HARQ Optimization for PDCP Duplication-Based 5G URLLC Dual Connectivity

Changsung Lee, Junsung Kim, Jaewook Jung, Jungsuk Baikand Jong-Moon Chung,＊

1Samsung Research, Seoul, 06765, Korea

2Samsung Electronics, Suwon, 16677, Korea

3School of Electrical and Electronic Engineering, Yonsei University, Seoul, 03722, Korea

Abstract: Packet duplication (PD) with dual connectivity (DC) was newly introduced in the 5G New Radio (NR) specifications to meet the stringent ultra reliable low latency communication (URLLC) requirements.PD technology uses duplicated packets in the packet data convergence protocol(PDCP) layer that are transmitted via two different access nodes (ANs) to the user equipment (UE) in order to enhance the reliability performance.However, PD can result in unnecessary retransmissions in the lower layers since the hybrid automatic retransmission request (HARQ) operation is unaware of the transmission success achieved through the alternate DC link to the UE.To overcome this issue, in this paper, a novel duplication-aware retransmission optimization (DRO) scheme is proposed to reduce the resource usage induced by unnecessary HARQ retransmissions.The proposed DRO scheme can minimize the average channel use while satisfying the URLLC requirements.The proposed DRO scheme derives the optimal HARQ retransmission attempts for different ANs by solving a nonlinear integer programming (NLIP) problem.The performance of the proposed DRO scheme was evaluated using MATLAB simulation and is compared to the existing 5G HARQ support schemes.The simulations results show that the proposed DRO scheme can provide a 14.71% and 15.11% reduced average channel use gain compared to the selective data duplication upon failure (SDUF) scheme and latency-aware dynamic multi-connectivity algorithm (LADMA) scheme, respectively, which are the existing 5G PD schemes that use HARQ.

Keywords: 5G networks; URLLC; dual connectivity; retransmission; packet duplication

1 Introduction

Ultra reliable low latency communication (URLLC) is one of the fifth generation (5G) mobile service modes that can support high reliability wireless networking with very lowtime delays.Although different design goals are required for different URLLC applications (e.g., autonomous driving,drone control, factory automation, remote surgery), one of the stringent quality of service (QoS)requirements for URLLC traffic is the 99.999% reliability within 1ms packet latency [1].To support URLLC traffic in 5G networks, novel design concepts have been proposed in the physical (PHY) and medium access control(MAC)layers.For example, a short transmission time interval(sTTI)and faster processing speeds were applied to achieve the low latency requirements, and hybrid automatic repeat request (HARQ) was enhanced to achieve the high reliability requirements.However, in addition to the novel PHY and MAC designs, changes in the upper layers are also required to meet the stringent URLLC service requirements.

Dual connectivity (DC) was first introduced in the 3GPP Release 12 specifications in order to enable user equipment (UE) to access two different access nodes (ANs) simultaneously.The initial objective of DC was to achieve downlink (DL) data rate enhancement by splitting the user plane (UP)data, however, DC was recently extended to support reliability enhancements by exploiting packet duplication (PD).In the 3GPP Release 15 specifications, the PD functionality was adopted as a part of the packet data convergence protocol (PDCP) layer defined under the name PDCP duplication [2].By duplicating URLLC packets in the PDCP layer and transmitting them through different ANs, an UE can obtain macro diversity.As a result, PD enables an UE to obtain not only higher transmission reliability, but reduced latency since it can alleviate the potential need for a retransmission.

However, PDCP duplication can give rise to unnecessary retransmissions if HARQ is operated in the lower layers.When duplicated packets are transmitted via different ANs, the packet delivered through one AN might be successful and the other packet sent through the other AN may be unsuccessful.In this case, the HARQ process will not be needed, but HARQ will be executed since the HARQ entity would be unaware of the successful packet delivery through the other AN.This redundant retransmission of HARQ will lead to a waste of radio resources.As the number of data traffic and connected devices in the mobile network increase, efficient management of radio resources become more essential in 5G networks [3,4].In order to overcome this problem, a novel HARQ optimization scheme that cooperates with the PDCP duplication process is proposed.The original contributions of this paper are summarized as follows.

•To minimize redundant wireless resource usage in PDCP duplication, the duplication-aware retransmission optimization (DRO) scheme is proposed.

•The expected channel use model and the corresponding optimization statement is derived considering the HARQ latency and packet reliability model.

