Quasi-Dynamic Frame Coordination For Ultra- Reliability and Low-Latency in 5G TDD Systems

TL;DR

This paper introduces a quasi-dynamic frame coordination method for 5G TDD systems that significantly reduces latency and interference, enhancing ultra-reliable low-latency communications with minimal coordination overhead.

Contribution

It proposes a novel hybrid frame design and RFC code-book for adaptive inter-cell coordination, improving URLLC performance over existing dynamic TDD approaches.

Findings

92% reduction in URLLC outage latency

189% increase in cell-edge capacity

Reduced coordination overhead to B-bit

Abstract

The fifth generation (5G) mobile technology features the ultra-reliable and low-latency communications (URLLC) as a major service class. URLLC applications demand a tight radio latency with extreme link reliability. In 5G dynamic time division duplexing (TDD) systems, URLLC requirements become further challenging to achieve due to the severe and fast-varying cross link interference (CLI) and the switching time of the radio frame configurations (RFCs). In this work, we propose a quasi-dynamic inter-cell frame coordination algorithm using hybrid frame design and a cyclic-offset-based RFC code-book. The proposed solution adaptively updates the RFCs in time such that both the average CLI and the user-centric radio latency are minimized. Compared to state-of-the-art dynamic TDD studies, the proposed scheme shows a significant improvement in the URLLC outage latency, i.e., 92% reduction gain, while boosting the cell-edge capacity by 189% and with a greatly reduced coordination overhead space, limited to B-bit.

Tables

Table 1. Table I: Simulation parameters.

Parameter	Value
Environment	3GPP-UMA, one cluster, 21 cells
UL/DL channel bandwidth	10 MHz, SCS = 30 KHz, TDD
Antenna setup	$N_t=8$ Tx, $M_r=2$ Rx
UL power control	LTE-alike, $\alpha =1,P0=-103$ dBm
Average user load per cell	$K^{\text{dl}}=K^{\text{ul}}=$ 10 and 20
TTI configuration	0.5 ms (7-OFDM symbols)
Traffic model	$\begin{array}{c}\text{FTP3, }{f}^{dl}\text{ = }{f}^{ul}\text{ = 400 bits}\\ {\lambda }^{\text{dl}}\text{ =167, and 620 pkts/sec}\\ {\lambda }^{\text{ul}}\text{ =334, and 620 pkts/sec}\end{array}$
$\begin{array}{c}\text{Offered average load per cell}\\ \text{DL:UL}\end{array}$	$\begin{array}{c}\text{ DL:UL = 1:2 (0.6:1.2) Mbps}\\ \text{DL:UL = 1:1 (5:5) Mbps}\end{array}$
Proposed HFCS setup	$\begin{array}{c}𝒩\text{ = 55}\text{ RFCs}\\ L\text{ = 7}\text{ groups}\\ \text{ B = 6 bits}\end{array}$

Equations

$${\text{y}}_{k,c_k}^{\text{dl}}=\underset{\text{Useful signal}}{\underbrace{{\text{H}}_{k,c_k}^{\text{dl}}{\text{v}}_k{s}_k}}+\underset{\text{BS to UE interference}}{\underbrace{\sum _{i\in {\text{K}}_{\text{dl}}\backslash k}{\text{H}}_{k,c_i}^{\text{dl}}{\text{v}}_i{s}_i}}+\underset{\text{UE to UE interference}}{\underbrace{\sum _{j\in {\text{K}}_{\text{ul}}}{\text{G}}_{k,j}{\text{w}}_j{s}_j}}+{\text{n}}_k^{\text{dl}},$$

$${\text{y}}_{c_k,k}^{\text{ul}}=\underset{\text{Useful signal}}{\underbrace{{\text{H}}_{c_k,k}^{\text{ul}}{\text{w}}_k{s}_k}}+\underset{\text{UE to BS interference}}{\underbrace{\sum _{j\in {\text{K}}_{\text{ul}}\backslash k}{\text{H}}_{c_k,j}^{\text{ul}}{\text{w}}_j{s}_j}}+\underset{\text{BS to BS interference}}{\underbrace{\sum _{i\in {\text{K}}_{\text{dl}}}{\text{Q}}_{c_k,c_i}{\text{v}}_i{s}_i}}+{\text{n}}_{c_k}^{\text{ul}},$$

