The Tactile Internet is a new service portfolio of the next generation of wireless networks, e.g., the fifth generation wireless networks (5G), where a novel communication paradigm is introduced. For instance, via the Tactile Internet touch sensation can be remotely transmitted. One of the most important requirements of Tactile Internet service is ultra low end-to-end (E2E)e.g., E2E delay should be less than one millisecond [1, 2, 3, 4]. These requirements cannot be guaranteed via existing wireless networks such as fourth-generation wireless networks (4G) . However, 5G platform via its own soft, virtualized, and cloud based architecture can be leveraged to implement the Tactile Internet services [1, 2].
For instance, via the concept of cloud radio access network (C-RAN) in 5G, spectral efficiency (SE) and energy efficiency (EE) along with cost can be efficiently optimized, where baseband processing is performed by the baseband units (BBUs) which is connected to remote radio heads (RRHs) via the fronthaul links [5, 6].Specifically, C-RAN reduces energy consumption and cost, and improves throughput in dense environment [7, 8]. Therefore, this RAN architecture is a proper environment for the implementation of the Tactile Internet services in hotspot areas.
On the other hand, proper design and selection of multiple access technologies have a significant impact on the SE and EE. One of the promising approaches for the next generation wireless networks is the power domain non-orthogonal multiple access (PD-NOMA)111We use PD-NOMA and NOMA interchagebly in the followings. which can significantly improve the SE and reduce the transmission delay . In NOMA, the available spectrum is shared among different users of one RRH. A NOMA transmitter requires implementing a method for superposition of different users signals. Consequently, to compensate interference between users, a complex method for decoding the signals are required at a NOMA receiver. Successive interference cancellation (SIC) technique is one of these methods which is implemented at the receiver to decode the desirable signal [10, 11].
In the next generation wireless networks, due to the introduction of various services, each to be provided with a high quality of service (QoS) via the virtualization techniques, the concept of slice has been defined for each service in which each slice is a bundle of users with a specific set of QoS requirements [12, 13]. The slice concept adds flexibility to utilizing resources which leads to higher SE and EE. However, in this concept, the isolation between slices should be preserved such that the activities of users of one slice do not have harmful effects on QoS of the users of other slices. One of the major issues in the slicing is how to translate the isolation concept to the proper notation for the networks’ procedures. There exists a large body of work for this translation such as dynamic and static methods [14, 12, 13].In this paper, we consider the minimum required rate of each slice as a means of preserving the isolation between slices [14, 15].
Obviously, for this setup due to the complexity of system architecture, diverse transmission parameters such as power, and different QoS requirements, resource allocation are highly essential which has drawn a lot of attention recently [16, 3, 17, 18, 19]. For instance, in , a resource allocation problem for Tactile Internet in the Long-Term Evolution-Advanced (LTE-A) is investigated where the average queuing delay and queuing delay violation in one base station (BS) are optimized. Orthogonal frequency division multiple access (OFDM) and single carrier frequency division multiple access (SC-FDMA) are considered for downlink (DL) and uplink (UL), respectively. The cross-layer resource allocation problem for Tactile Internet is prossed in 
for single BS where the packet error probability, maximum allowable queuing delay violation probability, and packet dropping probability are jointly optimized with the objective to minimize the total transmit power subject to maximum allowable queuing delays. In, queuing delay, packet loss induced by queuing delay violation, packet error, and packet drop caused by channel fading are considered for analyzing the E2E delay of RAN. In , the effect of frequency diversity and spatial diversity on the transmission reliability in UL is studied in the Tactile Internet service where the number of subcarriers, the bandwidth of each subcarrier, and the threshold for each user are optimized for minimizing the total bandwidth to ensure the transmission reliability. In , a multi-cell network based on frequency division multiple access (FDMA) with a fixed delay for backhaul is studied in the Tactile Internet service. Moreover, queuing delay, delay violation probability, and decoding error probability are considered for analyzing the E2E delay of Tactile Internet service.
In the above-mentioned works, a network is considered in which for each user one queue at the BS is assumed. Therefore, by increasing the number of users, a lots of queues are needed at the BS. However, given that the Tactile Internet is assumed to be implemented in the 5G framework, it is necessary to consider C-RAN architecture. There exists a set of RRHs in highly dense network which are connected to BBU center via fronthaul links. Moreover, previous works consider orthogonal multiple access techniques instead of NOMA. In addition, the results in the above works generally ignore the fronthaul delay. However, due to the importance of delay in the Tactile Internet, it is crucial to consider queuing delay in fronthaul as well, otherwise, the resulting allocation of the resources may not practically fulfill the requirement of Tactile Internet.
