System-Level Analysis of Full-Duplex Self-Backhauled Millimeter Wave NetworksThanks: The authors are with the Wireless Networking and Communications Group (WNCG), The University of Texas at Austin, Austin, TX. Email: {g.manan@utexas.edu, ipr@utexas.edu, jandrews@ece.utexas.edu}. Last modified: August 24, 2026.

Multihop networks have been an active area of research for a few decades [13, 14, 15], as they require fewer fiber connections to tessellate an area according to the given coverage criterion, for example, received signal-to-noise ratio (SNR) above a threshold . However, per-user throughput deteriorates and packet delays increase with the number of hops between the donor and a user-equipment (UE) [14, 16]. Even though IAB networks operating at mmWave frequencies benefit from larger bandwidth (offering high data rates on the backhaul) and reduced interference (from directional communication and raised integrated noise power), they are subject to the same fundamental throughput-coverage trade-off as conventional multihop networks. As the number of hops increase, the throughput and latency performance degrade due to packet relaying, buffering, and link multiplexing delays [6, 17].

As a result, to satisfy throughput and latency targets, efficient resource utilization at the IAB node is of utmost importance and has led to a variety of studies on route selection [9, 18, 19], link scheduling [20, 8], load balancing [21], and topology optimization [22, 11]. In [9, 8], authors propose reinforcement learning frameworks that aim to minimize end-to-end latency of packets, and [18] presents routing strategies that minimize the number of hops to improve throughput and latency. Link scheduling and power allocation solutions that leverage simulated-annealing are proposed in [20], the benefits of offloading UEs to IAB nodes is studied in [21], and genetic algorithm based schemes for IAB node placement and non-IAB link distribution are developed in [22]. For a comprehensive survey on recent developments in IAB, please see [23].

In this work however, we investigate how FD capability can alleviate the aforementioned resource bottleneck and latency issues present in IAB networks. Recent breakthroughs in self-interference cancellation using analog [24], digital [25], and spatial [26, 27] cancellation techniques can rid a receive signal of self-interference. In [28], the authors prototype a two-hop IAB network and show that the throughput for FD-IAB is almost twice that of HD-IAB. In [25], the authors evaluate a multi-user FD-IAB network and present user selection and digital self-interference cancellation techniques to maximize the received signal power at the user. These works, however, do not explore the network-level consequences of mmWave FD in an IAB deployment. In this paper, we aim to quantify the relative gain of FD-IAB over HD-IAB in terms of the UE throughput, network depth, and latency in realistic multihop IAB deployments.

Our technical contributions are summarized as follows.

A single optimization framework for HD-IAB and FD-IAB networks. We formulate an optimization problem to study latency and throughput—arguably the most important performance metrics—in a mmWave IAB network. In Section II, we model the IAB network as a Jackson network of queues, which allows us to leverage results from queueing theory to model packet delay as the sojourn time and throughput as the packet arrival rate. We use this queueing model to formulate a utility maximization with throughput and latency constraints that are both practical and tractable. Our network optimization problem is a convex program and is parameterized by the routing structure of the network, the link capacities, and scheduling restrictions on each BS, which in turn depend on whether the BS is capable of HD or FD communication. Solving the convex program returns the per-user throughput and the corresponding resource allocation scheme that maximizes a desired network utility.

Characterizing the network gain of FD-IAB over HD-IAB. In Section IV, we use the optimization framework to compare FD-IAB with HD-IAB and derive a closed-form expression for latency gain. We also derive the maximum number of hops that achieve a target latency and target throughput for each UE, for both FD-IAB and HD-IAB. In Section VI-C, we show numerical results for both rate gain and latency gain of FD-IAB over HD-IAB. The benefit of equipping the IAB nodes with FD transceivers is more subtle and powerful than the familiar link-level gains (i.e., what is typically less than a two-fold improvement in spectral efficiency). We show that the more hops between a UE and the donor, the more the UE has to gain from a FD-IAB deployment, both in terms of latency and throughput. For example, the throughput of a UE four hops from the donor can improve by $8\times$ if IAB nodes are upgraded to FD. This many-fold increase stems from the increased scheduling opportunities that FD provides at each IAB node, which reduces multiplexing delays that are particularly significant in the context of multihop routing.

