Resource Management for 5G NR Integrated Access and Backhaul: a Semi-centralized Approach
11 Resource Management for 5G NRIntegrated Access and Backhaul:a Semi-centralized Approach
Matteo Pagin, Tommaso Zugno,
Student Member, IEEE,
Michele Polese,
Member, IEEE,
Michele Zorzi,
Fellow, IEEE
This paper has been submitted to IEEE for publication. Copyright may change without notice.
Abstract
The next generations of mobile networks will be deployed as ultra-dense networks, to match thedemand for increased capacity and the challenges that communications in the higher portion of thespectrum (i.e., the mmWave band) introduce. Ultra-dense networks, however, require pervasive, high-capacity backhaul solutions, and deploying fiber optic to all base stations is generally considered to betoo expensive for network operators. The 3rd Generation Partnership Project (3GPP) has thus introducedIntegrated Access and Backhaul (IAB), a wireless backhaul solution in which the access and backhaullinks share the same hardware, protocol stack, and also spectrum. The multiplexing of different linksin the same frequency bands, however, introduces interference and capacity sharing issues, thus callingfor the introduction of advanced scheduling and coordination schemes. This paper proposes a semi-centralized resource allocation scheme for IAB networks, designed to be flexible, with low complexity,and compliant with the 3GPP IAB specifications. We develop a version of the Maximum WeightedMatching (MWM) problem that can be applied on a spanning tree that represents the IAB networkand whose complexity is linear in the number of IAB-nodes. The proposed solution is compared withstate-of-the-art distributed approaches through end-to-end, full-stack system-level simulations with a3GPP-compliant channel model, protocol stack, and a diverse set of user applications. The results showhow that our scheme can increase the throughput of cell-edge users up to 5 times, while decreasing theoverall network congestion with an end-to-end delay reduction of up to 25 times.
Matteo Pagin, Tommaso Zugno and Michele Zorzi are with the Department of Information Engineering, University of Padova,Padova, Italy. Email: { paginmatte, zugnotom, zorzi } @dei.unipd.itMichele Polese is with the Institute for the Wireless Internet of Things, Northeastern University, Boston, MA USA. Email:[email protected] a r X i v : . [ c s . N I] F e b I. I
NTRODUCTION
Future generations of cellular networks will provide groundbreaking network capacity, inconjunction with a significantly lower delay and ubiquitous coverage [1]. 5th generation (5G)and beyond deployments will support new mobile broadband use cases, e.g., Augmented (AR)and Virtual Reality (VR), and expand into new vertical markets, enabling an unprecedenteddegree of automation in industrial scenarios, Vehicle-to-Everything (V2X) communications,remote medical surgery and smart electrical grids.To this end, the 3rd Generation Partnership Project (3GPP) has introduced various technolog-ical advancements with the specifications of the 5G Radio Access Network (RAN) and CoreNetwork (CN), namely NR and 5G Core (5GC) [2]. In particular, NR features a user and controlplane split, a flexible Orthogonal Frequency Division Multiplexing (OFDM) frame structure,and the support for millimeter wave (mmWave) communications, while the CN introducesvirtualization and slicing [3].Specifically, the mmWave band, i.e., the portion of the spectrum between 30 GHz and 300GHz, represents the major technological enabler toward the Gbit/s capacity target. These fre-quencies are characterized by the availability of vast chunks of contiguous and currently unusedspectrum, in stark contrast with the crowded sub-6 GHz frequencies. However, mmWaves exhibitunfavorable propagation characteristics such as high isotropic losses and a marked susceptibilityto blockages and signal attenuation [4], [5]. These issues can be partially mitigated usingbeamforming through large antenna arrays, thanks to the small wavelengths and advances in low-power Complementary Metal-Oxide Semiconductor (CMOS) RF circuits [6]; nevertheless, theirintroduction alone is not enough for meeting the high service availability requirement. Therefore,mmWaves networks also need densification, to decrease the average distance between mobileterminals and base stations and improve the average Signal-to-Interference-plus-Noise Ratio(SINR). The theoretical effectiveness of this technique is well understood [7]; however, achievingdense 5G deployments is extremely challenging from a practical point of view. Specifically,providing a fiber backhaul among base stations and toward the CN is deemed economicallyimpractical, even more so in the initial 5G deployments [8].Recently, wireless backhaul solutions for 5G networks have emerged as a viable strategytoward cost effective, dense mmWave deployments. Notably, the 3GPP has promoted IntegratedAccess and Backhaul (IAB) [9], i.e., a wireless backhaul architecture which dynamically splits the overall system bandwidth for backhaul and access purposes. IAB has been integrated in thelatest release of the 3GPP NR specifications. Prior research has highlighted that IAB representsa cost-performance trade-off [8], [10], as base stations need to multiplex access and backhaulresources, and as the wireless backhaul at mmWaves is less reliable than a fiber connection. Inparticular, IAB networks may suffer from excessive buffering (and, consequently, high latencyand low throughput) when a suboptimal partition of access and backhaul resources is selected,thus hampering the benefits that the high bandwidth mmWave links introduce [8], [11]. Therefore,it is fundamental to solve these non-trivial challenges to enable a smooth integration of IAB in5G and beyond deployments.
A. Contributions
This article tackles the access and backhaul partitioning problem by proposing an optimal,semi-centralized resource allocation scheme for 3GPP IAB networks, based on the MaximumWeighted Matching (MWM) problem on graphs. It receives periodic Key Performance Indicator(KPI) reports from the nodes of the IAB deployment, constructs a spanning tree that representsthe deployment, and uses a simplified, low-complexity version of the MWM to partition thelinks between access and backhaul. After a feedback step, each node can then schedule theresources at a subframe-level among the connected devices. To the best of our knowledge, thisis the first MWM-based resource allocation framework for 3GPP IAB networks at mmWaves,designed with three goals, i.e., it is flexible, integrated with the 3GPP specifications, and haslow complexity.The flexibility makes it possible to easily adapt the resource allocation strategy to differentrequirements, use cases, and classes of traffic for 5G networks. We achieve this by developing ageneric optimization algorithm, which identifies with a configurable periodicity the access andbackhaul partition that optimizes a certain utility function. The selection of the utility functionprioritizes the optimization of different metrics, e.g., throughput or latency, which in turn canbe mapped to different classes of traffic. To achieve the second goal, i.e., the compliance withthe 3GPP IAB specifications, the resource allocation framework relies only on information thatcan be actually exchanged and reported in a 3GPP deployment. In this regard, we also reviewthe latest updates related to the 3GPP IAB standardization activities.Finally, the algorithm operates with a low complexity, i.e., we propose a version of the MWMalgorithm that can be applied on spanning trees with linear complexity in the number of nodes in the network infrastructure, and demonstrate its equivalence to the generic (and more complex)MWM. Additionally, the proposed framework also relies on a feedback exchange that is linearin the number of base stations, and is thus decoupled from the number of users.Furthermore, we evaluate the performance of the proposed scheme with an end-to-end, full-stack system-level simulation, using the ns-3 mmWave module [12] and its IAB extension [11].This represents the first evaluation of an optimized resource allocation scheme for IAB with asimulator that is based on a 3GPP-compliant protocol stack, uses 3GPP channel models, andintegrates realistic applications and transport protocols. The extended performance evaluationhighlights how the proposed scheme improves the throughput of a diverse set of applications,with a 5-fold increase for the worst case users, with different packet sizes and transport protocols,while decreasing the latency and buffering at intermediate nodes by up to 25 times for the smallestpacket sizes.
