Traffic-aware ONU grouping and downstream bandwidth allocation for multicast services in flexible-rate PONs

Research Article Vol. 17, No. 9 / September 2025 / Journal of Optical Communications and Networking 757 Traffic-aware ONU grouping and downstream bandwidth allocation for multicast services in flexible-rate PONs Xiang Lu,1,2Xinshui Wei,1Zhiyuan Zhong,1Jun Li,1,*AND Luis Velasco2 1School of Electronic and Information, Soochow University, Suzhou 215031, China 2Advanced Broadband Communications Center (CCABA), Universitat Politècnica de Catalunya (UPC), 08034 Barcelona, Spain *[email protected] Received 22 March 2025; revised 20 July 2025; accepted 21 July 2025; published 18 August 2025 In flexible-rate passive optical networks (PONs), optical network units (ONUs) with different channel conditions can achieve different data rates by implementing flexible transmission parameters, e.g., modulation format, thereby enhancing system capacity. Correspondingly, the downstream frame is divided into multiple subframes, each being received and processed only by ONUs in its targeted group. However, for downstream multicast services, the data need to be duplicated and encapsulated into multiple subframes for ONUs belonging to different groups, resulting in data redundancy and degradation of effective throughput (i.e., throughput without redundant data). To improve resource utilization, ONU grouping should be dynamically adjusted according to both channel conditions and time-varying network factors, e.g., traffic loads and multicast memberships. For this, we first enhance the current downstream scheduling protocol to support dynamic ONU grouping in a multicast scenario and propose a traffic-aware ONU grouping (TAOG) algorithm to improve effective throughput, which optimizes ONU grouping by considering time-varying network conditions. As ONUs belonging to different groups have different data rates, we further propose a group-based downstream time slot allocation (GBDTA) algorithm to adjust time slots for each service by considering their demands and ONU data rates. Exhaustive simulation results show that the integrated TAOG-GBDTA scheme adapts effectively to dynamic network conditions and, compared to conventional schemes, it effectively improves effective throughput, reduces redundancy, and achieves lower packet latency under various multicast scenarios. © 2025 Optica Publishing Group. All rights, including for text and data mining (TDM), Artificial Intelligence (AI) training, and similar technologies, are reserved. https://doi.org/10.1364/JOCN.562901 1. INTRODUCTION Time division multiplexing-passive optical networks (TDMPONs), operating on a point-to-multipoint (P2MP) architecture, are among the most commonly-used access network technologies. Notably, International Telecommunication Union-Telecommunication Standardization Sector (ITU-T) PONs, e.g., Gigabit PON (GPON) and XG(S)-PON, are widely deployed, especially in the scenario of fiber to the home (FTTH) [1]. To meet the stringent quality of service (QoS) requirements of emerging new broadband services, ITU-T has standardized the next-generation higher speed TDM-PON [2–7]. In a typical ITU-T TDM-PON, a single optical line terminal (OLT) is connected to multiple optical network units (ONUs) via an optical distribution network (ODN). In the upstream direction, each ONU transmits its data to the OLT in a burst mode within allocated time slots. In the downstream direction, the OLT aggregates data for all ONUs into a frame with a fixed length of 125 µs, referred to as the downstream PHY frame, and broadcasts it continuously to all ONUs. Each ONU identifies and retrieves its data from the downstream PHY frame. Furthermore, both upstream and downstream transmissions operate at a uniform, constant peak data rate across all ONUs [8,9]. In this regard, these TDM-PONs are categorized as fixed-rate PONs. However, the channel conditions between ONUs and the OLT can vary significantly due to optical path loss (OPL), influenced by factors such as the physical distance and the number of splitters [10–13]. In a fixed-rate PON, the constant peak data rate is constrained by the worst-case optical budget, rather than considering the actual channel conditions experienced by each ONU [10,13]. As a result, ONUs with better channel conditions operate below their potential capacity, thereby limiting the overall system capacity. To address this issue, a promising approach is to differentiate the peak data rates of ONUs by allowing ONUs with better channel conditions to operate at higher data rates while maintaining compatibility with those under poorer conditions. TDMPONs that allow ONUs to operate at varying peak data rates 1943-0620/25/090757-14 Journal © 2025 Optica Publishing Group Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on October 06,2025 at 09:44:35 UTC from IEEE Xplore. Restrictions apply. 