•The optimal number of HARQ retransmission attempts for different ANs is derived by solving a nonlinear integer programming (NLIP) problem based on a gradient descent method.

•The performance of DRO is compared to non-optimal retransmission control schemes such as selective data duplication upon failure (SDUF) and latency-aware dynamic multi-connectivity algorithm(LADMA),and the simulation results show that DRO can provide the lowest average channel use while satisfying the URLLC requirements.

2 Related Work

Several papers have focused on the reliability enhancement of PDCP duplication in 5G networks.In [5], the feasibility and effectiveness of PDCP duplication were investigated, where the numerical results show that PD can help satisfy the latency and reliability requirement of URLLC.In [6], a reliability-oriented multi-connectivity (MC) concept was proposed based on PDCP duplication, which, includes a novel admission mechanism to control the number of MC users.In [7], an analytical study on the reliability of MC was conducted while considering HARQ retransmission.In [8], the LADMA scheme was proposed to reduce the resource usage while satisfying the latency requirement, which activates MC only for users with a high latency violation probability.In [9], the SDUF scheme was proposed to reduce the radio resource consumption by activating DC only upon a transmission failure on the primary link.However, the number of HARQ retransmissions and flexible numerology which have an effect on the overall latency are not considered in the previous PDCP duplication works.

Other papers have studied on the HARQ retransmission aspect in 5G networks.These works focused on HARQ and efficent resource allocation for single connected users.In [10], an optimal resource allocation and repetition coding based HARQ retransmission scheme was proposed to minimize the necessary bandwidth required for URLLC traffic.In [11], a novel resource allocation scheme for retransmissions was proposed, where corresponding analytical models for loss rates were derived.In [12], a Pollaczek-Khinchine (P-K) formula based quadratic optimization (PFQO) scheme was proposed, which can minimize the URLLC bandwidth requirement based on the queueing analysis of HARQ.

3 System Model

Consider an UE which is connected to the master gNB (M-gNB) and the secondary gNB (S-gNB)simultaneously in the NR-NR DC (NR-DC) architecture [13], as shown in Fig.1.The indexi∈{1,2}denotes one of the dual-connected gNBs, wherei= 1 represents the M-gNB andi= 2 represents the S-gNB.Incoming URLLC packets of the M-gNB are duplicated in the PDCP layer, and transmitted via the M-gNB and S-gNB.Let γibe the signal-to-interference-plus-noise ratio (SINR) with respect to gNBi.For each gNB, incremental redundancy HARQ (IR-HARQ) is considered, where a packet is retransmitted if the UE fails to decode the received packet successfully.The index of transmission attempt for gNBiismi= 1, 2,...,Mi, whereMidenotes the maximum number of transmission attempts for one packet.For example,,mi= 1represents the initial transmission,andmi≥2represents a retransmission.For NR numerology, a flexible frame structure is considered.One slot consists of 14 orthogonal frequency division multiplexing (OFDM) symbols, where 1 OFDM symbol duration isTs= 2-μ/14 ms and subcarrier spacing (SCS) is Δf= 15×2μkHz (μ∈{0, 1, 2, 3, 4}).In addition,mini-slots are considered for sTTIs, which can consist of 1～13 OFDM symbols.

Figure 1: Example of the NR-DC architecture when packet duplication occurs in the PDCP layer

3.1 Packet Error Rate Model

A packet error occurs after a decoder fails to correct an erroneous packet, and thereby a retransmission of the erroneous packet is induced.Letpi,mbe the packet error rate (PER) at themth transmission attempt, which can be modeled as

wherean,mandgn,mare parameters of them-th transmission attempt when the modulation and coding scheme (MCS) isn, and γth,mis the SINR threshold when the PER is 1 [14].The example of MCS parameters is provided in Tab.1 [15].Consider that the channel is in then-th state if the received SINR falls into the range of [Γn,Γn+1).Based on the adaptive modulation and coding (AMC) scheme,MCSnis assigned for then-th channel state, where the SINR threshold to then-th channel state Γncan be expressed as below [15].

Table 1: Example of MCS parameters

3.2 Finite Block Length Model

Since URLLC packets are small in general, finite blocklength theory can be used to derive the amount of resource usage [16].An URLLC packet ofLbits can be transmitted with the PER ofpi,mbased on channel useriwith a given SINR γias

whereCi= log2(1 +γi) is the Shannon capacity based on the infinite blocklength andV(γi) =(log2(e))2(1 - 1/(1 +γi)2).As a result, the channel userican be approximated as in (4).

