Mobile Fog Computing-Assisted Resource Allocation for Two-Hop SWIPT OFDM Networks

1 School of Communication Engineering, Hangzhou Dianzi University, Hangzhou 310018, China 2State Grid Energy Research Institute Co., Ltd., Beijing 102209, China 3School of Computer and Communication Engineering, University of Science and Technology Beijing, Beijing 100083, China 4Beijing Computing Center, Beike Industry Park, Beijing 100094, China 5China Academy of Telecommunications Technology, Beijing 100191, China 6Epithelial Systems Biology Laboratory, Systems Biology Center, NationalHeart Lung and Blood Institute, National Institutes of Health, Bethesda, Maryland, USA


Introduction
In the past decade, cloud computing has emerged as a new paradigm.It enables computing, storage, and network managements to centralize in the clouds, which are referred to as data centers, cellular core networks, and so on.With the clouds, vast resources can be provided to resource-constrained devices to satisfy their requirements of computing and storage.However, there is an inherent limitation for cloud computing [1], i.e., the long propagation distance from the end user to the remote cloud center, resulting in very long latency.
Recently, a new trend has been happening; that is, the computing is pushed to the network edge devices due to their progressively enhanced computation capacity.This is called mobile fog computing (MFC) or mobile edge computing (MEC) [2,3], where the network edge devices perform computing tasks instead of remote clouds.Thanks to closer distance to end users, the latency is less and thus realtime tasks can be achieved via MFC, which is an effective supplement to cloud computing.MFC is applicable to delay sensitive tasks while cloud computing to sophisticated but delay-insensitive data processing work.
One potential application of MFC is in Internet of Things (IoT), such as wireless sensor networks (WSN).In WSN, sensor nodes are responsible for data gathering, and sink nodes for collecting and preprocessing data from surrounding sensor nodes and then delivering data to the remote clouds, which perform further complicated data processing and information mining.For WSN, resource allocation (RA) is a key approach to improve system performance.It is carried out according to different channel states and deemed to be a real-time task, as the wireless channels are fast time-varying channels and RA needs to be processed rapidly to adapt the dynamic channels.So MFC is a more appropriate option than cloud computing.
In the field of WSN, cooperative relay communication is deemed as an important technique, as it can guarantee that the far sensor nodes can complete communication with each other via intermediate relaying sensor nodes [4,5].On the other hand, orthogonal frequency division multiplexing (OFDM) is employed in wireless communication networks [6].The combination of relay and OFDM is able to significantly enhance the performance of the system [7,8].
MFC-assisted cooperative relay systems [9,10] and OFDM systems [11,12] have attracted much attention and been widely investigated.In [9], a fog-enabled cooperative communication network was considered, where multiple fog nodes were configured to support two-hop transmissions, and the optimal system performance was achieved by designing time reuse patterns.In [10], cooperative fog computing for the Internet of Vehicles (IoV) was studied, where the cooperation of fog nodes was explored to enhance the system performance.In [11], a MFC-assisted multiuser OFDM network was considered, and the total consumed energy of mobile users was minimized by jointly optimizing subcarrier and CPU time allocation.In [12], the joint subcarrier and power allocation problem in an MFC-based OFDM system was investigated to minimize the maximal delay of all devices.
Meanwhile, WSN are usually energy-constrained networks, and connecting sensor nodes to power grid is impossible sometimes.Batteries can be deployed in sensor nodes, but the batteries capacity is limited and may be hard to be replaced frequently.Recently, wireless power transfer has attracted much attention, in which energy-constrained devices can harvest energy by using wireless signals emitted by system nodes with sufficient energy source.Noting that wireless signal can simultaneously carry and transfer information and energy, this is deemed as the simultaneous wireless information and power transfer (SWIPT) [13].