To address the mentioned issues, we consider a C-RAN architecture serving a set of tactile users. The contributions of this paper are as follows, many of which have been considered for the first time in Tactile Internet:
We propose a C-RAN scenario in ultra dense environment in 5G platform. This will impose new constraints to the system as far as the number of queues is concerned. For the considered C-RAN architecture, we propose a practical queuing model for sequential queues in Tactile Internet that can be implemented in realistic networks. We consider a PD-NOMA scheme for our system model while all earlier works are based on OFDMA and FDMA schemes. Moreover, we consider slicing for Tactile Internet service in our work.
In contrast to [3, 17, 16] where fronthaul delay is ignored, we take this delay into consideration. Moreover, we consider dynamic adjustment of the access and fronthaul delays based on channel state information (CSI) for each pair of users instead of fixed maximum delay values per each transmission part of our setup and show that it can significantly reduce the required total transmit power.
The rest of this paper is as follows. In Section II, the system model is described. In Section III, we formulate the optimization problem. Numerical results and simulation are presented in Section IV. Finally, Section V concludes the paper.
Ii System Model
We consider a C-RAN network where all RRHs are connected to the BBU via fronthaul links. In this region, there exist several pairs of tactile users where each user aims to send its information to its own tactile user pair via the closest RRH through the uplink transmission link. Then, RRH sends the received data to the BBU via the fronthaul link. The BBU processes all received data and then sends data to the corresponding RRH of its paired tactile user. Finally, this RRH transmits the relevant message to the paired tactile user via downlink (DL) transmission link. Assume each RRH has only one queue for UL transmission and all the data of tactile users is stored in this queue. RRHs send all data to the BBU to process. Then BBU sends data to the corresponding RRH. In DL, we consider each RRH has a queue for each user for sending data to the paired users such as connected car and telemedicine applications.
As shown in Fig. 1, in the considered system model, we have RRHs, slices and pairs of tactile users. Slice contains tactile users and the total number of tactile users in our system model is equal to pairs of users. The terms of access link and fronthaul link often are used to express the RRH-user connection and RRH-BBU connection, respectively. In order to reduce the cost of cabling, wireless fronthaul is used instead of fiber fronthaul [20, 5]. Considering that fronthaul links are provided via wireless channels in ultra dense environment, we assume there exist two sets of subcarriers and for access and fronthaul links, respectively. Moreover, we define for simplicity. We consider a two-phase transmission; in the first phase, all tactile users send their data to the corresponding RRH and simultaneously all RRHs send their buffered data to one BBU via fronthaul links.
In the second phase, all RRH send data to the corresponding tactile users , and simultaneously, BBU sends buffered data to all RRH via fronthaul links. These two phases do not perform at the same frequency. Therefore, the proposed system model is based on the frequency division duplex (FDD) transmission mode in which each RRH can transmit and receive simultaneously in different frequencies. In order to isolate slices, a minimum required rate in each slice must be reserved [21, 14, 22]. By considering the above definitions, we can now proceed to review the system parameters.
Remark 1: Note that the complexity of SIC cancellation in NOMA is primarily a function of the number of code layers in superposition coding . By assuming that two users is allocated to each subcarrier, the complexity of SIC receiver is constant and negligible due to the fact that only two layers superposition coding is required for this case [24, 25]. In addition, since strong and low latency codes such as low density parity check code (LDPC) can be utilized for decoding in our setup, the decoding delay of the signals can be neglected . In particular, for this type of detection codes, a decoding delay is around which can be omitted in this setup [4, 26, 27]. Moreover, the LDPC can increase the reliability of decoding.
To estimate the CSI for DL transmission, pilot signals are transmitted via RRHs to all users. Then, each user sends the estimate channel sends to RRHs via feedback channels. To estimate the CSI for UL, pilot signals are transmitted via users to RRHs, and then, the estimate channels are sent to the users. For the CSI estimation, one of the proposed approaches in[28, 29, 30] can be applied.
Ii-a Access Links Parameters
We introduce a binary variablewhich is set to 1 if subcarrier is assigned to user in slice at RRH , i.e.,
Since we deploy PD-NOMA in this setup, each subcarrier can be allocated to users . Therefore, we have the following constraint
Here, for all , , , and , the achievable rate for user on subcarrier at RRH can be calculated as
where , in which , , and represent the transmit power, channel power gain from RRH to user on subcarrier in slice , and noise power, respectively. Also, is intra-cell interference which is equal to and is inter-cell interference which is equal to . Therefore, the total achievable rate in the access links at RRH is as follows
Due to the power limitation of each RRH in DL transmission, we have the following constraint
Moreover, due to the power limitation of each user, we have
Due to using NOMA in access, we have SIC constraint as follows
Ii-B Fronthaul Links Parameters
We introduce a binary variable denoting that subcarrier is assigned to RRH , and defined by
Assuming that PD-NOMA is also deployed for the fronthaul links, again each subcarrier can be allocated to at most RRHs, and hence, we have the following constraint
The achievable rate for each RRH on subcarrier is calculated as follows
where is defined as where is the interference among RRHs and is represented as . Therefore, the total achievable rate in BBU is obtained as follows
Due to the power limitation of each RRH in UL transmission, we have
Moreover, due to the power limitation of BBU, we have
Due to using NOMA in fronthaul, we have SIC constraint as follows
Ii-C Queuing Delay Model
The total delay of this architecture consists of three components: delay resulting from UL queues at RRH, BBU queue, and DL queues at RRHs, as shown in Fig. 2. Due to delay constraint in Tactile Internet service, we have
where , , , and are delays of UL queues at RRH, BBU queue, DL queues at RRH, and total delay, respectively.