$${\gamma }_k^{\text{dl}}=\frac{p_{c_k}^{\text{dl}}{\Vert {\text{H}}_{k,c_k}^{\text{dl}}{\text{v}}_k\Vert }^2}{{\sigma }^2+\sum _{i\in {\text{K}}_{\text{dl}}\backslash k}{p}_{c_i}^{\text{dl}}{\Vert {\text{H}}_{k,c_i}^{\text{dl}}{\text{v}}_i\Vert }^2+\sum _{j\in {\text{K}}_{\text{ul}}}{p}_j^{\text{ul}}{\Vert {\text{G}}_{k,j}{\text{w}}_j\Vert }^2},$$

$${\gamma }_{c_k}^{\text{ul}}=s\frac{p_k^{\text{ul}}{\Vert {\text{H}}_{c_k,k}^{\text{ul}}{\text{w}}_k\Vert }^2}{{\sigma }^2+\sum _{j\in {\text{K}}_{\text{ul}}\backslash k}{p}_j^{\text{ul}}{\Vert {\text{H}}_{c_k,j}^{\text{ul}}{\text{w}}_j\Vert }^2+\sum _{i\in {\text{K}}_{\text{dl}}}{p}_{c_i}^{\text{dl}}{\Vert {\text{Q}}_{c_k,c_i}{\text{v}}_i\Vert }^2},$$

$$\left\{\begin{array}[]{c}\underset{c}{\arg\min}\,\phi_{c}\left(\eta_{c}\right)=\frac{1}{C}\stackrel{{\scriptstyle[}}{{x}}=1,x\neq c]{C}{\sum}\varphi_{c,x}\left(\eta_{c},\eta_{x}\right),\\ \text{$\forall$}k\in\text{$\mathcal{K}$}_{\textnormal{ul}/\textnormal{dl}}:\underset{k}{\arg\min}\left(\varPsi_{c,k}\right),\,\varPsi_{c,k}\leq\epsilon\,\textnormal{ms},\end{array}\right.$$

$${\beta }_c\le \frac{\sum Z_c^{\text{dl}}}{\sum Z_c^{\text{dl}}+\sum Z_c^{\text{ul}}},$$

$${\Xi }_c^{\text{avg.}}=\frac{{\tilde{\beta }}_c\times {\Theta }_c^{\text{BS}}+\widetilde{{\mu }_c}\times {\Theta }_c^{\text{UE}}}{{\Theta }_c^{\text{BS}}+{\Theta }_c^{\text{UE}}},$$

$$\tilde{\beta}_{c},\,\tilde{\mu_{c}}=\left\{\begin{array}[]{c}\frac{1}{\beta_{c}},\frac{1}{\mu_{c}}\,\,\,\forall\,\Theta_{c}^{\textnormal{BS}},\Theta_{c}^{\mathit{\textnormal{UE}}}\leq\varrho\,\textnormal{dBm}\\ \beta_{c},\mu_{c}\,\,\,\forall\,\Theta_{c}^{\textnormal{BS}},\Theta_{c}^{\mathit{\textnormal{UE}}}>\varrho\,\textnormal{dBm}\end{array}\right.,$$

$${\beta }_c=\frac{\sum Z_c^{\text{dl}}}{\sum Z_c^{\text{ul}}}\times {\mu }_c,$$

$${\phi }_{c,x}(t)\le {\psi }_c(t).$$

$$\omega \left\{{\text{PF}}_{k_{\text{ul/dl}}}\right\}=\frac{r_{{}_{k_{\text{ul/dl}},rb}}}{{\overline{r}}_{k_{\text{ul/dl}},rb}},$$

$$k_{\text{ul/dl}}^{\ast }=\underset{k_{\text{ul/dl}}\in {\text{K}}_{\text{ul/dl}}}{\mathrm{argmax}}\omega \left\{{\text{PF}}_{k_{\text{ul/dl}}}\right\},$$

Full text

Index Terms— Dynamic TDD; 5G new radio; URLLC; Cross link interference (CLI); Traffic; UDP.

I Introduction

Ultra-reliable low-latency communication (URLLC) is a key driver of the fifth generation (5G) mobile networks [1]. Various URLLC use cases require one-way radio latency of one or several milliseconds with an outage probability below $10^{-5}$ [2]. As most of the 5G URLLC deployments are envisioned over the 3.5 GHz band, the time division duplexing (TDD) becomes a vital candidate transmission mode due to its frame adaptation, in order to dynamically match the sporadic URLLC capacity in both downlink (DL) and uplink (UL) directions [3].

With the 5G new radio (NR), the agile frame structure with variable transmission time interval (TTI) duration is introduced [3, 4]. Thus, 5G-NR TDD offers more adaptation flexibility with much faster link-direction update periodicity, that is slot-dependent instead of being frame-based, i.e., $\leq 1$ ms. However, the coexistence of different transmission directions in adjacent cells results in cross link interference (CLI) [5], i.e., base-station to base-station (BS-BS) and user-equipment to user-equipment (UE-UE) CLI, respectively. Hence, URLLC performance is highly impacted by the degraded decoding ability, due to the fast-varying CLI, and the waiting interval to the first DL/UL transmission opportunity.