758 Vol. 17, No. 9 / September 2025 / Journal of Optical Communications and Networking Research Article are referred to as flexible-rate PONs [12], and this concept is also under discussion in the ongoing work of ITU-T Study Group 15 (SG15) on Very High Speed PON [14]. In flexible-rate PONs, ONUs with similar channel conditions can be grouped to share the same transmission parameters, e.g., modulation formats and forward error correction (FEC) coding, enabling them to operate at a specific peak data rate. Furthermore, the peak data rate can be flexibly adjusted by modifying transmission parameters within the constraint of the power budget [11–20]. Upstream multi-rate reception is supported not only in multi-generation PON coexistence systems [21,22] but also natively in advanced standards such as NG-PON2 and 50 G-PON [4,5]. Since the OLT can receive data encoded with different modulation formats, the traditional upstream scheduling protocol can still be used after introducing flexible rates [23]. Notably, the differences in peak data rates among ONUs must be considered during time slot allocation to ensure efficient scheduling [15,24]. In contrast, in the downstream direction, since ONUs may employ different transmission parameters, the traditional broadcast-based downstream scheduling protocol in the fixed-rate PON becomes ineffective. To address this, the downstream PHY frame is divided into multiple subframes, each occupying multiple specific time slots [10,16]. The subframe differs in modulation and coding parameters, which are assigned to a group of ONUs with similar channel conditions and data rate requirements. Different from fixed-rate PONs, where ONUs need to process the whole downstream frame, ONUs in flexible-rate PONs only receive and process their own downstream subframe to reduce computing complexity and energy consumption [10,18,25]. Compared to fixed-rate PONs, flexible-rate PONs have greater potential for improving throughput efficiency by exploiting available link margin. Moreover, they reduce power consumption and lower ONU computational complexity, and thus are worthy of further investigation. Despite these advantages, flexible-rate PONs still face several open issues, such as the challenge of maintaining stable downstream synchronization under flexible modulation and the potential increase in system cost. Among these, downstream scheduling remains a key challenge, especially in multicast scenarios. On the one hand, multicast scenarios introduce additional redundancy into the current scheduling mechanisms, thereby impacting system performance. The growing demand for high-bandwidth video services, e.g., live video streaming, personalized short-form videos, and ultra-high-definition video-on-demand services, has significantly increased multicast traffic, further straining scheduling efficiency [26,27]. Based on broadcast mode, downstream services can be classified into unicast and multicast services. Unicast services, e.g., video-on-demand (VoD) streams, require dedicated data flows for individual user requests to ensure personalized content delivery. Thus, a unicast traffic flow is sent exclusively to its corresponding ONU in a fixed-rate PON. In contrast, a traffic flow from a multicast service, e.g., live streaming, may be broadcast simultaneously to multiple ONUs. The set of ONU members associated with a specific multicast service is referred to as its multicast membership. Nowadays, multicast delivery has already been implemented in FTTx broadband access networks [28–30]. In GPON and EPON systems, patching-based multicast mechanisms are commonly used: the first user initiates a multicast stream, and subsequent users receive the ongoing content via multicast and the missed portion through short unicast patches [28]. This method significantly reduces redundant transmissions and improves bandwidth utilization. Moreover, these multicast services often follow a Zipf distribution, where a small number of highly popular contents account for a large portion of total traffic. As a result, operators can pre-cache these popular contents in user-side buffers during off-peak hours, thereby improving delivery efficiency [29]. In addition, for live streaming scenarios, an SDN-aided NG-EPON architecture has been proposed to dynamically manage multicast groups based on user viewports and QoS demands [30]. However, in flexible-rate PONs, each ONU group can only process its own subframe. Consequently, when the memberships of a multicast service are distributed across multiple ONU groups, the multicast data must be duplicated and encapsulated into separate subframes. The amount of duplicated data increases with the number of ONU groups associated with the service, leading to significant redundancy and limiting effective throughput. A straightforward approach to reducing redundancy is to minimize the number of ONU groups. However, this may lower the data rates of ONUs with better channel conditions, which reduces the overall system capacity. Consequently, despite the decrease in redundancy, effective throughput may still degrade. Therefore, optimizing ONU grouping requires balancing the trade-off between minimizing redundancy and maintaining high data rates to maximize overall effective throughput. On the other hand, in multicast scenarios, allocating downstream bandwidth to satisfy the bandwidth requirements of services with different broadcast modes introduces additional challenges. In [31], a session-based priority bandwidth allocation mechanism was proposed to dynamically