SWIPT has been widely investigated [14][15][16][17][18][19][20][21][22][23][24].In [14], it was assumed that information decoding (ID) and energy harvesting (EH) are simultaneously carried out by using the received identical signals.Nevertheless, this is deemed not to be realized, and therefore some practical SWIPT receivers were also presented, such as time switching (TS) and power splitting (PS) in [13].In TS receivers, ID and EH are performed in two different phases, respectively.In PS receivers, the wireless signal is split into two streams: one stream enters into the energy receiver to harvest energy, and the other enters into the information receiver to obtain information.In [16], these SWIPT receiver architectures were applied to cooperative two-hop network, and the achievable rate performance was investigated.These architectures have also been widely studied in OFDM systems; see, e.g., [17][18][19][20].But, the existing work mainly concerns point-to-point OFDM systems.For example, in [17,19], the performances of throughput and weighted sum-rate were investigated for multiuser OFDM networks.In [20], max-min fair resource allocation was studied for multigroup multicast OFDM systems.In [18], a new SWIPT receiver architecture was proposed, where one part of the subcarriers was used for ID, and the other part was used for EH.
Recently, some work discussed the SWIPT-enabled twohop OFDM system.In [21,22], the authors considered amplify-and-forward (AF) relaying protocol and the achievable information rates were maximized for two-hop MIMO-OFDM AF relay system.In [23,24], the SWIPT-enabled two-hop OFDM decode-and-forward (DF) relay system was considered, but the subcarrier pairing over the two hops was not involved.In [15], a PS receiver architecture was considered and a RA algorithm was proposed to improve the achievable rate; however, the complexity of algorithm was so high that it was hard to be applied to computationconstrained fog system.
This paper investigates the SWIPT for a MFC-assisted two-hop OFDM network, in which a source node transmits information to a destination with the help of a DF relay.The source is assumed with fixed energy source, while the relay is an energy-constrained node and thus has to obtain energy from wireless signal emitted by the source and further forwards the information of source to destination.
The main contributions of this paper are given as follows.
Firstly, to achieve the simultaneous information and energy transmission, we adopt a different PS ratio PS (DPS) architecture, where a frequency selective power splitter splits the signal on each subcarrier into two streams and thus all subcarriers are of the different PS ratios, which can adaptively change.Further, a particular energy cooperation strategy is considered; i.e., the energy harvested on some subcarrier of the first hop is only used to forward the information received on the corresponding subcarrier, in order to reduce the excessive computational complexity.Unlike the existing work [15], it is assumed that PS receiver splits all subcarriers into two streams with identical PS ratio, which is called identical PS ratio PS (IPS) receiver architecture in this paper.
Secondly, in order to explore the system performance limit of the proposed DPS architecture, a RA optimization problem is formulated to maximize the achievable rate of the system by jointly optimizing the subcarrier pairing (SP), the PS ratios, and the PA at both source and the relay.As the problem is nonconvex and hard to solve, a low-complexity efficient RA algorithm is designed by decomposing it into three separate subproblems.The related computation is operated at the source node, which is generally a sink node of higher computation capacity in WSN.
Thirdly, extensive simulation experiments are performed to discuss the system performance.The results demonstrate that although there are some performance loss of achievable rate of the proposed DPS architecture compared with the existing IPS architecture, the computation complexity of DPS architecture is much lower than IPS architecture.So DPS architecture may be a better option for computationconstrained fog system.This paper is organized as follows.In Section 2, the network architecture and system model are presented, and then RA optimization problem is formulated.In Section 3, an efficient RA algorithm is designed.Simulation results are shown in Section 4 to discuss the performance of the DPS receiver architecture and RA algorithm.In Section 5, this paper is summarized.

