in closer Femto.As the UE battery power plays an important role in wireless communication, we allow the UEs to communicate the uplink information to the closer Femto.Thus the decouple system helps in saving power.Compared to coupled access system in floor #1 and floor #2, uplink power has reduced by 56% in decoupled access system.

C. Average Performance of the Proposed System
We performed the simulation for 30 different scenarios (i.e., UE distribution is varied arbitrarily by changing the seed value) with the fixed traffic pattern (i.e., 400 Kbps) and the following are the results.Fig. 30 shows the CDF of the maximum uplink SNR achieved in full power transmission by the transferred UEs in the decoupled access system and the coupled access system.On an average, compared to decoupled access system, maximum achievable uplink SNR has decreased by 64% in coupled access system.For a fixed uplink SNR threshold (i.e., 0 dB), on an average, the UE transmission power has reduced by 70% in decoupled EAI European Alliance for Innovation on Future Internet  access system when compared to the coupled access system.
We then vary traffic from 150 Kbps to 550 Kbps in the steps of 50 Kbps in 9 experiments in such a way that all UEs in any given experiment generate fixed traffic in the above interval of (150 Kbps to 550 Kbps).Each experiment was ran for 30 seeds.In Figure 31, X-axis shows the variation in traffic demand and Yaxis shows the percentage decrease in achievable uplink SNR in coupled access system when compared to the proposed decoupled access system.It is plotted with 95% confidence intervals..

Performance of the Proposed System when the UE Traffic Demand is non-uniform
In the following, we consider Femto deployment given by uniform UE distribution and study the performance of the proposed system in the case of non-uniform UE traffic demand (i.e., 300 to 500 Kbps).
As explained in fixed traffic pattern (Section 5.2), the same optimal placement model is used with the assumption of uniform UE distribution.The UE distribution on floor #1 and #2 is same as shown in Figs.18 and 19.We assumed that the traffic patterns are varied from 300 Kbps to 500 Kbps in the steps of 10 Kbps.That means UEs arbitrarily select some traffic pattern in the interval of (300 Kbps to 500 Kbps).Fig. 32 shows the user downlink connection before and after offloading on floor #1.Similarly Fig. 33 shows the user downlink connection before and after offloading on floor #2.Note that all the observations are as similar to what had been reported for the uniform UE distribution and only the offloading UEs will differ based on the traffic load (i.e.. traffic pattern).

Conclusions and Future work
In this paper, considering realistic constraints, we have provided an MILP model for the optimal placement of Femtos based on user occupant probabilities inside an enterprise building scenario to achieve desirable signal strengths for all the users.We conducted extensive experiments in MATLAB to demonstrate the benefits of proposed optimal placement model.We established DuD connections based on the shortest-path loss Femto for the uplink access and a less loaded neighboring Femtos for the downlink access.On average we observed 70% energy savings in decoupled access system when compared to the traditional coupled access system.In future, we intend to consider more complex scenarios involving cross-tier and co-tier interference in the system model and guarantee SINR even at the farthest point in the sub-regions.Also would like to employ different access modes of Femtos while providing an algorithm for optimal Femto placement.

Figure 1 .
Figure 1.Coupled Access System before offloading U t to Femto 1 .

Figure 2 .
Figure 2. Coupled Access System after offloading U t to Femto 1 from Femto 2 .

Figure 3 .
Figure 3. Decoupled Access System after offloading only donwlink of U t to Femto 1 from Femto 2 .

Figure 4 .Figure 5 .
Figure 4. Top view of a floor in the building.

Figure 6 .
Figure 6.Top view of sub-regions in the building.
ijk is the upper bound for ν f ijk .The Bilinear equation (33) holds good within the bound 0 ≤ ν f ijk ≤

Algorithm 1 1 :
Offloading Algorithm Input 1 : F: Set of all Femtos Input 2 : R: Total available resources in each Femto Input 3 : u f : Set of all UEs connected to Femto f Input 4 : SNR matrix, minRB matrix Input 5 : T otDemand f : Total RB demand in each Femto Input 6 : T ED 0 : Initial total excess demand Output: u f Initialization: Iteration Count ← 1; Arrange the Femtos in decreasing order of T otDemand values 2: for f = 1 : |F| do 3: while T otDemand f > R do 4:

Figure 11 .Figure 12 .
Figure 11.SNR distribution and Femto locations given by optimal Femto placement model for uniform UE distribution on the floor #1.

Figure 17 .
Figure 17.SNR distribution and Femto locations given by optimal Femto placement model for non-uniform UE distribution given in Fig. 15.

Figure 18 .
Figure 18.UE Distribution on floor #1 for uniform traffic pattern.

Figure 20 .F12Figure 21 .
Figure 20.UE Connectivity before and after Offloading on floor #1 for uniform traffic pattern.

Figure 22 .
Figure 22.T otDemand f of each Femto before and after offloading on floor #1 for uniform traffic pattern.

Figure 23 .
Figure 23.T otDemand f of each Femto before and after offloading on floor #2 for uniform traffic pattern.

Figure 24 .
Figure 24.Downlink user count in each Femto before and after offloading on floor #1 for uniform traffic pattern.

Figure 25 .
Figure 25.Downlink user count in each Femto before and after offloading on floor #2 for uniform traffic pattern.

Figure 26 .
Figure 26.Maximum Achieved Uplink SNR in coupled and decoupled access systems on floor #1 for uniform traffic pattern.

Figure 27 .
Figure 27.Maximum Achieved Uplink SNR in coupled and decoupled access systems on floor #2 for uniform traffic pattern.

Figure 28 .
Figure 28.Power emitted by UE in coupled and decoupled access systems on floor #1 for uniform traffic pattern.

Figure 29 .
Figure 29.Power emitted by UE in coupled and decoupled access systems on floor #2 for uniform traffic pattern.

Figure 30 .
Figure 30.CDF of the maximum Achieved Uplink SNR in coupled and decoupled access systems.

Figure 36 .
Figure 36.Maximum Achieved Uplink SNR in coupled and decoupled access systems on floor #1 and floor #2 for non-uniform traffic pattern.

Figure 37 .
Figure 37. Power emitted by UE in coupled and decoupled access systems on floor #1 and floor #2 for non-uniform traffic pattern.

Fig. 34
Fig. 34 (a) and Fig. 34 (b) show the required RBs for each Femto (T otDemand f ) on floor #1 and floor #2 before and after offloading.Fig. 35 (a) and Fig. 35 (b) show the user count on floor #1 and floor #2 for each Femto before and after offloading.We show the SNR value of the offloaded UEs when they transmit with full power in coupled and decoupled access systems.For the fixed uplink SNR, we also show the advantages in terms of the transmission power when UEs are in coupled and decoupled access systems.Fig. 36 (a) and Fig. 36 (b) show the maximum SNR value that the UEs can achieve in full power transmission when it connect to the serving Femto or the target Femto on floor #1 and floor #2.Similarly, Fig. 37 (a) and Fig. 37 (b) show the uplink power emitted by the downlink offloaded UEs to maintain SN R = 0 dB for floor #1 and floor #2.

Table 2 .
Simulation Parameters placement results for both uniform and non-uniform UE distributions inside the building.As UE density is factored into the proposed optimal Femto placement model, non-uniform UE distribution case could lead to cover holes inside the building.
and 19.The placement