allocate bandwidth among ONUs in TDM-PONs, addressing inefficiencies caused by multiple ONUs requesting the same multicast content. On this basis, Kim et al. [32] proposed a multicast-share weighted fair queuing (MS-WFQ) mechanism to target inter-session fairness, which dynamically adjusts the service weights of multicast services to optimize throughput while maintaining fairness across these services. However, such schemes are difficult to apply in flexible-rate PONs that employ ONU grouping methods, as they do not account for multicast redundancy and the varying data rates of ONUs. Consequently, they struggle to meet the requirements of services, which are crucial for maintaining high effective throughput in a flexible-rate PON. Therefore, downstream bandwidth allocation schemes should be enhanced to consider both multicast redundancy and data rate differences, thereby fully meeting service requirements and improving scheduling efficiency. To address these challenges, this paper first enhances the downstream scheduling protocol for flexible-rate PONs to support multicast scenarios. Given the limitations of existing ONU grouping and bandwidth allocation schemes, we introduce a novel downstream traffic-aware ONU grouping (TAOG) algorithm, complemented by a grouping-based downstream time slot allocation (GBDTA) algorithm. Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on October 06,2025 at 09:44:35 UTC from IEEE Xplore. Restrictions apply. Research Article Vol. 17, No. 9 / September 2025 / Journal of Optical Communications and Networking 759 Simulation results show that the combination of the proposed algorithms, referred to as the TAOG-GBDTA scheme, can effectively support multicast scenarios and significantly improve system performance in terms of effective throughput and latency. The rest of the paper is organized as follows. Section 2 introduces an enhanced downstream scheduling protocol and analyzes the effective throughput for flexible-rate PONs. Section 3presents the proposed TAOG-GBDTA scheme based on the proposed scheduling protocol. Section 4presents and analyzes the simulation results for the proposed algorithm. Finally, Section 5concludes the paper. 2. FLEXIBLE-RATE PON ARCHITECTURE AND PROTOCOL In this section, we first present a representative architecture of a flexible-rate PON. To support dynamic ONU grouping in a multicast scenario, an enhanced downstream scheduling protocol is proposed. Then, we further analyze the downstream effective throughput in a flexible-rate PON. A. Architecture Figure 1illustrates the high-level architecture of a flexible-rate PON, which consists of an OLT and multiple ONUs. The ONUs adopt different transmission parameters, e.g., modulation format, according to their OPL, which is influenced by factors such as physical fiber distance and the number of optical splitters along the ODN. For example, ONU1in Fig. 1is connected to the OLT via a single optical splitter, thus experiencing a low OPL, which enables utilizing the 4-level pulse amplitude modulation (PAM4) format to achieve higher data rates. In contrast, ONU2and ONU3are connected via two optical splitters, hence experiencing a higher OPL. Therefore, they employ the non-return-to-zero (NRZ) modulation format to maintain reliable performance under more challenging channel conditions. Based on their OPL, these ONUs are grouped into two groups: ONU1belongs to Group 1, while ONU2and ONU3belong to Group 2. However, the data rates for these ONUs can be adjusted dynamically, and the group assignments may change accordingly. To support this flexibility, digital signal processing (DSP) modules are introduced in both the OLT and the ONUs, enabling modulation format adaptation during transmission and reception. The operation of the DSP modules is managed by the PON transmission convergence (TC) layer, which consists of the service adaptation sublayer, the framing sublayer, and the PHY adaptation sublayer. These sublayers handle framing, scheduling, and group management, among others. In the downstream transmission, the OLT divides the frame into multiple forward error correction (FEC) codewords, i.e., data blocks encoded with FEC for error resilience, customized for different ONU groups. For instance, FEC codewords modulated with PAM4 are designated for Group 1, while those modulated with NRZ are directed to Group 2. Upon receiving the downstream frame, each ONU identifies the FEC codewords corresponding to its modulation format and reconstructs its assigned subframe. For example, Group 1 only processes the PAM4-modulated codewords (yellow blocks), while Group 2 processes the NRZ-modulated codewords (green blocks). This ensures that each ONU group handles only its designated data, thereby reducing DSP complexity. The flexible-rate PON architecture determines a special downstream scheduling scheme that is distinct from that of conventional fixed-rate PONs. The downstream scheduling protocol plays a critical role in implementing dynamic ONU grouping, especially in multicast scenarios. In the next subsection, we introduce the downstream scheduling protocol for fixed-rate PONs, followed by our proposed enhanced downstream scheduling protocol tailored for flexible-rate PONs. Tx( ) OLT: Optical line terminal ONU: Optical network unit Tx: Transmitter Rx: Receiver DSP: Digital signal processing TC: Transmission convergence Rx( ) Rx( ) ComTC Framing sublayer PHY adaptation sublayer Service adaptation sublayer Tx 4:1 Splitter Rx PON TC Framing sublayer PHY adaptation sublayer Service adaptation sublayer Filter DSP OLT 4:1 Splitter time rate NRZ NRZ time rate PAM4 PAM4 time rate NRZ NRZ PAM4 PAM4 Downstream PHY frame Rx Tx ONU1 PON TC DSP (PAM4) ONU 2 ONU 3 Tx Rx PON TC DSP (NRZ) Group 1 Group 2 Fig. 1. High-level architecture of a flexible-rate PON. Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on October 06,2025 at 09:44:35 UTC from IEEE Xplore. Restrictions apply. 760 Vol. 17, No. 9 / September 2025 / Journal of Optical Communications and Networking Research Article (b) Flexible-rate PON OLT TC layer ONU TC layer (Group 1: ONU 1) SA sublayer PHY sublayer SA sublayer FS sublayer PHY sublayer XGEM frames FS frame 1 FS frame 2 DS PHY frame FS frame 2 (a) Fixed-rate PON OLT TC layer ONU TC layer SA sublayer FS sublayer PHY sublayer SA sublayer FS sublayer PHY sublayer XGEM frames FS frame DS PHY frame FS frame U 1 M 2 U 2 FS Header (BWmap + PLOAMd) HM 1 U 1 M 2 U 2 FS sublayer DS bandwidth allocation ONU grouping and bandwidth allocation Traffic scheduler Traffic scheduler DS PHY frame Traffic classifier Traffic classifier HM 1 U 1 M 2 U 2 P H 1 M 1 U 1 PSBd H M 1 U 1 M 2 U 2 PHM 1 U 1 M 2 U 2 M 1 U 1 DS PHY frame 1 DS PHY frame 2 Traffic classifier M 1 U 2 FS frame 1 ONU TC layer (Group 2: ONU 2 and 3) M 1 U 1 M 2 U 2 M 2 MT 2 UT 2 MT 1 UT 1 Upstream scheduling and data rate adjusting OLT: Optical line terminal ONU: Optical network unit TC: Transmission convergence MT: Multicast traffic UT: Unicast traffic SA: Service adaptation FS: Framing DS: Downstream PHY: Physical interface adaptation BWmap: Bandwidth map PSBd: Physical synchronization block downstream PLOAMd: Physical layer operation, administration and maintenance downstream 1 3 2 3 5 46 7 8 5 4 67 8 M 1 U 1 M 2 U 2 MT 2 UT 2 MT 1 UT 1 M 1 1 2 Upstream scheduling Upstream scheduling and data rate adjusting UT 3 (OMCI) U 3 U 3 U 3 U 3 U 3 U 3 UT 3 (OMCI) U 3 H 2 M 1 M 2 U 2 U 3 H 2 H 1 M 1 U 1 U 2 PM 1 M 2 U 3 H 1 M 1 U 1 P H 2 U 2 PM 1 M 2 U 3 H 2 U 2 M 1 M 2 U 3 H 1 M 1 U 1 U 3 FS Header-G1 FS Header-G2 ONU 2&3 ONU 2 ONU 1&2 ONU 1 ONU 2&3 ONU 2 ONU 1&2 ONU 1 ONU 2 ONU 2 Fig. 2. (a) Downstream scheduling protocol in a fixed-rate PON; (b) enhanced downstream scheduling protocol in a flexible-rate PON. B. Enhanced Downstream Scheduling Protocol Figure 2(a) illustrates the downstream scheduling protocol according to the ITU-T 50G-PON Common TC layer specification [5]. In the OLT, service data units (SDUs), such as Ethernet frames, are first classified and added to different queues based on their service types (step 1). Then, in the service adaptation sublayer, SDUs are encapsulated into XGEM frames using an XG-PON encapsulation method and assigned a Port-ID, which is included in the header (step 2). The XGEM Port-ID identifies its corresponding service type and broadcast mode. The XGEM Port-ID of a unicast XGEM frame corresponds to a single ONU, whereas the XGEM PortID of a multicast XGEM frame can be recognized by multiple ONUs. For illustrative purposes, Fig. 2(a) shows two multicast (M1and M2) and three unicast (U1,U2, and U3) traffic flows belonging to different services. Multicast XGEM frames are marked with dashed-line blocks, while unicast XGEM frames are represented by solid-line blocks. Among these, U3represents the ONU Management and Control Interface (OMCI) traffic, which delivers configuration and management information from the OLT to each ONU individually. One function of OMCI is to assign or withdraw the logical identifiers for each ONU (XGEM Port-IDs), which allows the OLT to control whether an ONU can receive a specific unicast or multicast traffic flow. As a result, all multicast XGEM frames need to be transmitted only once, and therefore, each service requires one single dedicated time slot. In the framing sublayer, a bandwidth allocation module allocates downstream bandwidth to each traffic, determining the XGEM frames that need to be scheduled (step 3). As per the bandwidth allocation information, the traffic scheduler aggregates these XGEM frames and encapsulates them into an FS frame by adding an FS header (step 4). The FS header contains both the upstream bandwidth allocation information (i.e., the BWmap field) and the downstream physical layer operation, administration, and maintenance (PLOAMd) field for ONUs. The PLOAMd messages account for carrying management and control instructions for ONUs, such as ONU activation or ranging. Since a multicast XGEM frame can be recognized by multiple ONUs, each multicast traffic (M1and M2) is allocated one single dedicated time slot in an FS frame. Subsequently, in the PHY adaptation sublayer, this FS frame is further encapsulated into a 125 µs downstream PHY frame after FEC encoding, scrambling, and bit interleaving (step 5). After adding a downstream physical synchronization block (PSBd), the downstream PHY frame is broadcast to all ONUs in the flexible-rate PON. Upon receiving the PHY