Network Architecture and System Model
The considered network architecture is shown as in Figure 1, which is divided into three layers, i.e., data layer, fog layer, and cloud layer.Data layer comprises data nodes, which are responsible for gathering data from surrounding environment.Cloud layer contains vast resources to store and process the amount of data from data layer; meanwhile, it also sends control information to data nodes to instruct their operations.Fog layer is a bridge between data layer and cloud layer, which means that on one hand, it is responsible for collecting and preprocessing data from data layer and delivering data to cloud layer to further process it; on the other hand, it is responsible for forwarding control information from cloud layer to data layer.The information between cloud layer and fog layer is transmitted on wired channels while the information between fog layer and data layer is transmitted on wireless channels.
In this paper, we consider the transmission of control information from cloud layer to data layer.The cloud layer first sends control information to fog layer; fog layer stores the information and then forwards it to data layer.The reason of introducing the fog layer instead of directly using cloud computing is that the wireless channels are deemed to be fast time-varying channels and thus information transmission and RA task have to be performed rapidly to adapt the dynamic channels.It is worth noting that due to the enhanced computing capacity, the fog layer has the capability of performing RA algorithm according to the channel state information (CSI).
To study the information transmission from fog layer to data layer, a MFC-assisted two-hop OFDM network is considered, which consists of one source (S) in fog layer, and one destination (D) and one relay (R) in data layer, as shown in Figure 1. S desires to send information to D with the help of R. No direct link exists between S and D. S is of steady energy supply by connecting to power grid in fog layer and   denotes its power.R is an energy-constrained node operating in half-duplex mode and deploying DF relaying protocol, so it has to obtain energy from the signals of S and then uses the harvested energy to help S to forward information to D. The PS receiver architecture is adopted at R so that it can split the received RF signals into two streams to perform EH and ID, respectively.For such a SWIPT-enabled communication network, each transmission is based on frame of length T, which is divided into two subphases of equal length.
In the first subphase, S sends OFDM symbols to R. The received signal at R on subcarrier  can be expressed as where   and ℎ  , respectively, represent transmitted symbol and channel coefficient on subcarrier i and N is the number of subcarriers. , represents the additive white Gaussian noise (AWGN) from the antenna on subcarrier  at R, which is of zero mean and variance  2 ,att . , represents the transmission power at S on subcarrier  and satisfies In the second subphase, using the stream for ID and EH, R, respectively, decodes the received information and harvests the energy and then reencodes the received information and forwards the reencoded information to D. Subcarrier pairing is adopted, so the information of the first hop received on subcarrier  can be transmitted on the subcarrier  in the second hop.The signal on subcarrier  received at D can be expressed as where   is channel coefficient on subcarrier  at D, and  , is the power on subcarrier  at R.  , is AWGN from the antenna on subcarrier  at D, which is of zero mean and variance  2 ,att .To realize SWIPT, a PS receiver architecture is proposed, where all subcarriers are of different PS (DPS) ratios and can adaptively adjust, which is called DPS architecture.To implement DPS architecture, an analog adaptive passive frequency selective power splitter is required [21,24].In this paper, we consider a particular energy cooperation strategy which makes the computational complexity of the RA algorithm of the DPS architecture significantly decrease and is very meaningful for some communication scenarios where the processing capacity of communication nodes is limited, such as MFC-based WSN.
Let the PS ratios    and    represent the fraction of the signal power used for ID and EH received on subcarrier i, respectively, which satisfy the constraints of Thus, the harvested energy on subcarrier  at R is given by where  denotes the EH efficiency.We consider such an energy cooperation strategy adopted in [21,24], in which the energy harvested on subcarrier  of the first hop is only used to forward the information received on subcarrier  and thus the available power  , on subcarrier  at R can be inferred as The achievable information rate between S and D for DF relay system on a subcarrier pair (, ) can be expressed as [4] ) , where  2  =  Substituting ( 6) into ( 7), then we can write (7) as ) , Thus, the achievable rate of the system can be expressed as where  = {   ,    , ∀} is PS policy and satisfies the constraint (4).P = { , ≥ 0, | ∀, } is power allocation (PA) policy and satisfies (2).S = { , ∈ {0, 1} | ∀, } is SP policy, which represents that if the first hop subcarrier  is matched with the second hop subcarrier j,  , =1; else  , = 0. Further, one first hop (second hop) subcarrier can only match with one second hop (first hop) subcarrier.That is, With the objective of maximizing the achievable information rate of the system, by jointly optimizing the SP, the PA, and the PS ratio, the optimization problem is formulated as (P1): max P,S,   (P, S, ) 4) , (10) . (11)