To the best of our knowledge, no prior work has assessed the performance of the URLLC outage with the 5G-NR dynamic TDD technology. The state-of-the-art TDD proposals consider joint multi-cell scheduling, cell muting, and enhanced power control [6, 7] to minimize the average network CLI. Furthermore, advanced massive multi-antenna processing and beam-forming [8] are envisioned as vital to counteract the CLI by utilizing the channel hardening phenomenon. Opportunistic inter-cell coordination algorithms [9] are also proven attractive to boost the cell capacity of the dynamic TDD systems; however, at the expense of a sub-optimal URLLC outage performance.

In this work, we propose a hybrid-frame based coordination scheme (HFCS) for 5G-NR dynamic TDD systems. The proposed HFCS introduces a multi-objective and slot-dependent dynamic user scheduling. A hybrid radio frame structure and sliding radio frame configuration (RFC) code-book are designed to virtually extend the degrees of freedom of the TDD dynamicity. Thus, the URLLC users with the worst radio conditions always guarantee semi-preemptive, i.e., immediate scheduling over pre-set time slots, and CLI-free transmissions, leading to a significant reduction of the URLLC tail latency. The proposed coordination scheme shows a significant enhancement in the URLLC outage performance as well as maximizing the ergodic capacity, and with a confined coordination overhead span.

The performance of the proposed scheme is assessed by realistic system level simulations, due to the complexity of the 5G-NR and addressed problem herein. The major functionalities of the physical and media access control layers of the 5G-NR are incorporated and calibrated against latest 3GPP assumptions, including UL and DL channel modeling, hybrid automatic repeat request (HARQ), adaptive modulation and coding selection (MCS) and dynamic user scheduling.

This paper is organized as follows. Section II introduces the system modeling of this work while Section III presents the problem formulation. Section IV details the proposed solution and Section V discusses the numerical results of the proposed scheme. Finally, conclusions are drawn in Section VII.

II System Modeling

A macro 5G-NR TDD system is considered, with a single cluster of $C$ cells, each with $N_{t}$ antennas. Each cell has an average of $K^{\textnormal{dl}}$ and $K^{\textnormal{ul}}$ uniformly-distributed DL and UL active UEs, respectively, each with $M_{r}$ antennas. We assume a URLLC dedicated network where the sporadic FTP3 traffic is adopted with finite packet sizes of $\mathit{f}^{dl}$ and $\mathit{f}^{ul}$ bits, and Poisson arrival processes \textnormal{\lambda}^{\textnormal{dl}} and \textnormal{\lambda}^{\textnormal{ul}}, in the DL and UL directions. Accordingly, the average offered load per cell in DL direction is: K^{\textnormal{dl}}\times\textnormal{\mathit{f}^{dl}}\times\textnormal{\lambda}^{\textnormal{dl}} and in UL direction as: K^{\textnormal{ul}}\times\textnormal{\mathit{f}^{ul}}\times\textnormal{\lambda}^{\textnormal{ul}}.

We assume an RFC of 10 sub-frames, each can be DL, UL or a special sub-frame. UEs are dynamically multiplexed by the orthogonal frequency division multiple access with 15 kHz sub-carrier spacing. The smallest scheduling unit is the physical resource block (PRB) of 12 consecutive sub-carriers. Furthermore, we adopt a user scheduling per a mini-slot duration of 7-OFDM symbols for faster URLLC transmissions.

Furthermore, an arbitrary master cell is initially identified in each cluster, where other cells are considered as slaves. All slave cells within the cluster are bidirectionally inter-connected to the master cell through the Xn interface.

We define \text{\mathfrak{B}}_{\textnormal{dl}}, \text{\mathfrak{B}}_{\textnormal{ul}}, \text{\mathcal{K}}_{\textnormal{dl}} and \text{\mathcal{K}}_{\textnormal{ul}} as the inclusive sets of cells and UE with DL and UL transmission directions, respectively. Hence, the pre-decoding received signal at the $k^{th}$ UE, where k\text{\in\text{$\mathcal{K}$}{\textnormal{dl}}}, c_{k}\text{\in\text{$\mathfrak{B}$}{\textnormal{dl}}}, is expressed by