B. State of the Art
This section reviews relevant research on resources allocation in a multi-hop wireless network,deployed through either IAB or other wireless mesh solutions [13].The literature adopts different approaches to model and solve the resource allocation problem.The first, discussed in [14]–[20] is based on conventional optimization techniques. Specifically,the authors of [14] present a simple and thus tractable system model and find the minimalnumber of Next Generation Node Bases (gNBs) featuring a wired backhaul that are needed tosustain a given traffic load. Their work is further extended in [15], which provides an analysis ofthe performance benefits introduced by additional, fiber-less gNBs. In [16], the mobile networkis modeled as a noise-limited, k -ring deployment. Such model is then used to obtain closed-form expressions for the max-min rates achieved by User Equipments (UEs) in the network.Moreover, [17] proposes a system model which leads to an NP-hard optimization problem, eventhough it considers single-hop backhaul networks only, and uses deep Reinforcement Learning(RL) to reduce its computation complexity. In [18], the joint routing and resource allocationproblem is tackled via a Linear Programming (LP) technique. Notably, this work assumes thatdata can be transmitted (received) toward (from) multiple nodes at the same time. Similarly,the authors of [19] formulate a Time Division Duplexing (TDD), multi-hop resource allocationoptimization problem which leverages the directionality of mmWave antennas, albeit in thecontext of Wireless Personal Area Networks (WPANs). Since such problem is also NP-hard, a sub-optimum solution is found. Finally, [20] focuses on joint link scheduling, routing andpower allocation in multi-hop wireless networks. As in previous cases the obtained optimizationproblem is not tractable: in this instance such obstacle is overcome by studying the dual problemvia an iterative approach.The second approach relies on stochastic geometry to model IAB networks [10], [21]. Specif-ically, [21] determines the rate coverage probability of IAB networks and compares differentaccess/backhaul resource partitioning strategies. Similarly, [10] provides a comparison of orthog-onal and integrated resource allocation policies, although limited to single-hop wireless networks.Another significant body of literature leverages Markov Chains (MCs) to study IAB networks;some of these works can be interpreted as a direct application of such theory [22], [23], whileothers [24]–[27] exploit a more complex framework. The papers which belong to the formerclass are based on the pioneering work of [28], which inspects the stability of generic multi-hopwireless networks and formulates a throughput-maximizing algorithm known as back-pressure . Inparticular, [22] focuses on the optimization of the timely-throughput, i.e., takes into account thatpackets usually have an arrival deadline. Such problem is then addressed by formulating a MarkovDecision Process (MDP), leading to a distributed resource allocation algorithm. Similarly, [23]proposes an algorithm that also targets throughput optimality but, contrary to the back-pressurealgorithm, manages to avoid the need for per-flow information. On the other hand, the body ofliterature which belongs to the latter class uses the MC-derived Network Utility Maximization(NUM) framework first introduced in [29] and [30]. Specifically, the authors of [24] focus onsatisfying the Ultra-Reliable Low-Latency Communication (URLLC) Quality of Service (QoS)requirements by jointly optimizing routing and resource allocation. Then, the problem is solvedusing both convex optimization and RL techniques. In [25], an in-depth analysis of a mmWave,multi-hop wireless system is presented, proposing and comparing three different interferenceframeworks, under the assumption of a dynamic TDD system. This work is extended in [26]and [27], which consider respectively a Spatial Division Multiple Access (SDMA) and a Multi-User (MU)-Multiple Input, Multiple Output (MIMO) capable system.Finally, only a small portion of the literature [8], [11], [31] analyzes the end-to-end perfor-mance of IAB networks. Specifically, the authors of [11] extend the ns-3 mmWave module,introducing realistic IAB functionalities which are then used to characterize the benefit ofdeploying wireless relays in mmWave networks. Their work is extended in [31], where pathselection policies are formulated and their impact on the system performance is inspected. A further end-to-end analysis of IAB networks is carried out in [8], providing insights into thepotentials of this technology and the related open research challenges.Concluding, the literature exhibits the presence of algorithms relying on a varying degreeof assumptions on the network topology and the knowledge of system. Furthermore, most ofthe aforementioned studies lack an end-to-end, full-stack system-level analysis of the proposedsolution. Conversely, this paper proposes a semi-centralized resource allocation scheme, whichalso has a low complexity, both computationally and in terms of required feedback. Moreover,we provide considerations on how our proposed solution can be implemented and deployed instandard-compliant 3GPP IAB networks, and compare such solution to the state of the art withan end-to-end, realistic performance analysis. C. Paper structure
The remainder of this paper is organized as follows. Sec. II describes our assumptions andthe system model. Then, Sec. III presents a novel scheme for resource partitioning in mmWaveIAB networks, along with considerations on how it can be implemented in 3GPP NR. Finally,Sec. IV describes the performance evaluation results and Sec. V concludes this paper.II. IAB
NETWORKS
The following paragraphs identify the characteristics and constraints of mmWave IAB, ac-cording to the 3GPP design guidelines presented in [9] and the specifications of [32].
A. Network topology
In general, an IAB network is a deployment where a percentage of gNBs (i.e., the IAB-nodes)use wireless backhaul connections to connect to a few gNBs (i.e., the IAB-donors) which featurea wired connection to the core network, as can be seen in Fig. 1. Moreover, these deploymentsexhibit a multi-hop topology where a strict parent-child relation is present. The former can berepresented by the IAB-donor itself or an IAB-node; the latter by either UEs or downstream IAB-nodes. In [9], no a priori limit on the number of backhaul hops is introduced. As a consequence,3GPP argues that IAB protocols should provide sufficient flexibility with respect to the numberof backhaul hops. Moreover, the Study Item (SI) on IAB [9] highlights the support for both thetopologies depicted in Fig. 1a, i.e., Spanning Tree (ST) and Directed Acyclic Graph (DAG) IAB.Clearly, the former exhibits less complexity but, at the same time, poses some limits in terms of (a) ST and DAG topologies. (b) System model notation.Figure 1. Comparison of the IAB network topologies analyzed in [9] and related notation. network performance: the possible presence of obstacles may result in a service interruption, dueto the unique backhaul route established by the UEs. On the other hand, a DAG topology offersrouting redundancy, which can be used not only to decrease the probability of experiencing a“topological blockage,” but also for load balancing purposes.
B. Multiple access schemes and scheduling
An in-band, dynamic partitioning of the access and backhaul spectrum resources is currentlypreferred by 3GPP [9], [32], together with half-duplex operations of the IAB-nodes. Moreover,most of the literature suggests that 5G mmWave systems will operate in a TDD fashion [4], [33].This choice is mainly driven by the stringent latency requirements which the next generation ofmobile networks will be required to support, and by the usage of analog or hybrid beamform-ing. The usage of Frequency Division Duplexing (FDD), in conjunction with the presence oflarge chunks of bandwidth, would lead to severe resource under-utilization and make channelestimation more difficult.Based on these considerations, the system model exhibits a TDD, Time Division MultipleAccess (TDMA)-based scheduling where the access/backhaul interfaces are multiplexed in ahalf-duplex manner. It follows that at any given time instant, each node of the IAB networkcannot be simultaneously involved in more than one transmission or reception. In particular,IAB-nodes cannot schedule time and frequency resources which are already allocated by theirparent for backhaul communications which involve them. Finally, the introduction of resourcecoordination mechanisms and related signaling is explicitly supported in the IAB specificationdrafts [9], [32]. Nevertheless, these solutions must reuse as much as possible the available NRspecifications and require at most minimal changes to the Rel.15 5GC and NR.