Impact of imperfect self-interference cancellation on network performance. In Section VI-C we present numerical results to quantify the effect of imperfect self-interference cancellation on the throughput performance of a FD-IAB network. We show that even if the residual self-interference is $10$ dB above the noise floor, a six-fold rate improvement can be achieved by FD-IAB over its HD counterpart. Reducing the residual-interference-to-noise ratio (RINR) below $0$ dB can result in $8$ times rate gain for a user at the fourth hop. The FD gains saturate for RINR below $-5$ dB, suggesting that—from a network perspective—further self-interference cancellation is likely not worthwhile. These insights can drive physical layer design decisions for FD in IAB networks.

Notation: ${\mathbf{A}}_{:,i}$ denotes the $i$-th column of matrix ${\mathbf{A}}$, ${\mathbf{A}}_{i,:}$ denotes the $i$-th row of matrix ${\mathbf{A}}$, ${\mathbf{A}}^{T}$ denotes the transpose of a matrix, $[{\mathbf{A}}]_{i,j}$ denotes the $(i,j)$-th element of ${\mathbf{A}}$, and ${\mathbf{A}}^{\dagger}$ denotes the hermitian transpose of a matrix. The $i$-th element of vector ${\mathbf{a}}$ is denoted by ${\mathbf{a}}_{i}$. $\mathbbm{1}_{\mathcal{A}}$ denotes the indicator function over set $\mathcal{A}$, and $|\mathcal{A}|$ denotes the cardinality of set $\mathcal{A}$.

Latency is a key performance metric in modern networks. 5G cellular networks, for example, are designed to support end-to-end packet delays on the order of milliseconds for mission critical and tactile internet applications [37]. In (10), the delay threshold of a network is captured by $\delta$ and constraint (10e). For a network designer, it is useful to know the minimum delay a network can support and how it relates network parameters, such as its routing matrix ${\mathbf{F}}$, link capacities ${\mathbf{C}}$, and scheduling matrix ${\mathbf{G}}$. Mathematically, this is equivalent to finding the minimum $\delta$ such that problem (10) is feasible.

Minimizing $\delta$ without any constraints on ${\bm{\lambda}}$, however, could lead to solutions that are practically undesirable where UEs receive zero throughput. We address this by introducing a new constraint

to ensure a minimum average arrival rate $\lambda_{\min}$ is met at each UE, which can be tuned by network engineers. While one can solve for the minimum $\delta$ using numerical solvers, the delay constraint (10e) is non-convex when optimizing over $\delta$, ${\bm{\lambda}}$, and ${\bm{\mu}}$, and hence intractable. This motivates us to substitute (10e) with a tighter constraint by replacing ${\mathbb{P}}[\tilde{D}_{m}h_{m}\leq\delta]\geq\eta$ with ${\mathbb{P}}[D_{v}h_{m}\leq\delta]\geq\eta,\forall(u,v)\in\mathsf{R}(m)$. Combining all of this leads to the formulation of problem (12), which includes minimizing over $\delta$ and incorporates modified constraints (12b) and (12e).

The new latency constraint (12e) ensures that the per-hop delay is less than $\delta/h_{m}$, which is a stricter requirement than (10e), and $h_{m}$ is as defined in (8). Rearranging the terms in (12e) and applying the change of variable $t=-\log(1-\eta)/\delta$, problem (12) can be reformulated as the following linear program.