[TABLE]

where \boldsymbol{\textnormal{{H}}}_{k,c_{i}}^{\textnormal{dl}}\in\text{\mathcal{C}}^{M_{r}\times N_{t}} is the DL fading channel from the cell serving the $i^{th}$ UE, to the $k^{th}$ UE, \boldsymbol{\textnormal{{v}}}_{k}\in\text{\mathcal{C}}^{N_{t}\times 1} , \boldsymbol{\textnormal{{w}}}_{j}\in\text{\mathcal{C}}^{M_{r}\times 1} and $s_{k}$ denote the single-stream zero-forcing precoding vector at the $c_{k}^{th}$ cell, precoding vector at the $j^{th}$ UE, and transmitted data symbol of the $k^{th}$ UE, respectively, \boldsymbol{\textnormal{{G}}}_{k,j}\in\text{\mathcal{C}}^{M_{r}\times M_{r}} represents the the cross-link channel between the $k^{th}$ and $j^{th}$ UEs. $\boldsymbol{\textnormal{{n}}}_{k}^{\textnormal{dl}}$ denotes the additive white Gaussian noise at the $k^{th}$ UE. In the UL direction, the received signal at the $c_{k}^{th}$ cell, where c_{k}\text{\in\text{$\mathfrak{B}$}{\textnormal{ul}}} from k\text{\in\text{$\mathcal{K}$}{\textnormal{ul}}}, is modeled by

[TABLE]

where \boldsymbol{\textnormal{{Q}}}_{c_{k},c_{i}}\in\text{\mathcal{C}}^{N_{t}\times N_{t}} denotes the cross-link fading channel between the cells that serve the $k^{th}$ and $i^{th}$ UEs, respectively, k\text{\in\text{$\mathcal{K}$}_{\textnormal{ul}}} and i\in\text{\mathcal{K}}_{\textnormal{dl}}. The pre-detection signal-to-interference-noise-ratio (SINR) in the DL direction at the $k^{th}$ UE $\gamma_{k}^{\textnormal{dl}}$ and in the UL direction at the $c_{k}^{th}$ cell $\gamma_{c_{k}}^{\textnormal{ul}}$, are given by

[TABLE]

[TABLE]

where $p_{c_{k}}^{\textnormal{dl}}$ and $p_{k}^{\textnormal{ul}}$ denote the transmission powers of the $c_{k}^{th}$ cell and the $k^{th}$ UE, respectively. Finally, the received UL/DL signals are decoded by the linear minimum mean square error interference rejection combining receiver (LMMSE-IRC) [4] vector a, expressed as: $\hat{s}_{k}^{\kappa}=\left(\boldsymbol{\textnormal{{a}}}_{k}^{\kappa}\right)^{\textnormal{H}}\boldsymbol{\textnormal{y}}_{k}^{\kappa},$ \text{\mathcal{X}}^{\kappa},\kappa\text{\in}\{\textnormal{ul},\textnormal{dl}\}, with $\left(\bullet\right)^{\textnormal{H}}$ as the Hermitian operation.

III Problem Formulation

The URLLC latency and reliability requirements are further challenging to achieve in 5G-NR dynamic TDD systems, mainly due to the link-direction switching time and the degraded URLLC decoding performance. The former is significantly minimized by the flexible 5G frame structure; however, the latter still remains an open issue.

In fully dynamic TDD macro networks, neighboring cells may have simultaneous cross-directional transmissions, leading to a strong CLI which varies per the link-direction update periodicity. With the 5G-NR, such periodicity is slot-based, i.e., $\leq 1$ ms, leading to highly varying CLI fluctuations. As a result, URLLC UEs inflict significantly degraded decoding performance. In particular, lower-power URLLC UL transmissions suffer from a strong CLI from adjacent higher-power DL transmissions, leading to several HARQ re-transmissions prior to a successful decoding, not satisfying the URLLC targets.

Let $u_{c}$ and $d_{c}$ present the estimated numbers of UL and DL slots during a given RFC while $u_{c}^{\textnormal{opt.}}$ and $d_{c}^{\textnormal{opt.}}$ are the respective optimal numbers. Hence, the proposed HFCS defines a programming optimization problem as:

R\triangleq\underset{c}{\arg\max}\stackrel{{\scriptstyle[}}{{c}}=1]{C}{\sum}\min\left(\textnormal{u_{c}$,$u_{c}^{\textnormal{opt.}}}\right)\digamma_{c}^{u}+\min\left(\textnormal{d_{c}$,$d_{c}^{\textnormal{opt.}}}\right)\digamma_{c}^{d},

subject to:

[TABLE]

where $R$ is total capacity of each cluster, $\digamma_{c}^{u}$ and $\digamma_{c}^{d}$ denote rate utility functions of the UL and DL transmissions, i.e., capacity gain due to an UL or DL transmission. $\phi_{c}\left(\eta_{c}\right)$ and $\varphi_{c,x}\left(\eta_{c},\eta_{x}\right)$ represent the average and actual slot misalignment of the requested RFC by the $c^{th}$ cell $\eta_{c}$ and between the RFCs of the $c^{th}$ and $x^{th}$ cells, i.e., $\eta_{c}$ and $\eta_{x}$, respectively, $\forall x\neq c$, and $\varPsi_{c,k}$ is the one-way radio latency of the $k^{th}$ UL or DL user which is confined by $\epsilon$ ms.

For best RFC adaptation and highest ergodic capacity, $u_{c}=u_{c}^{\textnormal{opt.}}$ and $d_{c}=d_{c}^{\textnormal{opt.}}$ should be arbitrarily set in (5). However, $u_{c}^{\textnormal{opt.}}$ and $d_{c}^{\textnormal{opt.}}$ may introduce a large inter-cell slot misalignment $\phi_{c}$, resulting in severe CLI within the cluster, and thus, a significant degradation of the overall capacity $R$ and URLLC latency performance. As such problem is non-convex, we propose a heuristic approach using complexity-efficient coordination with hybrid-frame design, multi-objective user scheduling and a sliding-based RFC code-book.

IV Proposed HFCS Coordination

The proposed HFCS combines a hybrid RFC design, multi-objective distributed user scheduling, and a cyclic-offset-based RFC code-book. A pre-defined RFC code-book is constructed and presumed pre-known to all cells within the cluster, where all RFCs have a set combination of static and dynamic slots. At each RFC update instant, each slave cell selects the one RFC from the code-book that most satisfies its individual link-direction selection criterion. The slave cells signal the index of the selected RFC to the master cell where it seeks to improve the joint capacity. Thus, it may slightly change the RFCs requested by slave cells. Accordingly, the master cell feeds-back the updated RFC indices to the slave cells, to be adopted until the next RFC update. During each RFC period, each cell considers a dual-objective dynamic user scheduling.

IV-A Proposed Inter-Cell Coordination Scheme

Hybrid RFC design and sliding RFC code-book

A hybrid RFC design is adopted, where each RFC is divided into arbitrary static and dynamic slot sets (SSS, DSS). A SSS denotes the radio slots which are fixed across all RFCs in the code-book, i.e., static TDD slots with CLI-free transmissions. However, a DSS implies fully dynamic radio slots.

Accordingly, a pre-defined RFC code-book of $\mathcal{N}$ unique RFCs is constructed such that it is divided into $L$ groups. The RFCs within each group share the same DL:UL slot ratio, i.e., $d_{c}:u_{c}$; though, with a different placement during the DSS. For instance, the DSS of each RFC is a cyclic-shift of the other RFCs, as depicted in Fig. 1. The structure of the SSS, DSS, and size of the RFC code-book are design parameters.

At slave cells – Traffic and latency adaptation

During each RFC update instant, each slave cell selects the one RFC from the code-book which best satisfies its link-direction selection criterion. Without loss of generality, we consider the DL/UL buffered traffic size including pending HARQ re-transmissions as the major criterion to select the RFC, and accordingly the best $d_{c}:u_{c}$ ratio. Then, the traffic load threshold $\beta_{c}$ is defined as

[TABLE]

where $\sum Z_{c}^{\textnormal{dl}}$ and $\sum Z_{c}^{\textnormal{ul}}$ imply the aggregate traffic in the DL and UL directions, respectively. With $\beta_{c}=0.5$, if $\sum Z_{c}^{\textnormal{dl}}\gg\sum Z_{c}^{\textnormal{ul}}$, a cell selects an RFC with a majority of DSS DL slots. Although, the free selection of the best RFCs requested by each slave cell may result in severe CLI, hence, several HARQ re-transmissions may be inflicted, leading to a significant radio latency and reliability. On the other side, an abrupt change of these RFCs to reduce the average CLI leads to significant queuing delays up to the first DL/UL transmission opportunities. Thus, to address the constraints in (5), each cell adaptively estimates a dynamic sliding threshold $\psi_{c}(t)$, where $t$ is the link-direction update time, with which it instructs the master cell about the maximum allowable change of its desired RFC, in order to achieve an adequate joint URLLC and ergodic capacity performance.