frame, each ONU synchronizes with the PSBd field and extracts the FS frame (step 6). The FS frame is then further de-encapsulated into XGEM frames at the framing sublayer, where the BWmap fields are extracted from the FS header and utilized by the ONU for upstream scheduling (step 7). A traffic classifier then Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on October 06,2025 at 09:44:35 UTC from IEEE Xplore. Restrictions apply. Research Article Vol. 17, No. 9 / September 2025 / Journal of Optical Communications and Networking 761 categorizes the data frames and selects their corresponding XGEM frames based on the XGEM Port-IDs, restoring them into SDUs in the service adaptation sublayer (step 8). Notably, in this protocol, every ONU must receive and process the entire downstream PHY frame. In contrast to the fixed-rate PON, the downstream scheduling protocol in a flexible-rate PON must account for the varying transmission parameters across ONU groups. Note that to avoid potential detection errors and limit DSP complexity, in this paper, each ONU is restricted to belong to only one group at a time in a flexible-rate PON. Consequently, data for different ONU groups are independently encapsulated according to their respective transmission parameters. To achieve this, after processing in the service adaptation sublayer, XGEM frames belonging to the same ONU group are aggregated and encapsulated into an independent FS frame. For multicast traffic, if all its ONU members belong to the same group, the corresponding FS frame can directly carry its multicast XGEM frames without duplication. Otherwise, the XGEM frames must be duplicated and encapsulated into nFS frames, where nrepresents the number of ONU groups spanned by the multicast service. This implies that a multicast service will be assigned a different time slot in multiple distinct FS frames. Figure 2(b) illustrates the enhanced downstream scheduling protocol to support multicast scenarios. Assume that ONU1belongs to Group 1, ONU2and ONU3 belong to another group. The multicast XGEM frame M1 is intended for ONU1and ONU2. Therefore, M1is duplicated and encapsulated into both FS Frame 1 and FS Frame 2 (step 4). In contrast, multicast traffic 2 has ONU members ONU2and ONU3in the same group, and thus M2is only encapsulated into FS Frame 2. Note that the duplication of XGEM frames occurs only during their encapsulation into FS frames. As a result, no additional queues are required in the OLT to store data frames for different groups, eliminating the need for modifying the queue management mechanism. Since ONU grouping is determined by transmission parameters, regular updates of these parameters are necessary to support dynamic ONU grouping. To achieve this, each FS frame includes the transmission parameter information within the PLOAMd field for the ONUs of the group periodically. This information, together with the BWmap, is generated by a novel ONU grouping and bandwidth allocation module (step 3) and copied to all FS frame headers by the traffic scheduler to ensure consistent updates across all ONU groups (step 4). We prefer the PLOAMd field rather than OMCI for delivering grouping updates, as PLOAM messages offer lower signaling latency, thereby enabling prompt ONU reconfiguration. The FS frames from different groups are subsequently aggregated in the PHY adaptation sublayer and encapsulated into a complete downstream PHY frame (step 5). During this process, a codeword interleaving method between FS frames can be performed to reduce DSP complexity [25]. Upon receiving the downstream PHY frame, the ONU reconstructs the FS frame for its group (step 6). Subsequently, the FS frame is de-encapsulated at the framing sublayer (step 7). In addition to the BWmap field used for upstream scheduling, the PLOAMd field in the FS header provides updated transmission parameters. To enable dynamic ONU grouping, the ONU uses this information to adjust its modulation format for receiving downstream data. Meanwhile, the extracted XGEM frames are processed at the service adaptation sublayer (step 8), where the traffic classifier determines whether the frames are intended for the receiving ONU. C. Downstream Effective Throughput Analysis To evaluate the potential performance improvements brought by dynamic group optimization, we further analyze the downstream effective throughput under three different ONU grouping methods: (a) a fixed-rate PON, (b) a flexible-rate PON, and (c) a flexible-rate PON with ONU grouping optimization. Figure 3(a) illustrates an example of a downstream PHY frame in a fixed-rate PON without ONU grouping. For simplicity, the overhead is omitted. The colored blocks represent the bandwidth allocated to different services, which is determined by a fixed data rate r0and frame duration t0.Ui jand Mi j indicate the jth unicast and multicast service targeting ONU i, respectively. Suppose that there are six ONUs and five traffic flows (from four unicast services and one multicast service) that require transmission. Since both unicast and multicast data are transmitted only once, no redundancy is generated. In this scenario, the downstream effective throughput equals the sum of the bandwidth allocated to each