3.3 Packet Reliability Model

Since PDCP duplication is used in the NR-DC architecture, the reliability of the URLLC packet can be enhanced.LetPout,i.be the packet loss rate (PLR) of gNBi.Since a packet is considered to be lost afterMitransmission attempts, the packet delivery failure probabilityPout,ican be defined as in (5).

In the NR-DC architecture, the URLLC packet fails if both the original packet and the duplicated packet are lost at the M-gNB and S-gNB.Consequently, the packet reliability in the NRDC architectureRDCcan be modeled as in (6).

3.4 Latency Model

LetTone,ibe the one-shot transmission latency from gNBi, which is defined as the duration from the time that a packet arrives at the lower layer of gNBi, to the time that a packet is decoded at the UE.Therefore,Tone,ican be expressed as

whereTgNB_procis the gNB processing time (i.e., time to prepare transmission and make a scheduling decision) andTDL_alignis the packet alignment time (i.e., time until the next monitoring occasion (MO)).In addition,TDL_Tx,iis the packet transmission time andTUE_procis the UE processing time (i.e., time it takes to decode a received packet and prepare an ACK).LetTHARQ,ibe the maximum latency afterMitransmission attempts of gNBi.Since a packet is retransmitted by HARQ after a decoding failure,THARQ,ican be expressed as

whereTACK_alignis the ACK alignment time, andTACK_Txis the ACK transmission time.In Fig.2, an example of the latency timeline is presented based on HARQ retransmission.

Figure 2: URLLC latency timeline based on HARQ retransmission mechanism

4 Proposed Scheme

The proposed DRO technique is a selective duplication scheme [8,9], where a duplicated packet of the S-gNB is transmitted only when an initial transmission of the M-gNB fails.Therefore, unnecessary transmissions of the duplicated packets are prevented and the total number of transmission attempts can be effectively reduced.Fig.3 shows an example of the selective duplication scheme when an initial transmission of the M-gNB fails.An incoming packet is duplicated in the PDCP layer of the M-gNB and the duplicated packet is forwarded to the S-gNB only after HARQ NACK reception.Otherwise,if the initial transmission of the M-gNB is successful, ACK is transmitted instead of NACK and the duplicated packet in the M-gNB is discarded.In addition to selective duplication, DRO derives the optimal number of maximum transmission attempts to minimize the radio resource consumption.

Figure 3:Signaling flow of selective duplication, where duplicated packet forwarding and transmission is carried out only after HARQ NACK reception

4.1 Problem Formulation

LetrDCbe the total channel use to transmit an URLLC packet in the NR-DC system.The objective of the proposed DRO scheme is to derive the optimal number of maximum transmission attempts for each gNB that can minimize the expected channel use E[rDC] under the URLLC QoS requirements.Therefore, the optimization statement can be formulated as

subject to (C1) :RDC≥1 -δ

(C2) :THARQ,i≤d∀i∈{1,2}

(C3) :ri=khisi∀i∈{1,2}

(C4) :hi≤Wi∀i∈{1,2}

where (C1) represents the URLLC reliability condition, in which the reliabilityRDCshould be greater than the reliability requirement 1 -δ, (C2) represents the URLLC latency condition, in which the maximum transmission delayTHARQ,ishould be smaller than the latency requirementd.Due to the selective duplication scheme, packet transmission from the S-gNB is delayed for one-shot transmission and the feedback time from the M-gNB and additional signaling delayTPD_delay.Therefore,the maximum latency of each gNB can be rewritten as (10) and (11).

In addition, (C3) represents the channel use condition.The channel use in the time-frequency plane can be derived by the required bandwidth resourcehi, the required time resourcesi, and the number of channel use per unit time per unit bandwidthk[10].Since the packet transmission time is allocated based on the OFDM symbol duration,TDL_Tx,ican be expressed as

Lemma 1.Suppose pi=pi,1=pi,2=···=pi,Mifor all i∈{1,2}.If the maximum number of transmission attempts on the i-th connection Miincreases from Mito Mi+ 1, then the expected channel useE[rDC].increases by-r1lnp1(p1)M1/(1 -p1)for i= 1, and-p1r2lnp2(p2)M2/(1 -p2)for i= 2.□

Proof.The expected channel use E[rDC] can be expressed using the expected number of transmission attempts E[mi] as follows.