Let $\Theta^{\textnormal{BS}}$ and $\Theta^{\mathit{\textnormal{UE}}}$ denote the BS-BS and UE-UE CLI at the BS and UE, respectively. These CLI estimates can be obtained at the BS through radio feedback links from UEs; however, there is no a standardized mechanism of the CLI measurement reporting available yet. Then, each BS calculates the average experienced CLI using an arbitrary filter function. In this work, we assume a weighted average filter as

[TABLE]

[TABLE]

[TABLE]

where $\beta_{c}$ and $\mu_{c}$ are the BS-BS and UE-UE CLI weights, and $\varrho$ is a CLI threshold. The ratio of both weights is set to the ratio of the buffered traffic as in (9), such that a cell with $\sum Z_{c}^{\textnormal{dl}}\gg\sum Z_{c}^{\textnormal{ul}}$, and accordingly a DL-heavy RFC, shall impose severe BS-BS CLI to adjacent cells. Hence, under this condition, $\Xi_{c}^{\textnormal{avg.}}$ is biasedly maximized and the RFC adaptation is enforced towards the CLI minimization. In the delay domain, cells measure the head of line delay (HoLD) within their DL and UL transmission buffers. HoLD indicates an estimate of the maximum time required to transmit the last packet in the buffer, based on the expected UL/DL transmission constraints of the current RFC. Such metric is of a significant importance with URLLC since a packet can be considered of no use if its latency deadline is not fulfilled. Hence, to reduce the average HoLD, selected traffic-based RFCs should be used without a significant change in order to quickly transmit the data buffers.

Thus, we propose a simple and dynamic sliding threshold for a best-effort trade-off between CLI and radio latency. Fig. 2 shows a numerical example of such approach. A SSS to DSS ratio of 8:12 is assumed. Thus, the slot misalignment threshold is bounded by the size of the DSS. Accordingly, the range of the CLI and HoLD values is quantized over the DSS size. For an arbitrary cell, if the average CLI, experienced over the previous measurement cycle, is at maximum, e.g., $\Xi_{c}^{\textnormal{avg.}}=-60$ dBm, it implies a tight slot misalignment threshold should be enforced to promptly reduce such severe CLI over the upcoming RFC period, e.g., $\psi_{c}(t)=1$ slot. However, if such cell simultaneously inflicts a large HoLD, e.g., HoLD = 52 ms, the slot misalignment constraint shall be relaxed, e.g., $\psi_{c}(t)=10$ slots, in order not to allow the master cell to change the RFC of this cell, hence, having faster transmissions for the respective traffic. Without loss of generality, we apply a fair averaging of both misalignment thresholds, i.e., $\psi_{c}(t)=5$ slots.

Finally, at each RFC update periodicity, slave cells signal the master cell with the requested RFC indices of \textnormal{B}=\log_{2}\left(\text{\mathcal{N}}\right) bits on the *Xn interface *as well as the maximum allowable slot misalignment thresholds $\psi_{c}(t).$

**At master cell – CLI minimization **

When the master cell receives all RFC information from slave cells, it first identifies a *common RFC, *which is requested by the majority of the slave cells. If not feasible, the master cell randomly selects any reported RFC as the common one, to which all other RFCs shall maintain the respective slot misalignment thresholds. Thus, for each RFC $\eta_{c}$ of the $c_{k}^{th}$ cell, master cell calculates the slot misalignment to the common RFC $\delta_{x},\forall x\neq c$ as in (5). Then, the master cell does not alter such requested RFC if the following condition is fulfilled:

[TABLE]

Hence, the respective slave cell utilizes its best matching RFC to its latency and capacity outage. Otherwise, the master cell slides over all RFCs within the same group as the desired RFC of the $c^{th}$ cell $\eta_{c}$. Accordingly, it estimates the corresponding slot misalignment values and considers the one RFC with $\varphi_{c,x}(t)$ that has the closest linear distance to the requested $\psi_{c}(t).$ If the slot misalignment constraint in (10) is satisfied, master cell adopts such RFC as the updated RFC of the current cell. This way, an acceptable average CLI is guaranteed at the slave cells while still preserving the same requested traffic service ratio $d_{c}:u_{c}$, leading to a significant improvement of the capacity and outage latency performance.