service. Figure 3(b) presents an example of a downstream PHY frame in a flexible-rate PON with a fixed ONU grouping method. In this case, the system achieves its maximum system capacity. Six ONUs are divided into three groups: Group 1, Group 2, and Group 3. ONU1 and ONU2 in Group 1 achieve the highest data rate r1due to favorable channel conditions, while ONU3 and ONU4in Group 2 operate at a lower peak data rate r2. ONU5and ONU6in Group 3 experience the poorest channel conditions and operate at the lowest peak data rate r3=r0. Assuming the FS frame of each group is allocated an equal (b) (a) (c) G1(ONU1&2) G2(ONU3&4) G3(ONU5&6) G1(ONU1&2) G2(ONU3) G3(ONU4&5&6) U3 Fig. 3. Examples of a downstream PHY frame in (a) a fixed-rate PON, (b) a flexible-rate PON, and (c) a flexible-rate PON with ONU grouping optimization. Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on October 06,2025 at 09:44:35 UTC from IEEE Xplore. Restrictions apply. 762 Vol. 17, No. 9 / September 2025 / Journal of Optical Communications and Networking Research Article time slot duration (i.e., t1=t2=t3), the flexible-rate PON can transmit the same data in less time (indicated by the red dashed blocks), thereby increasing the maximum system throughput. However, the multicast data M1directed to ONUs in three different groups (i.e., ONU1, ONU4, and ONU6) must be transmitted separately for each group, occupying up to three time slots (M1 1,M4 1, and M6 1), which introduces unnecessary redundancy and significantly reduces the effective throughput. Figure 3(c) shows an example after optimizing the grouping by moving ONU4from Group 2 to Group 3 (i.e., lowering its data rate to r3), so the redundant transmission M4 1can be eliminated. Compared to the solution in Fig. 3(b), this adjustment further improves the effective throughput. However, moving ONU1to Group 3 to reduce further redundancy is not advisable, as lowering its data rate would increase the time slot required for the unicast service U1 1, thus decreasing the overall effective throughput. Therefore, optimizing ONU grouping involves a trade-off between minimizing redundancy and maintaining high ONU data rates. Furthermore, considering that not only the multicast traffic load but also multicast memberships can dynamically change as ONUs join or leave a multicast session, it is necessary to design a traffic-aware algorithm to optimize ONU grouping dynamically. In addition, the examples in Figs. 3(b) and 3(c) adopt a simple bandwidth allocation strategy, where bandwidth is evenly allocated among ONU groups. In practice, downstream bandwidth allocation in flexible-rate PONs is more complex and can significantly impact the system effective throughput. To address this, the next section proposes a scheme that integrates optimized ONU grouping with an advanced downstream bandwidth allocation algorithm. 3. TRAFFIC-AWARE ONU GROUPING AND DOWNSTREAM BANDWIDTH ALLOCATION Based on the proposed downstream scheduling protocol, we further design a TAOG-GBDTA scheme combined with two heuristic algorithms for flexible-rate PONs. Figure 4shows the TAOG-GBDTA framework in the TC layer of OLT. In this framework, the proposed TAOG algorithm and GBDTA algorithm optimize the ONU grouping and downstream bandwidth allocation, respectively. As shown in Fig. 4, both the TAOG and GBDTA algorithms receive traffic parameters from the multicast information manager module in the upper layer during each scheduling cycle, which has a fixed length of 125 µs. These parameters include the traffic load of services and the mapping table between services and their multicast memberships. Based on these parameters, the TAOG algorithm periodically generates optimized ONU groupings, which are fed into the GBDTA algorithm and provided to the traffic scheduler for generating PLOAMd messages. The interval between these updates is referred to as an optimization cycle, which is set as an integer multiple of the scheduling cycle. Note that the optimization cycle is configurable and can be set significantly longer than the scheduling cycle to match hardware constraints and ensure enough time for ONUs to switch groups via PLOAMd-based updates. In Fig. 4, blue arrows indicate the inputs and outputs that occur at each scheduling cycle, while green arrows TAOG: Traffic-aware ONU grouping GBDTA: Grouping-based downstream time slot allocation ONU grouping and bandwidth allocation Processed per scheduling cycle Processed per optimization cycle Optimized ONU grouping Time slot allocation Multicast information manager Traffic scheduler Traffic parameters TAOGGBDTA Optimized ONU grouping Fig. 4. TAOG-GBDTA framework. represent those at each optimization cycle. The GBDTA algorithm is responsible for performing downstream bandwidth allocation. When a new ONU grouping is provided, GBDTA updates the relevant parameters accordingly to ensure optimal bandwidth allocation decisions. The output decisions are fed into the traffic scheduler to generate FS frames for each ONU group. The detailed descriptions of these two algorithms are provided in the following subsections. A. Traffic-Aware ONU Grouping The TAOG algorithm is designed to iteratively optimize ONU grouping in response to time-varying