Given Theorem 1, we can compute $\delta^{*}_{\mathrm{HD}}$ and $\delta^{*}_{\mathrm{FD}}$ for HD and FD deployments, respectively. The latency gain $\ell$ of FD-IAB over HD-IAB can then be expressed as

Here, ${\mathbf{G}}_{\mathrm{HD}}$ and ${\mathbf{G}}_{\mathrm{FD}}$ denote the scheduling matrices for the HD and FD deployments, respectively, and ${\mathbf{C}}_{\mathrm{HD}}$ and ${\mathbf{C}}_{\mathrm{FD}}$ denote the corresponding the capacity matrices. Depending on the quality of self-interference cancellation and the structure of the IAB routing tree $\mathcal{T}$, the bottleneck BS—which achieves the minimum in (14)—could be different for FD and HD deployments. Simplified expressions can be obtained if one considers the structure of the routing matrix ${\mathbf{F}}$ or particular cases of ${\mathbf{C}}$.

To illustrate a simple example, we refer to the IAB routing tree given in Fig. 2a to have

and $\tilde{{\mathbf{h}}}=[1,3,2,3,3]^{T}$ (defined in (14)). The HD and FD scheduling matrices are as given in (4), respectively. If we assume for this example that all edges have equal capacity $C$ and ${\mathbf{C}}_{\mathrm{HD}}={\mathbf{C}}_{\mathrm{FD}}=C{\mathbf{I}}$, then using Theorem 1 we have

and the latency gain is given by $\ell=t^{*}_{\mathrm{FD}}/t^{*}_{\mathrm{HD}}$, which is plotted in Fig. 3. This simple but insightful example illustrates that IAB with FD nodes can support $\ell=2.5$ times tighter latency constraints when the minimum arrival rate to each UE $\lambda_{\min}$ is $10\%$ of the link capacity $C$. When $\lambda_{\min}$ is just over a fifth of $C$, the latency gain is $\ell=12$ (nearly quintuples) and increases substantially for small increases in $\lambda_{\min}$ thereafter. This behavior is explained by the simple fact that FD-IAB can support higher values of $\lambda_{\min}$ which are infeasible for HD-IAB, driving the latency gain $\ell$ toward infinity.

In (15), we derived latency gain $\ell$ for a general IAB routing tree. In the following, we will derive a closed-form expression for (14) under both HD and FD deployments of a line network, along with the corresponding latency gain $\ell$.

Note that, since $\mathcal{T}$ represents a line network, $K+1$ is also the number of hops between the donor and the furthest UE. It is interesting to note that the bottleneck BS is either the first or the last IAB node. This can be attributed to the function $f(\cdot)$ (see Appendix A) and its property that it is either non-increasing or non-decreasing depending on whether $\lambda_{\min}$ is above or below a threshold value given in Theorem 2. On close inspection, we see that $f$ is the ratio of the fraction of time the BS is idle and the time it takes for the BS to serve one unit of the total arriving traffic (i.e., $\lambda_{\min}{\mathbf{F}}\bm{1}$) under the delay constraints. The idle time is an increasing function of $k$ and a decreasing function of $\lambda_{\min}$, whereas $\lambda_{\min}{\mathbf{F}}\bm{1}$ is an increasing function of $k$. This explains the peculiar property of $f$. One can also interpret $f$ as the product of the service rate of the BS and the fraction of time it is idle—in some sense it is the effective service rate. The inverse of which is the average delay experienced by the arriving flow at the BS. The objective of (13) (or equivalently (12)) is to find the bottleneck BS with the maximum service time within the resource allocation constraints. The delay threshold $\delta$ an IAB network can support is limited by this bottleneck BS.

In this section, we present analytical expressions for the maximum depth $K^{\max}$ of a line network that can support given latency and throughput targets. Note that the result of Theorem 2 can also be written as

Let $K^{\max}_{\mathrm{HD}}$ and $K^{\max}_{\mathrm{FD}}$ represent the maximum number of IAB nodes for the HD and FD line networks, respectively, that can guarantee a target delay threshold $\delta_{\mathrm{target}}$ and minimum arrival rate $\lambda_{\min}$. For notational convenience, let $\kappa_{\mathrm{HD}}$ and $\kappa_{\mathrm{FD}}$ denote the break points from (22) and (23) as follows.