If the slot misalignment constraint is not yet feasible across all RFCs from the same group as the requested one, master cell progressively slides to the other RFC groups from the RFC code-book with the nearest possible $d_{c}:u_{c}$ ratio to the requested ratio, e.g., $d_{c}:u_{c}=4:12\underset{\textnormal{slide\,to}}{\rightarrow}d_{c}^{{}^{\prime}}:u_{c}^{{}^{\prime}}=3:13$, and repeats the same process. Herein, the master cell partly relaxes the target outage requirements of the salve cells due to the abrupt change in the $d_{c}:u_{c}$ ratio. However, such outage degradation is bounded across a limited number of slots during the RFC and is reversely proportional to the size of the RFC code-book $\mathcal{N}$. As a last best-effort resort, if the constraint in (10) could not be satisfied across all RFCs, either from same or different group(s), the master cell considers the one RFC with the closest possible estimated slot misalignment to desired $\psi_{c}(t)$, and then, it signals all slave cells within the cluster over the* Xn interface* with the updated RFC indices that should be used over the upcoming RFC periodicity.

IV-B *Distributed multi-objective user scheduling *

During each RFC periodicity, each cell applies a slot-dependent dynamic user scheduling. During the DSS instances, cells may adapt an arbitrarily capacity maximizing user scheduling. Without loss of generality, and since we assume an equally-prioritized URLLC setup, we adopt the proportional fair (PF) criterion $\omega$ in both the time and frequency domains to maintain a global scheduling fairness as

[TABLE]

[TABLE]

where $r_{{}_{k_{\textnormal{ul/dl}},rb}}$ and $\overline{r}_{k_{\textnormal{ul/dl}},rb}$ denote the instantaneous and average delivered rates of the $k^{th}$ UL/DL user. However, during the SSS periods, each cell preemptively interrupts its individual time-domain scheduling metric by immediately allocating the users with the worst radio conditions, i.e., potentially cell-edge users. These users are identified based on the reported channel quality indication (CQI) reports. To avoid threshold-based user identification, the UL/DL time-domain scheduler sorts active users in an ascending-order list in terms of their reported CQI levels, i.e., users from the top of the list are of worst radio conditions, thus, scheduler grants them a higher priority for immediate scheduling during the CLI-free SSS. In the frequency domain, the PF metric is used to preserve fairness among cell-edge URLLC users. Thus, cell-edge URLLC users achieve a better decoding ability with faster transmissions, avoiding the latency-costly HARQ re-transmissions.

IV-C Comparison to the state-of-the-art TDD studies

We evaluate the performance of the proposed scheme against the state-of-the-art coordinated TDD proposals as:

**Non-coordinated TDD **(NC-TDD): no RFC coordination is assumed. Cells independently and dynamically in time pick the RFCs from the code-book which most meet their individual traffic demand, as in (6). Hence, maximum TDD RFC flexibility is achieved with no coordination overhead; however, associated with potentially a large slot misalignment and severe average CLI levels accordingly.

Sliding code-book based coordinated TDD (SCC-TDD) [9]: in our prior work, we introduced a simple inter-cell coordination algorithm, mainly for broadband services, to significantly reduce the average slot misalignment, based on a preset global misalignment threshold $\varOmega$, and hence, the aggregate CLI, resulting in greatly improved ergodic capacity. Though, it has been demonstrated not suitable for URLLC transmissions due to the monotonic scheduling objective.

CLI-free coordinated TDD (CFC-TDD): cells dynamically select their respective RFC according to (6). A sophisticated BS-BS and UE-UE coordination is artificially assumed. That is, BSs and UEs exchange PRB mapping, UE MCS and precoding information, for them to perfectly suppress the BS-BS and UE-UE CLI. However, such coordination introduces a significant control overhead over both the back-haul and radio interfaces, respectively. In [10], a 3GPP technical study introduces a sub-optimal CFC-TDD approach with a lower overhead space. However, CFC-TDD holds an optimal theoretical baseline, where both maximum TDD RFC flexibility and CLI-free transmissions are always guaranteed.

V Performance Evaluation

The major simulation assumptions are presented in Table I. During each TTI, each cell dynamically multiplexes users over system PRBs using the PF metric, if it is within the DSS of the current RFC or by preemptive cell-edge user allocations when it is within the SSS. We consider a fully dynamic MCS selection and adaptive Chase-combining HARQ re-transmissions, where the HARQ feedback is always prioritized over new transmissions. The post-detection SINR levels are estimated by the LMMSE-IRC receiver, where the average interference is identified by its mean covariance. Finally, we assess the proposed solution under the latency-efficient user data-gram protocol for several offered cell loading conditions.