network conditions, thereby improving effective throughput. Specifically, the algorithm begins with an initial ONU grouping, i.e., the currently adopted ONU grouping, and explores new ONU groupings by making incremental adjustments, i.e., reassigning one ONU to a different group or modifying its data rate. At each step, the algorithm evaluates whether the new solution can improve system effective throughput by an acceptance threshold. As the algorithm progresses, the threshold gradually decreases, making it increasingly selective and ultimately converging on an optimized grouping solution. The TAOG algorithm is detailed as follows. At the beginning of each scheduling cycle, traffic-related parameters are updated. Let 3,, and 0denote the sets of all services, ONUs, and groups, respectively. Both unicast and multicast services are included in 3, as a unicast service can be considered a special multicast case with a single ONU member. For each service i∈3, the newly arrived traffic in the previous scheduling cycle is denoted by di, and the cumulative traffic load from the end of the last optimization cycle to the present moment is denoted by pi. These parameters are collectively represented by vectors D= [d1,d2, ... , dNS] and P= [p1,p2, ... , pNS], respectively, where NSrepresents the total number of services. Pis updated as (P+D)during each scheduling cycle. Besides, the multicast membership of Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on October 06,2025 at 09:44:35 UTC from IEEE Xplore. Restrictions apply. Research Article Vol. 17, No. 9 / September 2025 / Journal of Optical Communications and Networking 763 each service is updated using the multicast management messages received by the OLT, such as the JOIN/LEAVE requests from the higher-layer Internet Group Management Protocol (IGMP) [33]. Upon receiving such requests through the data plane, the OLT updates its multicast membership table and reconfigures the XGEM port-ID of affected ONUs accordingly via OMCC. The mapping between services and their ONUs can be described as a matrix M, defined as M=     m1,1m1,2... m1,NO m2,1m2,2... m2,NO . . .. . ..... . . mNS,1mNS,2... mNS,NO      ,(1) where NOdenotes the total number of ONUs, and mi,j∈ {0,1}indicates whether ONU jis a member of service i. Specifically, mi,j=1 indicates that ONU jis a member of service i, and mi,j=0 otherwise. Similarly, the current ONU grouping is described by a matrix GC, expressed as GC=     g1,1g1,2... g1,NG g2,1g2,2... g2,NG . . .. . ..... . . gNO,1gNO,2... gNO,NG      ,(2) where NGdenotes the number of groups, and gj,k∈ {0,1} indicates whether ONU jbelongs to group k(i.e., gj,k=1 means ONU jbelongs to group k, and gj,k=0 otherwise). As mentioned in Section 2, each ONU is assumed to belong to only one group at a time. Accordingly, each ONU is restricted to a single group during the optimization, and gi,kmust satisfy X k∈0 gj,k=1,j∈.(3) Then, the OLT checks an internal countdown timer τ, which tracks the remaining scheduling cycles before the next grouping optimization. If τ=1, the optimization process is triggered, and τis reset to τG, which denotes the predefined optimization cycle. Otherwise, τis decremented by one, and the algorithm terminates for this scheduling cycle. During grouping optimization, the performance of the current grouping GCis evaluated. For this, a matrix L= [li,k]is generated, expressed as L=MGC,(4) where the element li,kindicates the number of ONU members of service iin group k. In the enhanced downstream scheduling protocol, the ONU members of a service belonging to the same group share the same time slot, and redundancy occurs when ONU members are from different groups. To account for shared time slots, each li,kin Lis transformed into l0 i,k; thus, a new matrix L0is generated. l0 i,kis calculated by l0 i,k=1,if li,k≥1 0,otherwise ,(5) where l0 i,kindicates whether service ihas any ONU members in group k. In other words, each non-zero l0 i,kcorresponds to one real traffic flow that must be scheduled, including the redundant ones. Consequently, the total number of flows equals the sum of all l0 i,kin L0. Metrix L0is then used to calculate the total traffic load and redundancy for each service, which is denoted by a vector B= [b1,b2, ... , bNG]and expressed as B=PL0.(6) Let the vector R= [r1,r2, ... , rNG]denotes the data rates of the ONU groups. Using Band R, the total time slots for group kin the last optimization cycle are calculated as (bk/rk). Furthermore, the effective throughput under the current ONU grouping GCin the last optimization cycle, referred to as ETc, is expressed as ETC=Pi∈3pi Pk∈0 bk rk .(7) Note that the numerator, which represents the total amount of accumulated traffic to be served, is fixed during each optimization cycle, while the denominator reflects the total time required to serve this traffic under the current ONU grouping GC. Therefore, the optimization on effective throughput can be formulated by minimizing the denominator, referred to as the effective service time (ESC). To minimize the effective service time ESC, an optimized grouping solution GBshould be found. To achieve this, the algorithm follows an iterative process to explore potential solutions, and the number of iterations is determined by Imax. ESC and GCare initially set to the minimum effective service time ESBand the optimal ONU grouping solution GB, respectively. At each iteration, a new ONU grouping solution GNis generated based on GCby randomly selecting an ONU and adjusting its data rate up or down by one level. Note that the adjustment should not exceed the peak data rate constraint of the ONU. The new effective service time ESNunder GNis then calculated by Eqs. (4)–(7). The difference in effective service time, referred to as δ, is expressed as δ=ESN−ESC.