Ii-C1 UL Queuing Delay
where is the statistical QoS exponent of the RRH. A larger indicates a more stringent QoS and a smaller implies a looser QoS requirement. is the number of bits arrived at RRH queue defined as . The probability of queuing delay violation for RRH can be approximated as
for all where is the RRH delay, is the maximum delay, and is the non-empty buffer probability. Equation (5) can be simplified to
Therefore, we have
Ii-C2 BBU Queuing Delay
We consider a queue for all RRH at the BBU for processing data. Therefore, the formulas in the previous section can also be used for this section. The effective bandwidth for each queue in BBU is where is the statistical QoS exponent in the BBU and is the number of bits arrived at the queue in the BBU which is defined as The probability of queuing delay violation at the BBU can be approximated as
where is the non-empty buffer probability. Equation (6) can be simplified to
Therefore, we have
Ii-C3 DL Queuing Delay
The effective bandwidth for each user in RRH is defined as where is the statistical QoS exponent of the user in RRH and is the number of bits arrived at user queue in RRH which is defined as . The probability of queuing delay violation for user can be approximated as
where is the user delay in RRH and is the non-empty buffer probability. Equation (7) can simplified to
Therefore, we have
In order to avoid bit dropping, the output rate of queues must be greater than the input rate of queues. Therefore, we have two following constraints
Iii Optimization Problem Formulation
In this section, our aim is to allocate resources to minimize the overall power consumption in our setup by considering a bounded delay constraint to satisfy the E2E delay requirements. Based on the mentioned constraints C1-C14, the optimization problem can be written as
The optimization variables in (8) are subcarrier and power allocation for different users in access and fronthaul as well as in both UL and DL where , , , and
are the transmit power, the access subcarrier allocation, fronthaul subcarrier allocation, and delay vector for users, respectively. C15 is the rate constraint for isolation of slices. According to the delay and SIC constraints in the optimization problem, the one of the outputs of the optimization problem is the pair of NOMA users that satisfy SIC and delay constraints. In fact, the SIC constraint defines the conditions for executing SIC for two distinct users. If this condition is not met, each subcarrier is exclusively allocated to each user.
Constraints C1-C8 are linear functions of the optimization variables. In problem (8), the rate is a non-convex function, which leads to the non-convexity of the problem. In addition, this problem contains both discrete and continuous variables, which makes the problem more challenging. Therefore, we resort to an alternate method to propose an efficient iterative algorithm [33, 34] with three sub-problems, namely, subcarrier allocation sub-problem, power allocation sub-problem, and delay adjustment sub-problem which will be explained in the followings.
Iv An Efficient Iterative Algorithm
As mentioned earlier, to solve (8), we deploy an iterative algorithm that divides the problem into three sub-problems and solve them alternately [33, 34]. This procedure is presented in Algorithm.1. Let be the iteration number and , , and be the initial values. In each iteration, we solve each sub-problem with considering the optimization parameters of other sub-problems as fixed values derived in the previous steps. The iteration stops when the error in Step 5 is less than a predetermined threshold, i.e., , or the number of iterations exceeds a predetermined value i.e., . The solution of the last iteration is then declared as the solution of (8).
Proposition 1 : The presented iterative algorithm which is described in Algorithm.1 converges.
See Appendix A. ∎
Iv-a Subcarrier Allocation Sub-Problem
With assuming fixed value and , the subcarrier allocation sub-problem is written as follows
While (9) has less computational complexity than (8), still it suffers from non-convexity due to the interference in rate functions. In addition, this problem contains discrete variables. We apply time sharing method and relax discrete variables as and . To solve this problem, we use the difference of two convex (DC) functions to transform the problem into a convex form. The subcarrier allocation sub-problem is transformed into the following form (See Appendix C):
Proposition 3: The proposed iterative algorithm based on the SCA method for subcarrier allocation step converges.
See Appendix B by considering fixed value for power (). ∎
Iv-B Power Allocation Sub-Problem
For the fixed value of , and , the power allocation sub-problem is obtained as follows
Similar to the subcarrier allocation sub-problem, in problem (11), the rate is a non-convex function, which leads to the non-convexity of the problem. Therefore, it is necessary to approximate (11) with a convex problem. To solve this problem, we use the DC approximation to transform the problem into a convex form. Therefore, the power allocation sub-problem is transformed as follows (See Appendix C):