Fig. 3 depicts a comparison of the complementary cumulative distribution function (CCDF) of the URLLC outage latency in the UL direction for all TDD coordination schemes under assessment, for an average offered load of 2 Mbps/cell with a DL:UL traffic ratio of 1:2. Furthermore, we present the latency performance of the best static-TDD case where the static pattern is pre-selected to perfectly match the DL-to-UL average traffic ratio, i.e., 6 DL mini-slots, 12 UL mini-slots and 2 guard mini-slots. The optimal CFC-TDD achieves the best URLLC latency performance, i.e., 42 ms at $10^{-5}$ outage probability. However, it comes under the ideal assumption of perfect elimination of any experienced CLI, and with an infinite coordination overhead, which is infeasible in practice. The proposed HFCS clearly provides a significant improvement of the UL URLLC latency, approaching the optimal CFC-TDD; however, with greatly reduced overhead span, mainly limited to \log_{2}\left(\text{\mathcal{N}}\right) bits. That is, it achieves $92\%$ and $67\%$ reduction gain in the UL outage latency compared to SCC-TDD and NC-TDD. The best static-TDD case out-performs proposed HFCS scheme, i.e., $9\%$ reduction in the outage latency, due to the absence of the CLI, approaching CFC-TDD; though, this comes with the assumption that the static RFC pattern is pre-defined to perfectly align with the traffic demands.

The significant latency improvements of the proposed HFCS are attributed to the guaranteed preemptive cell-edge user scheduling with CLI-free transmissions, where these users majorly control the latency tail, i.e., outage, performance. Thus, less costly HARQ re-transmissions are experienced. The SCC-TDD latency performance depends on the preset misalignment threshold $\varOmega.$ For instance, with a tight $\varOmega=3$, the master cell may aggressively change the requested RFC of a given slave cell, in order to only allow for an average misalignment of three slots. As a result, slave cells may adopt RFCs that do not best match their current traffic demands, leading to a more queuing delay to the first transmission opportunity. Finally, the NC-TDD offers a fair URLLC latency performance since the maximum possible TDD RFC flexibility is utilized; however, with severe CLI levels.

Similar observations are obtained from the URLLC outage latency in the DL direction, as shown in Fig. 4. All considered TDD coordination schemes provide a decent DL latency , i.e., $\leq 8$ ms. This is due to the larger desired DL transmission power, i.e., compared to the interfering UL power, hence, less impactful CLI. However, static-TDD case inflicts a longer queuing delay due to the fixed DL and UL slot placement.

Fig. 5 shows the empirical CDF (ECDF) of the post-receiver UL interference performance in dBm, including both cross and same link inter-cell interference, respectively. Due to the absence of the CLI, CFC-TDD offers an attractive interference performance. However, due to the dual-scheduling metrics during the DSS and SSS periods, the proposed HFCS achieves the same interference suppression capability as the optimal CFC-TDD for the critical lower percentiles below $20\%$, i.e., cell-edge users. Furthermore, the proposed HFCS offers $39\%$ and $45\%$ reduction of the post-receiver interference at the $20^{th}$ percentile, compared to SCC-TDD and NC-TDD, respectively. The SCC-TDD exhibits a monotonic interference suppression performance where cell-edge users, get most impacted, while NC-TDD inflicts the worst interference performance due to the extreme slot misalignment, hence, the sever CLI levels.

Fig. 6 presents the average cell throughput per TTI in the UL direction, with an average total offered load per cell of 10 Mbps. As can be noted, proposed solution boosts the cell-edge capacity, e.g., $189\%$ capacity gain is achieved against SCC-TDD at the $30^{th}$ percentile. The change of the distribution slope of the proposed HFCS is due to the slot-based dual scheduling objectives, i.e., joint latency-capacity scheduling. However, the proposed HFCS still exhibits a capacity loss of $45\%$ at the $95^{th}$ percentile compared to ideal SCC-TDD, due to the preemptive scheduling of cell-edge users during the SSS of each RFC, despite that they may not be the best capacity/fairness maximizing set of users. The fully dynamic NC-TDD fails to offer an acceptable cell-edge capacity due to the extreme CLI, i.e., $\sim 48\%$ of the scheduling TTI instances have no sufficient capacity.

VI Concluding Remarks

A quasi-dynamic coordination scheme has been introduced for ultra-reliable and low-latency communications (URLLC) in 5G TDD networks. The proposed solution combines hybrid radio frame design, distributed multi-objective user scheduling and a cyclic-offset-based radio frame code-book. Compared to the state-of-the-art coordinated TDD proposals from industry and academia, proposed scheme offers a significant improvement of the URLLC outage performance, e.g., $92\%$ latency reduction gain, in addition to achieving aggregated cell capacity gain of $189\%$, and with a limited control overhead space, bounded to $\textnormal{B-bit}.$

VII Acknowledgments

This work is partly funded by the Innovation Fund Denmark – File: 7038-00009B. Also, part of this work has been performed in the framework of the Horizon 2020 project ONE5G (ICT-760809) receiving funds from the European Union.