Resource Allocation Design
In this section, we first describe our proposed resource allocation (RA) algorithm for problem P1 and then we shall prove that it is able to achieve the global optimal solution of problem P1.
3.1.The Proposed Resource Allocation.Our proposed RA is described as Algorithm 1, which is divided into three separate subproblems.In what follows of this subsection (Section 3.1), the detailed process of each step in Algorithm 1 is described, and its global optimality is proven in Section 3.2.
(1) The Optimal SP S * .The proposed SP scheme is only based on the channel power gains.Firstly, according to the channel power gains of the two hops |ℎ  | 2 and |  | 2 , the first hop subcarriers and the second hop subcarriers are, respectively, sorted from highest to lowest.Next, the kth first hop subcarrier is matched with the kth second hop subcarrier, which is equivalent to the optimal S * satisfying that where ) , It is easy to find that in (14) the first term, i.e., log 2 (1 +   /(   2 ,att +  2 ,proc )), is a monotonically increasing function of   , and the second term, i.e., log 2 (1 +   ), is a monotonically decreasing function of   , so, to obtain the optimal solution, the two terms should be equal.Meanwhile, using   +   = 1, the optimal PS factor   can be calculated, according to This is a quadratic equation and its two roots are given as where only the one satisfying the constraints in problem P2 can be considered as the optimal solution.Since   +   = 1,   ,   ≥ 0, we have that 0 ≤   ,   ≤ 1.
It is easy to observe that the one of the two roots is always less than 0, so it is discarded.For the other one, we can prove that it satisfies the above constraint.
Thus, the optimal PS factors can be given by (1) Find the optimal S * from ( 12).
(3) Calculate the optimal P * with the obtained S * and  * from (22).
(3) The Optimal PA P * with the Obtained S * and  * .We have obtained the optimal PS factors   * and   * .As the optimal PS factors are related to the channel gain of the second hop from (18) So for given S * and  * , the PA problem can be formulated as (P3): where SP is the set of subcarrier pairs.From (18), it can be easily found that the optimal PS factors are not related to PA and thus  , is also not related to PA.So this problem is a classical water-filling PA problem, and we can obtain its optimal solution as where [] + = max{0, }, and ] is Lagrangian multiplier and can be solved using

Global Optimum of Our Proposed RA.
In this subsection, we shall prove that although Algorithm 1 is divided into three separate subproblems, it can still achieve the global optimal solution of problem P1, and the result is given by the following theorem.

Theorem 2. The RA in Algorithm 1 achieves the global optimum of problem P1.
Proof.To prove that the RA policy in Algorithm 1 can achieve the global optimal solution of problem P1, we only need to prove that each step of Algorithm 1 maintains the global optimum.From Lemma 1, we have known that the CG-sorted SP scheme in step 1 of Algorithm 1 gives the globally optimal SP policy.The scheme is only related to channel gains, which does not require the knowledge of the optimal PS and PA.
According to the derivation process in step 2 of Algorithm 1, the obtained PS is optimal under the given optimal SP, and it does not require the knowledge of optimal PA.Then in step 3 of Algorithm 1, the obtained PA is optimal under the given optimal SP and optimal PS.Since each step maintains the global optimum, Theorem 2 is proved.

Complexity Analysis. The complexity of step 1 of
Algorithm 1 depends on the adopted sorting method, which is ( log ) if the quick-sort method is applied.Moreover, the complexity of step 2 of Algorithm 1 is (), and the complexity of step 3 of Algorithm 1 is also () (the waterfilling over the sorted  , ) [25].Thus, the total computational complexity can be expressed by ( log ).For comparison, the computational complexity of RA algorithm for IPS architecture in [15] is ( log  + ( + 2)  ), where  is the number of loops in the algorithm [15], so it can be found that the computational complexity of proposed DPS architecture's RA algorithm is on the order of (+2)  / log  less than the IPS architecture.So the proposed DPS architecture may be more proper for MFC-assisted networks, where the devices are of lower computation capacity.