In Fig. 4a we present the effect of $\lambda_{\min}$ on maximum feasible network depth $K^{\max}+1$ as computed in Theorem 3, for different backhaul SNR and access SNR of $5$ dB. Theorem 3 assumes perfect self-interference cancellation and hence $R_{b}$ and $R_{a}$ are computed by applying the Shannon formula to SNR. We evaluate (26) and (25) for a packet size of $10$ KB, bandwidth $W=100$ MHz, and $w=5$ UEs per BS. Solid lines denote $K^{\max}_{\mathrm{FD}}+1$ and dashed lines denote $K^{\max}_{\mathrm{HD}}+1$. The plot shows that FD-IAB can increase maximum feasible network depth by about three times for smaller throughput targets, compared to its HD counterpart. The plot also shows that a $2$ dB increment in the backhaul SNR typically translates to an increment of about two hops in $K^{\max}_{\mathrm{FD}}$. The same increment in backhaul SNR typically results in an increment of one hop in $K^{\max}_{\mathrm{HD}}$, and sometimes no increment at all. In other words, the marginal return of improving the backhaul link quality is more for FD-IAB. Fig. 4b shows the variation of $K^{\max}$ with the latency target $\delta_{\mathrm{target}}$ for fixed $\lambda_{\min}$ and different backhaul SNR. The plot shows similar trends as Fig. 4a and the marginal return of increasing backhaul SNR is more profound. Theorem 3 can also be used to derive a closed-form expression for hop gain by following steps similar to Corollory 1.

We model mmWave channels using the Saleh-Valenzuela-based model [3]. Let $N_{\mathrm{tx}}$ and $N_{\mathrm{rx}}$ denote the number of antennas at the transmitter and receiver respectively. The $N_{\mathrm{rx}}\times N_{\mathrm{tx}}$ channel matrix ${\mathbf{H}}$ can be written as

which is simply the composition of propagating $N_{\textrm{cluster}}$ clusters, each having $N_{\textrm{ray}}$ rays. We denote by $h_{i,j}\sim\mathcal{C}\mathcal{N}(0,1)$ the complex gain of the $j$-th ray in the $i$-th cluster. $\textrm{AoA}_{i,j}$ is the angle of arrival of the ray at the receiver, and $\textrm{AoD}_{i,j}$ is the angle of departure at the transmitter. The scalar in front of the summation handles power normalization. The vectors ${\mathbf{a}}_{\mathrm{tx}}(\cdot)$ and ${\mathbf{a}}_{\mathrm{rx}}(\cdot)$ represent the array response vectors at the transmitter and the receiver, respectively.

We assume that the network is noise-limited and interference from neighboring BSs is negligible. Hence, received signals are corrupted by additive noise and—at FD IAB nodes—by residual self-interference remaining after cancellation. Due to lack of measurements and characterization of the mmWave self-interference channel, we do not assume a particular model for the self-interference channel and our contribution does not rely on specific characteristics of a model. Instead, we evaluate our system based directly on RINR, which captures the degree of residual self-interference plaguing a desired receive signal after analog, digital, and/or spatial cancellation. The SINR at a FD IAB node $v$ in such a case is expressed as

where $\mathrm{RINR}_{v}$ is the RINR at the receiver due to self-interference and $\mathrm{SNR}_{v}$ is the SNR the equivalent HD IAB node observes, written as

Here, $P_{\mathrm{tx}}$ denotes the transmit power, $L_{v}$ is the inverse pathloss, ${\mathbf{H}}_{v}$ is the channel matrix, ${\mathbf{f}}_{\mathrm{tx}}$, and ${\mathbf{w}}_{\mathrm{rx}}$ are the transmit and receive beamforming vectors, $W$ is the system bandwidth, $\sigma^{2}$ is the noise power spectral density. The SNR at each UE can be expressed analogously.