C. System model
According to these assumptions and referring to Fig. 1b., a generic IAB network can bemodeled as a directed graph G = {N , E } , where the set of nodes
N ≡ { n , n , . . . n |N | } comprises the IAB-donor, the various IAB-nodes and the UEs. Accordingly, the set of directededges E ≡ { e , e , . . . e |E| } ≡ { e n j → n k } j,k , where the edge e n j → n k originates at the parent node n j and terminates at the children n k , comprises in all the active cell attachments, either of mobileterminals to a gNB or from IAB-nodes towards their parent node. Since the goal of this paperis to study backhaul/access resource partitioning policies, this generic model can be actuallysimplified: in fact, all the UEs connected to a given gNB can be represented by a single nodein G without any loss of generality. Similarly, the same holds true for their links toward theserving gNB, which can then be represented by a single edge. Furthermore, this work focuseson ST topologies only. Nevertheless, the proposed framework can be easily extended to thecase of a DAG IAB network: it suffices to introduce a preliminary routing step where an ST G (cid:48) is computed from the actual DAG network G , for example by using the strategies presentedin [31]. Such process can be possibly repeated at each allocation instance, effectively removingany constraint on the network topology.We define as feasible schedule any set of links E (cid:48) ⊆ E such that none of them share acommon vertex, i.e., ∀ e n j → n k (cid:54) = e n l → n m ∈ E (cid:48) it holds that n j (cid:54) = n m and n l (cid:54) = n k . Let then f u be a utility additive map , namely, a function such that the overall utility experienced bythe system when scheduling edges e and e satisfies f u ( e , e ) = f u ( e ) + f u ( e ) . Let also W ≡ { w , w , . . . w |E| } be the set of positive weights whose generic entry w j represents theutility which is obtained when scheduling the j -th edge, namely, w j ≡ f u ( e j ) . Then, the overallutility of the system is U ≡ (cid:80) e k ∈ E (cid:48) f u ( e k ) = (cid:80) e k ∈ E (cid:48) w k . The goal is to find the feasible set E (cid:48)∗ which maximizes the overall utility, i.e., argmax E (cid:48) U . In computer science, this task is typicallyreferred to as the Maximum Weighted Matching problem [34].Finding the MWM of a given graph, in the general case, is not trivial from a computationalpoint of view. In fact, the fastest known MWM algorithm for generic graphs has a complexityof O ( | V || E | + | V | log | V | ) [35], posing serious limitations to the suitability of such algorithmto 5G and beyond networks, which target a connection density of 1 million devices per km .However, we argue that under the aforementioned assumptions on the system model, whichrestrict the network to an ST topology, it is possible to design an MWM-based centralized resource partitioning framework which exhibits linear complexity with respect to the networksize and which, as a result, is able to satisfy the scalability requirements highlighted by 3GPPin [9].III. S EMI - CENTRALIZED RESOURCE ALLOCATION SCHEME FOR
IAB
NETWORKS
This section presents an MWM algorithm for ST topologies (Sec. III-A), an efficient andMWM-based centralized resource partitioning framework for IAB networks (Sec. III-B) andsome considerations about its implementation (Sec. III-C).
A. MWM for ST graphs
As the first of our contributions, we present an algorithm, hereby called
T-MWM , whichcomputes the MWM of an ST in linear time. In particular,
T-MWM is a bottom-up algorithmwhich, upon receiving as input a weighted ST G described by its edge list E and the correspondingweight list W , produces as output a set of active edges E ∗ which are an MWM of G . Furthermore, E is from now on assumed to exhibit the following invariant: each IAB parent precedes itschildren in the list, hence avoiding the need for a recursion. This is automatically obtained aseach IAB child connects after its parent, and is thus added to the list in a subsequent position.Nevertheless, this assumption can be easily relaxed, albeit at the cost of losing as a side-effectthe bottom-up design.As can be seen in Alg. 1, the T-MWM algorithm basically performs two traversals. During thefirst one, the utility yielded by the various nodes and their children is computed. Then, duringthe second traversal, this knowledge is used for computing an MWM of the network.The correctness of this procedure can be easily proved. Consider the sub-tree of G whose rootis represented by the generic node n k . Let also F ( n k ) be the utility yielded by a MWM of suchsub-tree which activates a link originating from n k , and G ( n k ) , conversely, the utility providedwhen the MWM contains no links which originate from such node. Then, the correctness of thefirst phase of the algorithm, namely the computation of the F and G vectors, follows directlyfrom the following Lemmas. Lemma 1.
Let n k be an arbitrary internal node of G and { n j } k be the set of its children. Then,any MWM of G must contain an edge which has as one of its vertices either n k or an elementof { n j } k . Algorithm 1
Tree-Maximum Weighted Matching
Input:
A weighted ST G encoded by a list E , which associates each node in G to its edges, and the correspondingweights list W . Output:
An MWM E ∗ of G . procedure T-MWM ( E , W ) F ← ; G ← (cid:46) Initialize the utility vectors to zero vectors E ∗ ← {} (cid:46) Initialize the set of active edges as empty for each node n k ∈ E do (cid:46) Iterate over the various nodes, in ascendingorder w.r.t. to their depth in G maxU til ← ; maxEdge ( n k ) ← {} for each edge e k,j ≡ ( n k , n j ) do (cid:46) Iterate over its edges G ( n k ) ← G ( n k ) + F ( n j ) currU til ← W ( e k,j ) + G ( n j ) − F ( n j ) if currU til > maxU til then maxU til ← currU til ; maxEdge ( n k ) ← e k,j F ( n k ) ← G ( n k ) + maxU til end if end for end for for each node n k ∈ E do (cid:46) Iterate over the various nodes, in ascendingorder wrt to their depth in G if F ( n k ) ≥ G ( n k ) then E ∗ ← E ∗ ∪ e k, maxEdge ( n k ) F ( maxEdge ( n k )) ← − (cid:46) Ensure child does not get acti-vated multiple times end if end for return E ∗ end procedure Proof:
Suppose there exists an MWM E ∗ of G which does not contain any such edge. Thenthe set ˆ E ∗ ≡ E ∗ ∪ { e n k → n m } , where e n k → n m is the edge from n j to its (arbitrary) child n m isstill a feasible activation set, since no edge in E ∗ shares such vertices. Furthermore, since theweights are positive we have that f u ( ˆ E ∗ ) ≡ f u ( E ∗ ) + W ( e n k → n m ) > f u ( E ∗ ) , which is clearly acontradiction. Lemma 2.
For any internal node n k : G ( n k ) = (cid:80) { n j } k F ( n j ) F ( n k ) = (cid:80) { n j } k F ( n j ) + max { n j } k { W ( e n k → n j ) + G ( n k ) − F ( n k ) } where the set { n j } k comprises all the children of n k . Conversely, for leaf nodes F ( n l ) ≡ G ( n l ) ≡ . Proof: This lemma can be proved by induction over the height h k of the sub-tree corre-sponding to node n k . The base case is h k = 0 , i.e., when n k is a leaf node; in this case, trivially,both F ( n k ) and G ( n k ) are zero since no links exhibit n k as parent node and the sub-tree of G which originates in n k consists of n k only, respectively.Consider then the node n k having a sub-tree of height h k > . From Lemma 1 we know thateither n k or (at least) one of its children must be included in any MWM. If on the one handwe do not activate any edge which originates from n k , then no constraints on the children’sactivation are introduced. Therefore, in this case the maximum utility is obtained when all thechildren are active, hence G ( n k ) = (cid:80) { n j } k F ( n j ) . On the other hand, if we activate an edgefrom n k to one of its children n m then no additional edges which originate from the latter canbe added to E ∗ . It follows that the utility obtained in this instance reads: (cid:88) { n j (cid:54) = n m } k F ( n j ) + W ( e n k → n m ) + G ( n m ) and can be rewritten as: (cid:88) { n j } k F ( n j ) + W ( e n k → n m ) + G ( n m ) − F ( n m ) Such utility is maximized when n m is chosen as argmax { n j } k { W ( e n k → n j )+ G ( n j ) − F ( n j ) } , yielding: F ( n k ) = (cid:88) { n j } k F ( n j ) + max { n j } k { W ( e n k → n j ) + G ( n k ) − F ( n k ) } Finally, the validity of the last phase of
T-MWM follows from Lemma 3.
Lemma 3.