The expected number of transmission attempts for M-gNB can be derived as

Algorithm 1: DRO algorithm DRO Algorithm Input: SINR measurement γ1,γ2 Output: maximum transmission M1, M2 1 Initialize M1= 1, M2= 1 2 while do 3 for gNB i∈{1,2} do 4 Compute pi, ri, TDL_Tx,i 5 Compute maximum latency THARQ,i 6 end 7 Compute packet reliability RDC 8 if RDC≥1 -δ and THARQ,i≤d then 9 Terminate the loop (optimal solution)10 else 11 Compute∂E[rDC]/∂Mibased on (17)12 Select j = argmini(∂E[rDC]/∂Mi)13 Increase Mj←Mj+ 1 14 Update maximum latency THARQ,j 15 if THARQ,j＞d then 16 Decrease Mj←Mj- 1 17 Increase Mk←Mk+ 1 for k≠ j 18 Update maximum latency THARQ,k(Continued)

Algorithm 1: Continued 19 if THARQ,k＞d then 20 Terminate the loop (there exist no feasible solutions)21 end 22 end 23 end 24 end

However, the expected number of transmission attempts for the S-gNB can be derived as

since the S-gNB transmits a duplicated packet after transmission failure of the M-gNB.Therefore, the expected channel use E[rDC] can be derived as

Therefore, the partial differentials of E[rDC] can be derived as

In addition, the partial differential∂E[rDC]/∂Mi＞0 is satisfied sinceri＞0 and 0＜pi＜1.■

The approximation gap ofpi=pi,1=pi,2=···=pi,Miwas neglected from the computation for simplicity since it has almost no effect on the expected channel use and optimal solution (0.953% error on average) based on the simulation results.

4.2 DRO Scheme

SinceM1andM2are integer values, the optimization statement of (9) becomes a NLIP problem.In addition, the expected channel use.E[rDC] is a monotonically increasing function since∂E[rDC]/∂Mi＞0.Then,the minimum channel use can be achieved by increasing the maximum number of transmission attempts of the gNB which has the least gradient value∂E[rDC]/∂Miuntil satisfying the reliability condition while not violating the latency condition.Therefore, the optimal number of maximum transmission attempts can be found iteratively based on Lemma 1.Basically, the DRO scheme is conducted by the M-gNB periodically to update the maximum transmission attempts according to the dynamic channel condition.The pseudo code of the DRO sc.he is presented in Algorithm 1.

As the input, the SINR measurements with respect to the reference signal are used, which are reported periodically from the UE.First,the maximum number of transmission attempts are initialized to 1 in line 1, then the DRO scheme iterates a loop until it achieves the URLLC requirements in line 2.The maximum latencyTHARQ,iand the reliabilityRDCare computed basedonpi,ri,TDL_Tx,iinlines3-7.If the DRO scheme satisfies the URLLC requirements, it terminates the loop in lines 8-9.Otherwise, the partial differentials∂E[rDC]/∂Mi＞0 are computed based on (17) in lines 10-11.Then, the maximum number of transmission attempts for gNBjis increased by 1, wherejis an index of the gNB which least increases the expected channel use compared to the other gNBs in line 12-13.To examine if any violation of the latency requirements occurs, the maximum latencyTHARQ,iis updated in line 14.If the latency requirement is violated,Mjis decreased by 1 and the maximum transmission attempts of the other gNBMkis increased by 1 instead in lines 15-17.If the latency requirement is violated after updatingTHARQ,k, the loop is terminated since there exists no solution to satisfy the URLLC requirements in lines 18-20.Otherwise, the iteration is repeated based on line 2.

5 Performance Analysis

In this section, the performance of the DRO scheme is evaluated using MATLAB simulation,where the URLLC packet size was set toL= 32 bytes.The QoS parameters for URLLC packet transmission include the reliability requirement that was set to δ= 10-5, and the latency requirement which was set tod= 1 ms.Based on UE capability 2 and NR numerology μ= 2, the UE processing timeTUE_procwas set toN1= 0.161 ms and the gNB processing timeTgNB_procwas set toN2/2 = 0.098 ms [17].The packet alignment timeTDL_alignwas set to a 0～1 OFDM symbol duration since 7 MOs per slot was considered.In addition, an ideal backhaul is assumed and the ACK transmission timeTACK_Txwas set to 1 OFDM symbol duration, where the ACK alignment timeTACK_alignwas considered negligible.