Using the proposed optimization framework, we evaluate the gain of FD-IAB over HD-IAB for various depths and topologies. We begin with line networks with depths $K+1=2$, $K+1=3$, and $K+1=4$, and then evaluate a more general two-child tree network where the donor and each first-hop IAB node supports two child BSs. The IAB nodes are separated by $200$ m, the BSs transmit at $30$ dBm, and each BS is equipped with a uniform linear array (ULA) with $N_{\textrm{BS}}=64$ antennas. Each UE has a $16$-element ULA. Each BS serves five UEs which are dropped uniformly in a $100$ m radius around the BS. A sample drop for one donor and three IAB nodes—a network depth of $K+1=4$—line network is shown in Fig. 5 and the two-child tree network is illustrated in Fig. 11. The channels between pairs of devices are generated using (27), and ${\mathbf{f}}$ and ${\mathbf{w}}$ are drawn from a discrete fourier transform (DFT) codebook to maximize SNR (i.e., codebook-based beam alignment). For each Monte Carlo iteration, we drop a new set of UEs and generate channels such that $N_{\textrm{ray}}$ is drawn uniformly from the interval $[1,10]$ and $N_{\textrm{cluster}}$ is drawn from the interval $[1,6]$ [38]. The system operates at a carrier frequency of $30$ GHz with $100$ MHz bandwidth. For the pathloss, we use the model adopted in the 3GPP standard [39, page 26] for urban environment (UMa), which incorporates the effects of blockage, the multi-slope nature of the pathloss, and the elevation difference between the BSs and the users, through parameters that have been fit to real-world data. The additive noise power spectral density is $-174$ dBm/Hz plus a $10$ dB noise figure.

Packets of size $10$ KB arrive at the donor according to a Poisson process like the file transfer protocol (FTP) model 3 [33]. We divide the capacity of each edge by the mean packet size and normalize it to packets/sec, leading to units of delay in sec/packet. Note that these results can be transformed to any packet size by appropriate scaling. We use CVXPY [40] to solve our network design for $\eta=0.9$ (i.e., $90\%$ of the packets must be delivered to the destination UE within the target delay threshold $\delta$).

Using (10), we will now discuss the effect of deploying FD-IAB on network throughput, under the logarithmic utility function $U(\cdot)=\log(\cdot)$. As discussed earlier, the logarithmic utility achieves a healthy balance between network sum-rate and fairness, unlike pure sum-rate maximization which would result in greedy solutions that are unfair and practically undesirable, such as allocating all the resources to a first-hop UE with the best channel.

Rate gain with imperfect self-interference cancellation. For various depths of the described line network, we solve (10) for both HD and FD modes to obtain the optimal arrival rate vectors ${\bm{\lambda}}_{\mathrm{HD}}$ and ${\bm{\lambda}}_{\mathrm{FD}}$. Comparing FD to HD, the rate gain for UE $m$ is defined as $({\bm{\lambda}}_{\mathrm{FD}})_{m}/({\bm{\lambda}}_{\mathrm{HD}})_{m}$. Let $\mathcal{M}_{i}$ denote the set of UEs $i$ hops from the donor. Then, the sum-rate at $i$-th hop is defined as $\sum_{\mathcal{M}_{i}}{\bm{\lambda}}_{m}$ and the rate gain is defined as $\frac{\sum_{\mathcal{M}_{i}}({\bm{\lambda}}_{\mathrm{FD}})_{m}}{\sum_{\mathcal{M}_{i}}({\bm{\lambda}}_{\mathrm{HD}})_{m}}$. Fig. 6a shows the last-hop sum-rate and Fig. 7a shows the rate gain at the last hop versus RINR for various network depths. Naturally, as RINR increases, sum-rate and consequently rate gain suffer due to increased residual self-interference plaguing FD operation, diminishing the resource gains of FD over HD. Even with self-interference that is ten times stronger than noise (i.e., $\mathrm{RINR}=10$ dB), however, appreciable gains over HD are visible, especially for deeper networks. Reducing RINR below $-5$ dB does not yield meaningful rate improvements, as sum-rate saturates, which can drive physical layer FD design decisions from the perspective of a network operator.