Given an ST G , let G k be its sub-tree of root n k . Then, an MWM of G can be computedby performing, in a recursive fashion and starting from the root, the following procedure:
1) If F ( n k ) ≥ G ( n k ) , add to E ∗ the edge from n k to n m , where the latter is defined as n m ≡ argmax { n j } k { W ( e n k → n j ) + G ( n j ) − F ( n j ) } . Then, repeat recursively on all the sub-treescorresponding to n k ’s children { n j } k s . t . n j (cid:54) = n m and on the children of n m itself.2) If F ( n k ) < G ( n k ) , repeat recursively on all the sub-trees which exhibit the children of n k as their root.Proof: This Lemma follows directly from the definitions of F and G and the previousLemmas. Specifically, the above procedure always yields a feasible activation, i.e., a matchingof G . In fact, in either options we never recurse on a node which has already been activated,hence no pair of edges ∈ E can share any vertices. Furthermore, due to the properties of F and G and their validity for each sub-tree in G , the edges of E ∗ comprise a maximal matching, i.e.,they yield the maximum possible utility among all the feasible schedules.Regarding the computational complexity of the proposed algorithm, it can be observed thatduring the first phase the main loop effectively scans each edge of G , hence exhibiting acomplexity O ( | E | ) . Moreover, the second phase of T-MWM has complexity O ( | V | ) , since itloops through all the network nodes. Therefore, we can conclude that the overall asymptoticcomplexity of the algorithm is O ( | V | + | E | ) , or, equivalently, O ( | V | ) since in an ST the numberof edges equals | V | − . B. Semi-centralized resource partitioning scheme
Based on the system model introduced in Sec. II, and the
T-MWM algorithm, we present ageneric optimization framework which partially centralizes the backhaul/access resource parti-tioning process, in compliance with the guidelines of [9]. The goal of this framework is to aidthe distributed schedulers, adapting the number of OFDM symbols allocated to the backhaul andaccess interfaces to the phenomena which exhibit a sufficiently slow evolution over time, i.e.,large scale fading and local congestion. This optimization is undertaken with respect to a genericadditive utility function f u . An IAB network of arbitrary size is considered, composed of a singleIAB-donor, multiple IAB-nodes and a (possibly time-varying) number of UEs which connectto both types of gNBs. Furthermore, let the topology of the IAB network be pre-computed, forinstance by using the policies of [31], and assume that a central controller is installed on theIAB-donor. Figure 2. Creation of the IAB network graph. The original topology, exhibiting the actual cell attachments, is depicted on theleft. Conversely, the reduced topology is shown on the right.
The proposed framework can be subdivided into the following phases, which are periodicallyrepeated every T alloc subframes:1) Initial setup . This step consists in the computation of the simplified IAB network graph
G ≡ {V , E } . Specifically, V is composed of the donor, the various IAB-nodes and, possibly,of additional nodes which represent the set of UEs that are connected to a given gNB.Accordingly, E contains the active cell associations of the aforementioned nodes. Thisprocess, depicted in Fig. 2, must be repeated every time the IAB topology changes, i.e.,whenever a new UE performs its Initial Access (IA) procedure or an IAB-node connectsto a different parent due to a Radio Link Failure (RLF).2) Information collection . During this phase, the various IAB-nodes send to the centralcontroller a pre-established set of information for each of their children in G . For instance,this feedback may consist in their congestion status and/or information regarding theirchannel quality. To such end, the implementation of this paper uses modified versions ofpre-existing NR Release 16 Control Elements (CEs), as strongly recommended in the IABSI [9]. However, the scheme does not actually impose any limitations in such regard.3) Centralized scheduling indication . Upon reception of the feedback information, the centralcontroller calculates the set of weights W accordingly. Then, an MWM of G is computedusing the T-MWM algorithm and producing as output the activation set E ∗ , which maximizesthe overall utility of the system with respect to the chosen utility function. Subsequently, E ∗ is used in order to create a set of favored downstream nodes, i.e., of children whichwill be served with the highest priority by their parent, as depicted in Fig. 3. Finally, thesescheduling indications are forwarded to the various IAB-nodes which act as parents in theedges of E ∗ .
64 9
Figure 3. Computation of the MWM and of the corresponding scheduling indications. Distributed scheduling allocation . During this phase, the various IAB-nodes make use ofthe indications received by the central controller, if available, in order to perform the actualscheduling (which is, therefore, predominantly distributed). Specifically, the favored nodesare served with the highest priority, while the remaining downstream nodes are scheduledif and only if the resource allocation of the former does not exhaust the available OFDMsymbols.It is important to note that since G contains only the IAB-nodes, the donor and at most one“representative” UE per gNB, the proposed scheme effectively performs only the backhaul/accessresource partitioning in a centralized manner. On the other hand, the actual Medium AccessControl (MAC)-level scheduling is still undertaken in a distributed fashion, albeit leveragingthe indications produced by the central controller. The major advantages which this two-tierdesign exhibits, compared to a completely centralized solution, are the presence of a relativelylight signaling overhead and the ability to promptly react to fast channel variations, for instancecaused by small scale fading. C. Implementation of centralized allocation schemes in mmWave IAB networks
The remainder of this section discusses how the proposed scheme can be implemented in IABdeployments, with references to how the 3GPP specifications can support it.Basically, the resource allocation framework requires (i) a central controller, which is installedon the IAB-donor, or could be deployed in a RAN Intelligent Controller (RIC) following the O-RAN architecture [36]; and (ii) a scheduler which exchanges resource coordination informationwith the former and computes the weights for the resource allocation. In particular, referring tothe aforementioned phases of the proposed scheme, the following implementation considerationscan be made.
1) Initial setup:
The setup of the various centralized mechanisms is subdivided into two sub-phases: an initial configuration, where the relevant entities are initialized, and a periodic updateof the topology information. The former takes place when the IAB-nodes are first connectedto the network. During this phase, the controller acquires preliminary topology information, byleveraging the configuration messages which are already exchanged during the usual Rel.16IA procedure, generating a map which associates the depth in the IAB network to a pair ofchild-parent global identifiers (which from now on will be referred to as “IDs”). Since thisphase takes place when no UE has performed its IA procedure yet, the exchanged topologyinformation concerns the donor and IAB-nodes only.Moreover, the central controller is in charge of periodically updating the topology information.In order to minimize the signaling overhead, this process does not require any additional controlinformation: in fact, the status information which is already collected in a periodic manner canbe leveraged in such regard. Specifically, the periodic info received from the various IAB-nodes,which will carry a list of ID-value pairs, is analyzed. The child-parent associations are thencompared with the ones known by the controller, updating the latter whenever the two topologyinformation happens to differ.
2) Information collection:
The generation of the feedback information is performed in adistributed manner by both the IAB-nodes and the IAB-donor. To such end, the current imple-mentation features the forwarding of information on the channel quality and buffer status, in theform of Channel Quality Informations (CQIs) and Buffer Status Reports (BSRs) respectively. Thischoice is driven by both the will of maximizing the re-utilization of the NR Rel.16 specificationsand the goal of making use of MAC-level CEs only, hence avoiding the introduction of anyconstraint regarding the supported IAB-relaying architecture.In particular, the CQI and BSR information is generated by analyzing the corresponding CEswhich are received by the host gNB, and checking whether the source Radio Network TemporaryIdentifier (RNTI) belongs to an IAB-node or to a UE. In the first case, the corresponding IDis retrieved and an entry carrying such identifier along with its CQI/BSR value is generated.The feedback information concerning the UEs, instead, is averaged in the case of the CQIs andadded up for the BSRs, to obtain a single value for each gNBs. It can be noted that both CQIsand BSRs are available to the scheduler, since the UL buffer statuses are already periodicallyreported by the downstream nodes via their BSRs and the DL statuses can be easily retrieved bythe former, since the Radio Link Control (RLC) buffers reside on the same node as the scheduler itself, i.e., the gNB.Referring to the 3GPP specifications of [9], [37], [38], the buffers occupancy can then beforwarded to the IAB-donor by introducing a periodic-only BSR whose period is controlledby an ad hoc Radio Resource Control (RRC) timer. Similarly, the channel qualities can bereported by the various IAB-nodes via additional periodic CQIs which would carry only theCQI index, hence neglecting the Rank Index (RI) and Precoding Matrix Index (PMI), sincethis information would generate unnecessary signaling overhead. These CEs would preferablyleverage pre-existing NR measurements: the main novelty would be the introduction of theirperiodic reporting to the IAB-donor. To such end, the 5G CQI and BSR data-structures requirean additional field which carries the ID, if the chosen IAB-relaying architecture does not featurean Adaptation Layer [9]. Conversely, relaying solutions which support the latter can reuse theNR CEs and let such layer introduce an additional header.