Simulation Results
In this section, some simulation results are given to illustrate the performance of the presented DPS receiver architecture and RA algorithm.The noise powers are assumed follows:   2 att =   2 proc = −33 dBm and   2 = −30 dBm.The three network nodes (S, R, and D) are assumed to be placed on a straight line.The distance from S to D is reference distance and represented by  0 , where  0 = 10 m.The location of R is expressed as   / 0 , where   denotes the distance from S to R. ℎ  and   are, respectively, obtained from the distribution as where  is the path loss factor and set to be 3 and  is the number of taps and set to be 4. Firstly, we discuss the performance of our proposed Algorithm 1.For comparisons, the three other methods are also simulated, i.e., (1)   method: Equal PA without SP.EH efficiency  = 1.In Figure 2, we plot the achievable rates versus the total power   .It is easily seen that our proposed Algorithm 1 is superior to the three other methods.
We also show the effect of EH efficiency  on the achievable rate in Figure 3 and it can be found that when  = 1, the achievable rate is maximum.In the following simulations, to discuss the SWIPT-enabled system's performance limit, EH efficiency is always set to be  = 1.
Secondly, to figure out the system performance of the DPS receiver architecture, we compare the achievable information rates of the proposed DPS architectures and IPS architecture in [15].For IPS architecture, let  be the PS ratio used for ID, and the rest 1 −  part is used for EH;  should satisfy The energy obtained by R is The achievable rate from S to D on each subcarrier pair (, ) can be expressed as ) , and thus the achievable rate of the system is given by To maximize the achievable rate of the system, an optimization problem is formulated as  8) , ( 24) , ( 25) .

(28)
The solution of the problem is given in [15].In addition, conventional non-SWIPT two-hop OFDM system is also compared in order to show the difference between SWIPTenabled and non-SWIPT systems.For the non-SWIPT system, we use the optimal RA algorithm proposed in [8].
In Figures 4 and 5, the achievable rates of the DPS/IPS architectures and non-SWIPT system with respect to   and   / 0 are given, respectively.In these two figures, it can be found that the achievable rate of the non-SWIPT system is higher than SWIPT-enabled IPS/DPS architectures, and DPS architecture is worse than IPS architecture.Moreover, from Figure 5, one can find that, for the SWIPT-enabled system, when R is placed close to S or D, the system can obtain the better performance, and the proposed DPS architecture agrees with the existing IPS architecture.One can also find that, for conventional non-SWIPT system, the achievable information rate achieves maximum when R is placed at the midpoint on the line from S to D.
Finally, we also compare the average running time of IPS and DPS architectures in Figure 6.It shows that the running efficiency of the DPS architecture is far superior to the IPS architecture, which agrees with the analysis of computational complexity in Section 3.3, and thus, for computation-constrained MFC system, DPS architecture is a better option although there are some loss of the achievable rate compared with IPS architecture.

Conclusion
This paper investigated SWIPT for MFC-assisted two-hop OFDM network and proposed DPS receiver architectures.To study the system achievable rate limit, an efficient RA algorithm was given.In simulations, it was found that the achievable rate of the DPS architecture is worse than the existing IPS architecture; however, the computation complexity of DPS architecture is much lower than IPS architecture.So DPS architecture may be a better option for computationconstrained MFC system.