Deeper networks benefit more from FD deployment. In Fig. 7a and Fig. 6a, we observe that an IAB network with more hops has more to gain from FD deployment. At low RINR, the last-hop gain for a network with depth two is about $1.5$ whereas for a four-hop network the last-hop sum-rate improves from $300$ packets/sec to more than $2000$ packets/sec, an eight-fold rate improvement—far beyond the familiar potential doubling of capacity with FD. Fig. 6b and Fig. 7b present the sum-rate and rate gain across different hops for a four-hop network. With FD, users at all hops throughout the network enjoy a healthy rate, unlike in HD-IAB, where only first-hop users see a high rate (even having used a logarithmic utility). UEs deeper in the network see higher rate gains with FD, since their rates under HD-IAB are so poor.

Note that the rate gain for the first hop UE is about $0.5$. This is because of the shape of the constraint set in (10) for HD-IAB. With the HD restriction, and even with the logarithmic utility, the first-hop UEs are greedily served because the third- and fourth-hop UEs can meet the delay constraint with very small arrival rates. However, a more fair operating point can be reached with FD-IAB. Thus, even though it seems to be sacrificing performance at the first-hop, in reality FD-IAB improves the overall network performance, measured by the logarithmic utility.

FD-IAB allows for tighter delay constraints. In our previous results, we fixed a delay threshold $\delta$ in examining rate gain. Now we evaluate rate gain for changes in delay threshold $\delta$. In Fig. 8a and Fig. 9a, we plot last-hop sum-rate and last-hop rate gain as a function of delay threshold $\delta$ for various network depths. In doing so, we consider an $\mathrm{RINR}=-15$ dB in light of the diminishing gains at low RINR, with the understanding that increased residual self-interference would degrade the backhaul capacities and affect the rate gain as in Fig. 7. The rate improvement offered by FD-IAB is seen across network depths, with deeper networks experiencing a higher rate gain as before. By alleviating the multiplexing delay at the IAB nodes, FD-IAB improves the arrival rates at stricter delay constraints. In fact, delay constraints which were infeasible for HD-IAB can be made feasible by upgrading to FD. Note that in Fig. 9a, the gain at the fourth hop for $\delta=3$ msec/packet is infinite, because the corresponding HD-IAB deployment is infeasible (rate of $0$ packets/sec in Fig. 8a). The rate gain saturates with increasing $\delta$, because for large delay thresholds the per-user arrival rate is limited by the backhaul capacities and not the multiplexing delay at the IAB node.

Fig. 8b and Fig. 9b show the sum-rate and rate gain across different hops for a four-hop network. The observations are consistent with the story so far: UEs further from the donor have more to benefit from FD-IAB. At very small $\delta$, the gain tends to infinity since a delay constraint that cannot be met by HD-IAB can in fact be met with FD. Like before, we see FD-IAB is capable of delivering fairer service to UEs throughout the network, while still meeting the delay constraints across multiple hops.