3) Centralized scheduling indication:
Periodically, the controller located at the donor makesuse of the feedback received by the IAB-nodes to first compute the weights of the various networklinks and then generate the centralized scheduling indications. We propose the following policiesto compute the weights for the MWM problem:1)
Max Sum-Rate (MSR) . This policy maximizes the overall Physical (PHY)-layer through-put, i.e., the utility function is f MSR u ≡ (cid:88) e i → k ∈ E ∗ c i, k , and the weight assigned to the edge from node i to node k reads w i, k ≡ c i, k , where c i, k isthe capacity of the link e i → k .2) Backlog Avoidance (BA) . This resource partitioning strategy aims at avoiding congestion.Therefore, the system utility is: f BA u ≡ (cid:88) e i → k ∈ E ∗ q i, k , where the weight w i, k reads q i, k , namely, the amount of buffered data which would reachits next hop in the IAB network by crossing the link e i → k .3) Max-Rate Backlog Avoidance (MRBA) . This represents the most balanced option amongthe three, since it exploits favorable channel conditions while also preventing networkcongestion and favoring network fairness. The weight assigned to link e i → k is: w i, k ≡ c i, k + η · q i, k · (cid:18) µµ thr (cid:19) k , where η , µ thr and k are arbitrary parameters and µ represents the number of subframeswhich have elapsed since the last time edge e i → k has been marked as favored.Once the weights are computed, the controller obtains an MWM of the network via animplementation of the aforementioned T-MWM . This function outputs the activation set E ∗ , i.e.,a map associating the ID of the parent gNB to the one of its favored downstream node. Notably,this set does not necessarily contain scheduling indications for each IAB-node in the network:an entry corresponding to a given gNB is present if and only if such node is indeed active in theMWM. First, this map is used by the controller in order to keep track of which link has not beenfavored and for how long; this information may then be used to introduce a weight predictionmechanism, improving the robustness of the scheme with respect to the information collectionperiod. Finally, these scheduling indications are forwarded to the corresponding IAB-nodes.
4) Distributed scheduling allocation:
The last phase of the resource allocation procedureconsists in the distributed MAC-level scheduling. Before assigning the available resources, thevarious schedulers check whether any indication has been received from the controller. Based onthis condition, the buffer occupancy information is then split into two groups. The first containsthe BSRs related to the favored RNTI (if any), with the caveat that if the latter indicates thecumulative access link, then this set contains the BSRs of all the UEs attached to the host gNB,while the other comprises the remaining control information. The resource allocation processis then undertaken twice: first considering the set of favored BSRs only, then the remainder ofthese CEs.Thanks to this design, the favored link(s) is (are) scheduled with the highest priority, while therest of the network only gets the remaining resources. In such a way, the information receivedby the controller is actually used as an indication and not as the eventual resource allocation .For instance, the gNBs are free to override these indications whenever the buffer of the favoredchild is actually empty, due to discrepancies between its actual status and the related informa-tion available to the controller. In such a way, the unavoidable delay between the informationcollection and the reception of the scheduling does not lead to any resource underutilization.Moreover, this is achieved with minimal changes to the state of the art schedulers, making theproposed scheme relatively easy to implement and deploy in real-world networks. IV. P
ERFORMANCE EVALUATION
We implemented the proposed resource allocation scheme in the popular open source simulatorns-3, exploiting the mmWave module [12] and its IAB extension [11], to characterize the system-level performance of the proposed solution with realistic protocol stacks, scenarios, and userapplications.The ns-3 mmWave module is based on [39] and introduces mmWave channel models, includingthe 3GPP channel model for 5G evaluations [40], and highly customizable PHY and MAC layerimplementation, with an NR-like flexible OFDM numerology and frame structure. Additionally,the IAB module [11] models wireless relaying functionalities which mimic the specificationspresented in [9]. Specifically, this module supports both single and multi-hop deployment sce-narios, auto-configuration (within the network) of the IAB-nodes and a detailed 3GPP protocolstack, allowing wireless researchers to perform system-level analyses of IAB systems in ns-3.It is of particular relevance to understand how the scheduling operations are implemented inthe IAB module, since they offer not only the baseline for the proposed scheme, but also validguidelines for real-world deployments. The current ns-3 IAB schedulers exhibit a TDMA-basedmultiplexing between the access and backhaul interfaces. Moreover, scheduling decisions areundertaken in a distributed manner across the IAB network, i.e., each gNB allocates the resourceswhich its access interface offers (to both UEs and IAB-nodes) independently of the other gNBsin the network. In fact, in an IAB network these scheduling decisions are almost independent ofone another: if a parent node schedules the backhaul interface of a downstream node, clearly thelatter will be constrained in its own scheduling decisions, as it will not be allowed to allocatethe time resources which have already been scheduled for backhaul transmissions by its parent.Therefore, in a tree-based, multi-hop wireless network the various gNBs need to know in advancethe scheduling decisions performed by their upstream nodes: to solve this problem, the authors ofthe IAB module for ns-3 introduced a “ look-ahead backhaul-aware scheduling mechanism ” [11].Such mechanism features an exchange of Downlink Control Information (DCI) between theaccess and backhaul interfaces: in such a way, any time resources already scheduled by theparent for backhaul communications can be marked as such by the corresponding downstreamnode, preventing any overlap with other transmissions. Furthermore, the look-ahead mechanismrequires the schedulers of the various gNBs to commit to their resource allocation for a giventime T at a time T − k , where k − is the maximum distance (in terms of wireless hops) of − −
50 0 50 100 150 − IAB-donor IAB-node UE (a) S
IMULATION PARAMETERS P ARAMETER V ALUE
Number of runs N runs T sim T alloc { , , } subframesLayer 4 protocol { UDP, TCP } UDP packet size s UDP { , , , } BWeight policy f u { MSR, BA, MRBA } (b)Figure 4. On the left, a realization of the simulation scenario is depicted; the dotted lines represent the cell-attachments of theIAB-nodes. On the right, a brief list of simulation parameters is provided. any node from the donor. In such a way, the DCIs will have time to propagate across the IABnetwork and reach the farthest node at time T − , thus allowing its scheduler to perform theresource allocation process at least one radio subframe in advance. A. Simulation scenario and parameters
The purpose of these simulations is to understand the performance of the proposed resourcepartitioning framework in the context of its target deployment, i.e., a multi-hop IAB network.As a consequence, the reference scenario consists of a dense urban deployment with a singleIAB-donor and multiple IAB-nodes, as depicted in Fig. 4a. In particular, the various gNBs aredistributed along an urban grid where the donor is located at the origin while the IAB-nodesare deployed along the street intersections, with a minimum inter-site distance of 100 m. TheIAB-nodes attachments are computed using the so-called
HQF policy presented in [31]; however,this choice does not introduce any loss of generality since such parameter is fixed for all theruns. A given number of UEs are deployed within the surroundings of these base stations, withan initial position which is randomly sampled from circles of radius ρ and whose centers arethe various gNBs.Both the IAB-donor and the IAB-nodes are equipped with a phased array featuring 64 antennaelements, and transmit with a power of 33 dBm; conversely UEs are equipped with 16 antennaelements and their transmission power is restricted to 23 dBm. Notably, the presence of additionalantenna elements at the gNBs is a key (but reasonable) assumption, as it allows base stationsto achieve a high beamforming gain. In turn, it is possible to achieve a high capacity, which is fundamental to avoid performance bottlenecks, given the absence of a fiber backhaul. The UEsdownload data which originates from sources that are installed on a remote host; both the UserDatagram Protocol (UDP) and the Transmission Control Protocol (TCP) are used. For the UDPsimulations, the rate of the sources is varied from 4 to 40 Mbps to introduce different degrees ofsaturation in the network. Therefore, in these simulations only DL traffic is considered. Finally,the performance of the proposed policies is hereby compared with the baseline of [11], indicatedas “Dist.” by examining end-to-end throughput, latency, and a network congestion metric. B. Throughput
The first metric which is inspected in this analysis is the end-to-end throughput at the appli-cation layer. As a consequence, only the packets which are correctly received at the uppermostlayer of the destination node in the network are taken into account. In particular, for each UEand each simulation run, the long-term average throughput is computed as follows: S APP k,n ≡ B ( T sim , k, n ) T sim where B ( t, j ) is the cumulative number of bits received up to time t by the k -th UE, during the n -th simulation run. Then, the distribution of S APP , namely, the vector containing the collectionof the S APP k,n values across the different runs and UEs, is analyzed.Figs. 5a and 5b report the Empirical Cumulative Distribution Function (ECDF) of S APP , for aUDP packet size of 100 and 500 bytes, respectively, and the policies introduced in Sec. III-C. Inthe former, we can notice that the introduction of the centralized framework increases by up to15% the percentage of UEs whose throughput matches the rate of the UDP sources. Moreover, byfocusing on the leftmost portion of Fig. 5a we can observe another interesting result, concerningthe throughput experienced by the UEs which do not fulfill their QoS requirements. In fact,with respect to the first quartile the distributed scheduler achieves the worse performance, with25% of the UEs obtaining a throughput smaller than 3.3 Mbps. The centralized frameworksignificantly improves these results, even though the extent of such improvements varies quitedramatically across the different policies. Compared with the distributed case, the MSR policyachieves a higher throughput with respect to all the percentiles, albeit exhibiting the same highvariance of the former. Instead, the BA and MRBA policies have a dramatic impact on thesystem performance, introducing a 5-fold increase of the worst case throughput coupled with asignificantly lower variance. . . . . UE throughut [Mbps]
Distr MSRBA MRBA (a) s UDP = 100 B, i.e., UDP rate of 8 Mbps. . . . . UE throughut [Mbps]
Distr MSRBA MRBA (b) s UDP = 500 B, i.e., UDP rate of 40 Mbps.Figure 5. Per-UE end-to-end throughput ECDFs. The dashed line represents the rate of the UDP sources.