Appendix
Proof of Lemma 1 (a) Two-Subcarrier Case (N=2).Firstly, it is assumed that two hops' channel gains satisfy Observing Algorithm 1, we know the second step and the third step can be applied to any given subcarrier pairing policy in fact, although they are derived from the optimal subcarrier pairing policy.So the achievable information rate of the system using sorted subcarrier pairs (1,1) and (2,2) can be expressed as and the achievable rate using nonsorted subcarrier pairs (1,2) and (2,1) can be expressed as where  1  * ,  2  * are the optimal powers for nonsorted pairing scheme.
To prove Lemma 1, we need to prove  sort >  nonsort ; i.e., (1 +  ).Note that, from (18), we can observe that for given noise power, the optimal PS factor   * , is only related to B, that is, to the channel gain |  | 2 of the second hop since  = |  | 2 / 2  .Thus, for   , we only reserve the subscript .One sees that  , =     ; thus (A.3) is equivalent to Secondly, we can prove that  1 >  2 for the assumption 18), the derivative of   * , with respect to  can be computed as ) . (A.5) One can easily find that ( ); that is,  1 >  2 .Thirdly, for the two-subcarrier case, the explicit solutions of optimal PA can be obtained.When only total power constraint in (2) is considered and inequality constraints are ignored, the optimal  * 1 ,  * 2 for sorted pairing scheme can be derived as Similarly, the optimal  1  * ,  2  * for nonsorted pairing scheme can be derived as It is worth noting that, due to nonnegative power constraint in (2), (A.where inequality is obtained from the aforementioned condition  1  2 −  2  1 > 0 and  1 >  2 . Since (A.10) is a monotonically increasing function of   in the whole interval, (A.10) is always more than 0.
In summary, it is proved that, for all cases, (A.4) always holds.So, for two-subcarrier case, Lemma 1 is proved.
(b) Multisubcarrier Case (N > 2).The two-subcarrier case can be generalized to the multisubcarrier case.A proof by contradiction is adopted.For an -subcarrier relay system with  > 2, suppose the optimal pairing does not follow the sorted pairing rule of Lemma 1, so there are at least two pairs of incoming and outgoing subcarriers that are mismatched according to their channel gains.Without loss of generality, it is assumed that there are two pairs ( 1 ,  1 ) and ( 2 ,  2 ) satisfying Using the result for  = 2, it is found that pairing subcarrier  1 with subcarrier  2 and pairing subcarrier  2 with subcarrier  1 can achieve a higher rate than the nonsorted pairings.Hence, by using this new pairing while maintaining the other subcarrier pairs invariant, the total achievable rate can be increased.This contradicts our assumption on the optimality of a nonsorted pairing scheme.

Figure 1 :
Figure 1: Network architecture of MFC-assisted two-hop OFDM system.

Figure 3 :
Figure 3: Achievable information rate versus EH efficiency  with  = 64 and R located at the midpoint between S and D for DPS architecture.

Figure 4 :Figure 5 :
Figure 4: Comparison of achievable information rates of DPS architecture, IPS architecture, and non-SWIPT system versus   with  = 4 and R located at the midpoint between S and D.

Figure 6 :
Figure 6: Comparison of running time of IPS and DPS architectures versus the number of subcarriers  with R located at the midpoint from S to D and   =10dBm.
(7)signal processing noise on any subcarrier of R and D. In(7), the first part log 2 (1 +    |ℎ  | 2  , /(    2 ,att +  2 ,proc )) represents the mutual information from S to R on subcarrier i, and the second part log 2 (1 + |  | 2  , / 2  ) represents the mutual information from R to D on subcarrier .The coefficient 1/2 in (7) is because each frame is composed of two subphases of equal length.
2 ,att + 2 ,proc represents the total noise power of D on each subcarrier,  2 ,proc and  2 ,proc , respectively, represent the power Number(|  | 2 ) represents the serial number of |  | 2 inside all |  | 2 for  ∈ {1, 2, . .., } with an degressive sorting sequence for  ∈ {ℎ, }, respectively.This scheme is called the channel gain-(CG-) sorted SP scheme.The optimality of this scheme is given as follows.To prove this lemma, two-subcarrier case is first considered and proved.Then it is further extended to general multisubcarrier case.See the Appendix for details.To simplify the expressions, let  = |ℎ  | 2 ,  = |  | 2 / 2  and denote  , as   .Since the subcarrier pair is fixed, we further drop the indexes i,j in this subsection.
(2) The Optimal PS  * with the Obtained S * .First, it is easily found that the problem can be decomposed into  subproblems due to the independence of each subcarrier pair.For any given subcarrier pair (, ), the subproblem can be expressed as (P2): Figure 2: Achievable information rate versus   with  = 64 and R located at the midpoint between S and D for DPS architecture.