FD-IAB supports higher per-UE rates. To study improvements in per-UE rate with delay constraint, we solve the linear program (13) for both FD-IAB and HD-IAB modes and employ our definition of latency gain (15). We perform simulations fixing $\mathrm{RINR}=-15$ dB and examine the $(\lambda_{\min},\delta^{*})$-pairs that can be met by HD and those that can be met by FD. Fig. 10b illustrates these results for various network depths, where curves offering higher $\lambda_{\min}$ at lower $\delta^{*}$ are more desirable (i.e., toward the lower right). For a two-hop network, both the HD and the FD deployments handle the latency and throughput constraints equally well. With more than two hops, however, the power of FD in this network setting is brought to light. For a given latency target, FD-IAB can deliver a minimum rate $\lambda_{\min}$ far greater than HD-IAB: for $\delta^{*}=20$ msec/packet, FD offers over $1.5$ times higher $\lambda_{\min}$ in a four-hop network and over $1.3$ times higher $\lambda_{\min}$ in a three-hop network. Minimum rate targets $\lambda_{\min}$ that can be met by HD-IAB networks only with prohibitively high delays, can also be met by FD-IAB while simultaneously offering low latency. For instance, in a three-hop network, HD-IAB can deliver $\lambda_{\min}=900$ packets/sec only be introducing delays well beyond what is practical; a FD-IAB network can do so while meeting a latency of approximately $\delta^{*}=5$ msec/packet. As a result, latency gain can tend toward infinity.

While line networks are expected to be common for initial IAB deployments, the presented framework is in no way limited to them. To briefly illustrate this, we use our framework to study the IAB deployment in Fig. 11, which consists of one donor node, $K=6$ IAB nodes, and five UEs per BS. We refer to this as a two-child tree network since the donor and each first-hop IAB node deliver backhaul to two IAB nodes. While not explicitly shown due to space constraints, the key takeaways from the discussion on line networks holds, namely: (i) rate gain saturates for RINR below around $-5$ dB, (ii) UEs further from the donor have more to benefit from an FD-IAB deployment, and (iii) FD-IAB can support larger $\lambda_{\min}$ for the same minimum feasible delay $\delta^{*}$ as was shown in Fig. 10.

Nonetheless, there are some additional insights to draw from this two-child tree network. To start, we consider Fig. 12a, which shows the mean per-user arrival rate at various hops in the network as a function delay threshold. Unlike line networks, the donor and first-hop IAB nodes now deliver backhaul to more than one IAB node. Consequently, a HD-IAB deployment cannot support strict latency requirements at the second- and third-hop with a practically viable arrival rate. This is because first-hop IAB nodes must multiplex incoming and outgoing backhaul links, significantly worsening the multiplexing delay and the radio resource bottleneck. First-hop users, however, can enjoy high rates—even under strict delay thresholds—since they are not subjected to these relaying delays. The disparity in user service is evident in Fig. 12a as the delay threshold is made stricter, with first-hop arrival rate increasing as users deeper in the network see their service degrading.

With FD-IAB, however, the resource bottleneck is alleviated and the network can serve its multihop users much more fairly and with satisfactory rates. We can see from Fig. 12a that FD-IAB can support delay thresholds infeasible for its HD counterpart, with a healthy throughput across hops. The corresponding rate gain of FD-IAB over HD-IAB is shown in Fig. 12b. FD-IAB can meet very strict delay thresholds for second- and third-hop users when HD-IAB cannot, driving the rate gain toward infinity as $\delta$ is decreased. As $\delta$ increases, the rate gain saturates like in Fig. 9b. For first-hop users, we observe a rate gain less than one, which may seem undesirable at first glance but like Fig. 7b is justified by FD-IAB achieving a fairer distribution of arrival rates across hops.

Comparing these results of a two-child tree network to that of the simulated line network, we notice a few key differences. First, comparing Fig. 12b to Fig. 9b, we see that the rate gain of the second- and third-hop users in the two-child tree network saturates to noticeably less than $2$, while that for the line network saturates to nearly $2$. This is difficult to explain precisely, though it is likely largely due to the fact that the donor and each first-hop IAB node must multiplex outgoing backhaul transmissions—which line networks do not suffer from. FD cannot directly alleviate these multiplexing costs, suggesting that the network may be bottlenecked by the resources consumed when juggling multiple backhual links. This motivates future work to more thoroughly study this disparity across network topologies and investigate how to best deploy IAB networks when IAB nodes are FD-equipped.