These results can be explained as follows: since a UDP packet size of 100 bytes does notsaturate the capacity of the access links, the main performance bottleneck of this configurationis represented by the buffering of the aggregated traffic on the intermediate backhaul links.Therefore, the MSR policy provides only minimal improvements compared to the performance ofthe distributed scheduler, since it simply favors the links which exhibit a higher SINR. Conversely,the prioritization of the most congested links which is introduced by the other two strategiessuccessfully tackles the former problem. In particular, the BA policy exhibits the highest worstcase throughput, albeit at the cost of satisfying the QoS requirements of approximately 20% ofthe UEs. On the other hand, the bias towards high SINR channels introduced by the MRBAstrategy has the opposite effect, improving mostly the higher percentiles but also outperformingMSR and the baseline in the lower percentiles.By increasing the UDP packet size to 500 bytes, the network becomes noticeably saturated,as depicted by Fig. 5b; in fact, in this instance only a minority of the UEs achieves a throughputwhich is comparable to the source rate. With this configuration, the BA strategy achieves the worstperformance, providing a significantly lower throughput across all the percentiles. On the otherhand, the differences among the behavior of the remaining strategies are less evident. In particular,the MSR policy exhibits only a slight improvement over the distributed solution, albeit noticeableacross the whole ECDF. The MRBA, conversely, introduces performance benefits which mostlyaffect the bottom percentiles only. However, with this strategy only a limited portion of theUEs achieves the target throughput of 40 Mbps. As a consequence, we can conclude that withthe configuration depicted in Fig. 5b the network is approaching the capacity of the mmWavechannels. Therefore, buffering phenomena are likely occurring at each intermediate IAB-node.
100 200 300 400 5002468
Packet size [B] E E t h r oughpu t [ M bp s ] Distr MSRBA MRBA (a) First quartile.
100 200 300 400 500510152025
Packet size [B] E E t h r oughpu t [ M bp s ] Distr MSRBA MRBA (b) Third quartile.Figure 6. End-to-end throughput quartiles, for s UDP ∈ { , , , } B. Moreover, we can say that in a saturated network the congestion is so severe that prioritizingthe bottleneck links is not enough: we also need to take into account the channel conditions andprioritize the links which not only are congested, but also have the “biggest chance” of gettingrid of the buffered data due to the temporary better channel quality.Finally, Fig. 6 presents the first and third quartiles of S APP as a function of the UDP packetsize s UDP . It can be noted that, with respect to the first quartile, the MRBA outperforms allthe other policies by delivering a throughput which is up to 90% higher than the one obtainedby the distributed scheduler. In fact, Fig. 6b shows how MRBA achieves also the best thirdquartile, albeit the improvement over the distributed solution is less dramatic. Furthermore, wecan observe how the positive impact of the BA strategy is inversely proportional to the saturationin the network. We can then conclude that the bias it introduces loses its effectiveness as thebuffering phenomena start to affect the majority of the IAB-nodes.
C. Latency
Just like the aforementioned metric, the latency is measured end-to-end at the application layer.Thanks to this choice, the resulting delay accurately represents the system-level performance, asit includes the latency which is introduced at each hop in the IAB network.In particular, for each packet correctly received at the uppermost layer of its final destination,the following quantity is traced: D APP i ≡ (cid:88) l k ∈ E i D l k i where E i comprises the links in the IAB network that are crossed by the i -th packet, whilethe term D l k i indicates its point-to-point latency over the path link l i . Finally, these values are ,
000 1 ,
500 2 ,
000 2 ,
500 3 ,
000 3 , . . . . Per UE delay [ms]
DistrMSRBAMRBA . . . (a) ECDF, for s UDP = 100 B.
100 200 300 400 5005001 , , , Packet size [B] E E d e l a y [ m s ] Distr MSRBA MRBA (b) Third quartile, for s UDP ∈ { , , , } B.Figure 7. Per-UE end-to-end delay statistics. collected for each of the various runs into the vector D APP and its statistical properties areinspected.Fig. 7a shows the empirical ECDF of D APP for a packet size of 100 bytes. It can benoticed that, in this case, the 90th percentile achieved by the BA and the MRBA policies areapproximately 20 % smaller than the one obtained by the distributed scheduler. Moreover, thesestrategies manage to dramatically reduce the number of packets received with extremely highdelay, i.e., in the order of seconds, showing the dramatic impact of buffering in the baselineconfiguration. Conversely, the MSR policy provides the best performance with respect to thebest case delay only, although it still outperforms quite significantly the distributed strategy.These trends are exacerbated by Fig. 7b, which shows the third quartile of D APP as a functionof the UDP packet size s UDP . In fact, we can notice that the effectiveness of the BA policyis inversely proportional to the network saturation; the opposite holds true with respect to theMSR strategy. It follows that, for UDP rates in the order of 5 to 10 Mbps, the network is mainlyplagued by local congestion which causes the insurgence of buffering in some of the nodes.Conversely, as the rate of the UDP sources increases the system shifts to a capacity-limitedregime, a phenomenon which explains the dominance of the MSR and MRBA policies.
D. Network congestion
The network congestion is measured by collecting, every T alloc subframes, the RLC buffersstatus of the various nodes into the vector B RLC . It must be noted that, since RLC AcknowledgedMode (AM) is used, these values will indicate data which is related to both new packets andpossible retransmissions.
100 200 300 400 50000 . . · Packet size [B] R L C bu ff e r[ B ] (a) Medians, toward UEs.
100 200 300 400 50001234 · Packet size [B] R L C bu ff e r[ B ] Distr MSRBA MRBA (b) Medians, toward IAB-nodes. · Node depth R L C bu ff e r[ B ] · (c) Third quartile vs. depth in the IABnetwork, for s UDP = 200 B.Figure 8. Buffer occupancy statistics, for s UDP ∈ { , , , } B. Figs. 8a and 8b show the median of B RLC , for traffic flows whose next hop in the networkis represented by either UEs or IAB-nodes respectively. The BA strategy achieves the worstperformance in this metric, leading to unstable systems in the cases of s UDP = { } B. Areason for this behavior can be found in the “locality” of the BA policy criteria and the lack ofinfluence of the past allocations on the weights. These characteristics may lead to favoring thesame link in a repeated manner, hence offering little remedy to the end-to-end congestion. Onthe other hand, the buffer occupancy achieved by the MSR strategy depicts a system behaviorwhich, in accordance with previous observations, is extremely similar to that of the distributedcase. Interestingly, with these configurations the network congestion occurs primarily at the donorand, in general, on the backhaul links towards IAB-nodes. This phenomenon can be explained asfollows: even though, on average, the channel qualities of the backhaul links experience a betterSINR, the maximum number of such links which can be concurrently activated is limited, dueto the TDD configuration. Therefore, the MSR policy may introduce a bias towards the accesslinks instead, since their activation yields the highest sum capacity, despite their lower channelquality. Finally, the MRBA policy achieves the lowest amount of RLC buffering. Specifically,Fig. 8b shows that, compared to the MSR and distributed strategies, the median buffer occupancyamong backhaul links is up to 60% smaller, albeit at the cost of slightly more congested UEbuffers.Finally, Fig. 8c depicts the third quartiles of B RLC as a function of the depth of the corre-sponding gNB in the IAB network. It is possible to notice that, regardless of the policy in use,
500 1 ,
000 1 ,
500 2 , . . . . Per UE delay [ms]
Distr MSRBA MRBA (a) Delay ECDF.