In the simulation, NR gNBs form a hexagonal-shaped cellular coverage.The M-gNB is located at (-50, 0), and the S-gNB is located at (50), where the carrier frequency is set to 3.5 GHz.Since DC is activated in the cell edge regions and cell range extension (CRE) is considered, the UE is initially located at (-10, 15) and moves toward (20, 15).The performance of the DRO scheme was compared with LADMA [8] and SDUF [9] when the maximum transmission attemptsM=M1=M2is fixed to 2 and 3, respectively.A performance comparison of the reliability and the maximum latency performance (which are the two major URLLC QoS requirements) are presented in Figs.4 and 5,respectively.In addition, a performance comparison of the average channel use (which is an objective performance metric of the DRO scheme) is presented in Fig.6.

Figure 4: Simulation results with respect to the reliability performance

In Fig.4, the reliability of the DRO scheme is presented, where the black bold line represents the reliability condition (δ= 10-5).The distance of the horizontal-axis represents the travel distance of the UE from the initial position.The simulation results show that the DRO scheme can satisfy the reliability condition regardless of the distance of the UE movement, for the range of interest.Since the DRO scheme reduces the maximum transmission attempts adaptively to minimize the average channel use, the reliability performance is degraded when the distance is 2 and 20, but the satisfaction of the reliability condition is always guaranteed within the range of interest.In addition, if the maximum transmission attempts for the S-gNB is set to be larger than the M-gNB, the reliability of DRO increases as the UE moves further away from the M-gNB.While SDUF can satisfy the reliability condition for all distances regardless of the maximum transmission attempts, LADMA can satisfy the reliability condition only when the maximum transmission attempts are 3.Since LADMA relies on packet retransmissions of the M-gNB until the latency budget condition, the reliability of LADMA decreases as the UE moves further away from the M-gNB.

In Fig.5, the maximum latency and average latency of the DRO scheme is presented, where the black bold line in (a) represents the latency condition (d= 1).The simulation results show that the DRO scheme can satisfy the latency condition regardless of the distance of the UE movement.Since the DRO scheme can reduce the maximum transmission attempts when the distance is increased, the maximum latency is reduced.In addition, DRO shows the lowest average latency performance for the cell edge region compared to the other schemes by reducing redundant retransmissions from the unreliable M-gNB.In terms of the average latency, DRO can provide a 11.41% and 17.54% gain when compared to SDUF withM= 2 andM= 3, respectively, and can provide a 10.64% and 16.67% gain when compared to LADMA withM= 2 andM= 3, respectively.

Figure 5: Simulation results with respect to the latency performance

Figure 6: Simulation results with respect to the average channel use performance

In Fig.6, the average channel use of the DRO scheme is presented with respect to the distance of the UE movement.Based on the simulation results of Figs.4 and 5, the proposed DRO scheme can satisfy the URLLC QoS conditions for the entire range of interest.Furthermore, Fig.6 shows that the DRO scheme can provide the lowest average channel use compared to LADMA and SDUF for the entire range of interest.Under the feasible region of the URLLC QoS conditions, the DRO scheme can provide a 11.64% and 17.78% reduced average channel use when compared to SDUF withM= 2 andM= 3, respectively, and can provide a 12.44% and 17.77% reduced average channel use when compared to LADMA withM= 2 andM= 3, respectively.

6 Conclusion

In this paper, a PD aware HARQ retransmission scheme named DRO is proposed, which can minimize the average resource usage while satisfying the URLLC requirements in the NR-DC architecture.The proposed DRO scheme derives the optimal value of the maximum transmission attempts for the S-gNB is set to be larger than the M-gNB, the reliability of DRO scheme can provide a significant performance gain compared to the existing PD schemes, which include LADMA and SDUF.

Funding Statement:This work was supported by the Institute of Information & communications Technology Planning&Evaluation(IITP)grant funded by the South Korea government(MSIT,2021-0-00040, Development of intelligent stealth technology for information and communication resources for public affairs and missions).

Conflicts of Interest:The authors declare that they have no conflicts of interest to report regarding the present study.