50 100 15000 . . . . Per UE throughput [Mbps]
Distr MSRBA MRBA (b) Throughput ECDF.Figure 9. End-to-end delay and throughput statistics, for TCP layer 4 protocol. the amount of buffering at the various gNBs generally decreases as their distance to the donorincreases. This follows from the fact that nodes which have a lower depth exhibit, on average, abigger subtending tree; therefore the amount of traffic which makes use of their backhaul linksis significantly higher.
E. Performance with TCP traffic
This subsection extends the aforementioned analysis by inspecting the performance of theproposed scheme in the case of TCP traffic. Specifically, a TCP full-buffer source model isused, and the various centralized resource allocation policies are compared against the distributedscheduler.Fig. 9a shows the ECDF of the end-to-end delay experienced by the successfully receivedpackets. Similarly to the UDP case, the distributed scheduler exhibits the worse performance inthis regard. However, the behavior of the centralized policies is remarkably different. In particular,with this configuration the MSR policy provides the best results, followed quite closely by theMRBA and BA strategies. Fig. 9b, which depicts the statistics of the end-to-end throughputachieved by the various UEs, helps explain these results. The net effect of the BA and MSRpolicies is, approximately, a 20% increase of the peak throughput. Conversely, the MRBA strategycauses a redistribution of the achieved data rate, massively improving the lower quartiles (up tothe 80-th), albeit at the expense of the maximum throughput.Therefore, we can conclude that regardless of the specific policies used, the proposed schemeimproves the system performance by limiting the insurgence of local buffering, aiding the end-to-end congestion control mechanism offered by TCP. Furthermore, it can be noted that both a Per UE throughput [Mbps] D e l a y [ m s ] Distr MSR BA MRBA4 Subframes2 Subframes1 Subframe(a) Combined per UE end-to-end throughput first quartile and end-to-end delay third quartile, as a function of the centralized allocation period T alloc . Number of IAB-nodes ∈ V R un ti m e [ µ s ] (b) MWM runtime as a function of the number ofIAB-nodes in the network.Figure 10. Considerations on the formulated assumptions. prioritization of the most congested links and of the channels featuring a higher quality resultsin performance benefits in the average case, although it also causes a decrease of the networkfairness. On the other hand, the MRBA policy manages to optimize the backhaul/access resourcepartitioning, while introducing an increase in the throughput fairness at the same time. F. Further considerations
It is of particular relevance to analyze the performance of the centralized policies whenrelaxing the most restrictive hypothesis, i.e., the capability of reliably exchanging feedbackinformation in a timely manner, and to understand how restrictive such assumption actuallyis. To such end, Fig. 10a shows the performance of the proposed framework as a function ofthe centralized allocation period T alloc . In particular, each of the depicted points represents thejoint end-to-end throughput and delay achieved with the different configurations. As expected,in general the effectiveness of the various centralized policies progressively deteriorates as thefrequency of the scheduling indications decreases. Interestingly, the BA policy exhibits the lowestperformance degradation with respect to an increase of the allocation period, which suggeststhat this phenomenon has a slower evolution over time compared to the one exhibited by thechannels quality. Nevertheless, the key takeaway is that all of the proposed allocation strategiesoutperform the distributed solution, across both metrics. However, the trend depicted by Fig. 10aalso suggests that there exists a threshold value of T alloc after which the performance of theproposed frameworks brings only marginal performance benefits. Additionally, the running time of the MWM algorithm presented in Sec. III-A was analyzed,in order to understand whether it may partially invalidate the timely feedback assumption.Specifically, Fig. 10b presents the statistics of the various MWM execution times, obtained ona machine equipped with an i7-6700 4-core processor clocked at 3.4 GHz. The first observationwhich can be made is that this empirical analysis confirms the previously estimated asymptoticcomplexity, depicting a running time which exhibits a linear dependence on the number of gNBsin the network. Furthermore, it can be noted that the runtime of the MWM algorithm does notexceed 6 µ s, even for a significant number of IAB-nodes connected to the same IAB-donor. Asa consequence, we can conclude that the execution times of the centralized allocation processdo not pose any threat to the timely feedback assumption, since they are reasonably smaller thanthe duration of the minimum centralized allocation period.V. C ONCLUSIONS
In this paper we proposed a centralized resource partitioning scheme for 5G and beyondIAB networks, coupled with a set of allocation policies. We showed that the introduction ofthis light resource allocation cooperation dramatically improves the end-to-end throughput anddelay achieved by the system already, preventing (or at the very least limiting) the insurgenceof network congestion on the backhaul links. Specifically, the MRBA policy exhibits the mostpromising results, offering up to a 5-fold increase in the worst-case throughput and approximatelya 50% smaller worst-case latency, compared to the distributed scheduler. On the other hand, theeffectiveness of the BA and MSR policies varies quite significantly across the specific systemconfiguration and inspected metric.We provided considerations on the implementation of a centralized resource allocation con-troller in real world deployments. In particular, we acknowledged that the proposed scheme relieson the assumption of IAB-nodes being capable of exchanging timely feedback information withthe IAB-donor. Even though the amount of signaling data which the proposed solution requiresis quite low, and its performance is quite robust with respect to an increase of the centralallocation period, we argue that this remains a significant constraint. Moreover, such drawbackis exacerbated by the unfavorable mmWaves propagation characteristics. As a consequence, wedeem that centralized solutions, which rely on timely exchange of control information with theIAB-donor, are likely to require dedicated control channels, possibly at sub-6 GHz, in order togrant the utmost priority and reliability to the feedback information. Therefore, we can conclude that the aforementioned framework can bring dramatic performance benefits to IAB networks,although its introduction in 5G and beyond deployments requires additional research efforts.For this reason, as part of our future work we plan to design machine-learning algorithms whichpredict the network evolution at the IAB-donor. This improvement will allow us to relax thetimely feedback assumption, by increasing the minimum centralized allocation period which leadsto performance benefits over distributed strategies. Moreover, we foresee to implement mecha-nisms which adapt the parameters of the MRBA policy to the system load and configuration, andadditional resource partitioning strategies. Finally, the generalization of the proposed frameworkto SDMA systems will be studied. The use of such multiple access scheme should significantlyimprove the performance of mmWave wireless backhauling by introducing the possibility ofconcurrently serving multiple terminals, provided that they exhibit a sufficient distance amongthem. R EFERENCES [1] ITU-R, “IMT Vision – Framework and overall objectives of the future development of IMT for 2020 and beyond,” Sep.2015.[2] 3GPP, “NR; NR and NG-RAN Overall description; Stage-2,” 3rd Generation Partnership Project (3GPP), TechnicalSpecification (TS) 38.300, Jul. 2020, v16.2.0.[3] F. Z. Yousaf, M. Bredel, S. Schaller, and F. Schneider, “NFV and SDN—Key technology enablers for 5G networks,”
IEEEJournal on Selected Areas in Communications , vol. 35, no. 11, pp. 2468–2478, Nov. 2017.[4] F. Khan and Z. Pi, “mmWave mobile broadband (MMB): Unleashing the 3–300 GHz spectrum,” in , Princeton, NJ, USA, 2011.[5] S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter-Wave Cellular Wireless Networks: Potentials and Challenges,”
Proceedings of the IEEE , vol. 102, no. 3, pp. 366–385, March 2014.[6] I. A. Hemadeh, K. Satyanarayana, M. El-Hajjar, and L. Hanzo, “Millimeter-Wave Communications: Physical ChannelModels, Design Considerations, Antenna Constructions, and Link-Budget,”
IEEE Communications Surveys & Tutorials ,vol. 20, no. 2, pp. 870–913, Dec. 2017.[7] F. G´omez-Cuba, E. Erkip, S. Rangan, and F. J. Gonz´alez-Casta˜no, “Capacity Scaling of Cellular Networks: Impact ofBandwidth, Infrastructure Density and Number of Antennas,”
IEEE Transactions on Wireless Communications , vol. 17,no. 1, pp. 652–666, Nov. 2017.[8] M. Polese, M. Giordani, T. Zugno, A. Roy, S. Goyal, D. Castor, and M. Zorzi, “Integrated Access and Backhaul in 5GmmWave Networks: Potential and Challenges,”
IEEE Communications Magazine , vol. 58, no. 3, pp. 62–68, Mar. 2020.[9] 3GPP, “NR; Study on integrated access and backhaul,” Technical Specification (TS) 38.874, Jan. 2019, v.16.0.0.[10] C. Saha and H. S. Dhillon, “Millimeter Wave Integrated Access and Backhaul in 5G: Performance Analysis and DesignInsights,”
IEEE Journal on Selected Areas in Communications , vol. 37, no. 12, pp. 2669–2684, Dec. 2019.[11] M. Polese, M. Giordani, A. Roy, S. Goyal, D. Castor, and M. Zorzi, “End-to-End Simulation of Integrated Access andBackhaul at mmWaves,” in
IEEE 23rd International Workshop on Computer Aided Modeling and Design of CommunicationLinks and Networks (CAMAD) , Barcelona, Spain, 2018. [12] M. Mezzavilla, M. Zhang, M. Polese, R. Ford, S. Dutta, S. Rangan, and M. Zorzi, “End-to-End Simulation of 5G mmWaveNetworks,” IEEE Communications Surveys & Tutorials , vol. 20, no. 3, pp. 2237–2263, Third quarter 2018.[13] V. Gambiroza, B. Sadeghi, and E. W. Knightly, “End-to-end performance and fairness in multihop wireless backhaulnetworks,” in
Proceedings of the 10th Annual International Conference on Mobile Computing and Networking (MOBICOM) ,Philadelphia, PA, USA, 2004, pp. 287–301.[14] M. N. Islam, S. Subramanian, and A. Sampath, “Integrated Access Backhaul in Millimeter Wave Networks,” in , San Francisco, CA, USA, 2017, pp. 1–6.[15] M. N. Islam, N. Abedini, G. Hampel, S. Subramanian, and J. Li, “Investigation of performance in integrated access andbackhaul networks,” in
IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS) , Honolulu, HI,USA, 2018.[16] M. N. Kulkarni, A. Ghosh, and J. G. Andrews, “Max-min rates in self-backhauled millimeter wave cellular networks,” arXiv preprint arXiv:1805.01040 , 2018.[17] W. Lei, Y. Ye, and M. Xiao, “Deep reinforcement learning-based spectrum allocation in integrated access and backhaulnetworks,”
IEEE Transactions on Cognitive Communications and Networking , vol. 6, no. 3, pp. 970–979, September 2020.[18] M. E. Rasekh, D. Guo, and U. Madhow, “Interference-aware routing and spectrum allocation for millimeter wave backhaulin urban picocells,” in , Monticello, IL,USA, 2015.[19] M. Bilal, M. Kang, S. C. Shah, and S.-G. Kang, “Time-Slotted Scheduling Schemes for Multi-hop Concurrent Transmissionin WPANs with Directional Antenna,”
ETRI Journal , vol. 36, no. 3, pp. 374–384, Jun. 2014.[20] R. L. Cruz and A. V. Santhanam, “Optimal routing, link scheduling and power control in multihop wireless networks,” in , San Francisco,CA, USA, 2003.[21] C. Saha, M. Afshang, and H. S. Dhillon, “Bandwidth partitioning and downlink analysis in millimeter wave integratedaccess and backhaul for 5G,”
IEEE Transactions on Wireless Communications , vol. 17, no. 12, pp. 8195–8210, Dec. 2018.[22] R. Singh and P. Kumar, “Throughput optimal decentralized scheduling of multihop networks with end-to-end deadlineconstraints: Unreliable links,”
IEEE Transactions on Automatic Control , vol. 64, no. 1, pp. 127–142, Jan. 2019.[23] B. Ji, C. Joo, and N. Shroff, “Throughput-optimal scheduling in multihop wireless networks without per-flow information,”
IEEE/ACM Transactions on Networking , vol. 21, no. 2, pp. 634–647, Apr. 2013.[24] T. K. Vu, C.-F. Liu, M. Bennis, M. Debbah, and M. Latva-Aho, “Path selection and rate allocation in self-backhauledmmWave networks,” in
IEEE Wireless Communications and Networking Conference (WCNC) , Barcelona, Spain, 2018.[25] J. Garc´ıa-Rois, F. G´omez-Cuba, M. R. Akdeniz, F. J. Gonz´alez-Casta˜no, J. C. Burguillo, S. Rangan, and B. Lorenzo,“On the analysis of scheduling in dynamic duplex multihop mmWave cellular systems,”
IEEE Transactions on WirelessCommunications , vol. 14, no. 11, pp. 6028–6042, Nov. 2015.[26] F. Gomez-Cuba and M. Zorzi, “Optimal link scheduling in millimeter wave multi-hop networks with space division multipleaccess,” in , La Jolla, CA, USA, 2016.[27] F. G´omez-Cuba and M. Zorzi, “Optimal Link Scheduling in Millimeter Wave Multi-hop Networks with MU-MIMO radios.”
IEEE Transactions on Wireless Communications , vol. 19, no. 3, pp. 1839–1854, Mar. 2020.[28] L. Tassiulas and A. Ephremides, “Stability properties of constrained queueing systems and scheduling policies for maximumthroughput in multihop radio networks,” in , Honolulu, HI, USA, 1990.[29] F. Kelly, “Charging and rate control for elastic traffic,”
European Transactions on Telecommunications , vol. 8, no. 1, pp.33–37, Jan. 1997. [30] F. P. Kelly, A. K. Maulloo, and D. K. Tan, “Rate control for communication networks: shadow prices, proportional fairnessand stability,” Journal of the Operational Research society , vol. 49, no. 3, pp. 237–252, Apr. 1998.[31] M. Polese, M. Giordani, A. Roy, D. Castor, and M. Zorzi, “Distributed Path Selection Strategies for Integrated Access andBackhaul at mmWaves,” in
IEEE Global Communications Conference (GLOBECOM) , Abu Dhabi, United Arab Emirates,2018.[32] 3GPP, “NR; Integrated Access and Backhaul (IAB) radio transmission and reception,” Technical Specification (TS) 38.174,Jun. 2020, v0.1.0.[33] S. Dutta, M. Mezzavilla, R. Ford, M. Zhang, S. Rangan, and M. Zorzi, “Frame structure design and analysis for millimeterwave cellular systems,”
IEEE Transactions on Wireless Communications , vol. 16, no. 3, pp. 1508–1522, Mar. 2017.[34] B. Korte and J. Vygen,
Combinatorial Optimization . Springer Berlin Heidelberg, 2002.[35] H. N. Gabow, “Data structures for weighted matching and nearest common ancestors with linking,” in
Proceedings of theFirst Annual ACM-SIAM Symposium on Discrete Algorithms , San Francisco, California, USA, 1990.[36] L. Bonati, M. Polese, S. D’Oro, S. Basagni, and T. Melodia, “Open, programmable, and virtualized 5G networks: State-of-the-art and the road ahead,”
Computer Networks (COMNET) , vol. 182, Aug. 2020.[37] 3GPP, “NR; Medium Access Control (MAC) protocol specification,” Technical Specification (TS) 38.321, Jul. 2020, v16.1.0.[38] ——, “NR; Radio Resource Control (RRC); Protocol specification,” Technical Specification (TS) 38.331, Apr. 2020,v16.0.0.[39] N. Baldo, M. Miozzo, M. Requena-Esteso, and J. Nin-Guerrero, “An open source product-oriented LTE network simulatorbased on ns-3,” in
Proceedings of the 14th ACM international conference on Modeling, analysis and simulation of wirelessand mobile systems , Miami, Florida, USA, 2011.[40] T. Zugno, M. Polese, N. Patriciello, B. Bojovi´c, S. Lagen, and M. Zorzi, “Implementation of a spatial channel model forns-3,” in