(8) If δ < 0, GNis accepted as the new current solution. Otherwise, the acceptance of GNis determined probabilistically using the acceptance probability PA, which can be written as PA=e−δ AT ,(9) where AT denotes the acceptance threshold. A higher initial AT allows the algorithm to explore a wider solution space by increasing the probability of accepting suboptimal solutions, whereas a lower AT results in a more conservative search. To determine whether GNis accepted, PAis then compared against a random value Yranging from 0 to 1. If PA>Y, GNis accepted; otherwise, it retains GC. Following this, AT is reduced according to (AT =γ·AT), where γis the decay factor controlling the rate of reduction. A higher γensures a slower decay and more thorough exploration, while a lower γ Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on October 06,2025 at 09:44:35 UTC from IEEE Xplore. Restrictions apply. 764 Vol. 17, No. 9 / September 2025 / Journal of Optical Communications and Networking Research Article Algorithm 1. Traffic-Aware ONU Grouping (TAOG) Algorithm Input:τG,τ,AT,Imax ,γ,3,,0,D,P,M,GC,R Output:GB 1. P←P+D 2. if τ > 1then 3. τ←τ−1 4. elseif τ== 1then 5. τ←τG 6. Calculate L0and Bby Eqs. (4)–(6) 7. Calculate ESCby Eq. (7) 8. GB←GC; ESB←ESC 9. while i≤Imax do 10. Generate a new solution GN 11. Calculate ESNbased on GN 12. δ←ESN−ESC 13. if δ < 0then 14. GC←GN;ESC←ESN 15. else 16. PA←exp(−δ/AT) 17. Generate Y 18. if PA>Ythen 19. GC←GN;ESC←ESN 20. end if 21. end if 22. if ESC<ESBthen 23. GB←GC;ESB←ESC 24. end if 25. AT ←AT ·γ 26. i←i+1 27. end while 28. Reset P 29. end if accelerates convergence but risks getting trapped in a local optimum. At the end of each iteration, the best-known solution GBis updated if ESCis smaller than the previously recorded ESB; otherwise, GBremains unchanged. After completing the iterations, the best solution GBis selected based on the minimum observed effective service time. The elements in Pare reset to zero for the next optimization cycle. Finally, GBis passed to the GBDTA algorithm and the traffic scheduler for downstream bandwidth allocation and traffic scheduling. The pseudo-code for the proposed TAOG algorithm is shown in Algorithm 1. Steps (1–5) check whether grouping optimization is required in the current scheduling cycle. In steps (6–8), the effective service time for the current ONU grouping in the last optimization cycle is calculated and temporarily set as the optimal solution. Following this, steps (9–26) try to find an optimized ONU grouping solution by iteration. Steps (10–11) generate a new grouping solution, and steps (12–21) determine whether to accept the new solution. Steps (22–24) compare the new solution with the current optimal solution and update the optimal solution accordingly. Afterward, the algorithm reduces its acceptance threshold (step 25) and iterates again. Finally, Pis reset in step (28). B. Grouping-Based Downstream Time Slot Allocation The GBDTA algorithm is proposed to optimize the downstream bandwidth allocation in multicast scenarios, thereby further maximizing the effective throughput of the flexible-rate PON. Here, bandwidth specifically refers to the time slots allocated for data transmission. First, the OLT calculates the length of the available time slots for the entire downstream PHY frame (T) before the beginning of each scheduling cycle, which can be expressed as T=E−TPSBd −TFS −TP,(10) where Edenotes the total duration of a downstream PHY frame, TPSBd denotes the time slots occupied by the PSBd, TFS accounts for the time slots used for FS headers, and TP represents the time slots allocated for parity bits in the FEC codewords. Simultaneously, if the TAOG algorithm has generated a new ONU grouping solution, GCis updated. The matrix M, which reflects the multicast memberships of services, is also updated based on the latest network conditions. Whenever either GCor Mis updated, L0is recalculated using Eqs. (4) and (5). Subsequently, the OLT allocates downstream time slots to each service. To prevent the overuse of downstream bandwidth by any single service due to the variation in bandwidth demand, each service must be allocated a maximum available time slot. Here, the maximum available time slot of service iis determined by its bandwidth weight ωi, which is defined as the ratio of the required time slot for service ito the total required time slots for all services, expressed as ωi=TSReq i Pn∈3TSReq n ,(11) where TSReq idenotes the requested time slot of service i. For a multicast service, more than one time slot may be required in different FS frames, where different modulation formats are applied during transmission. Thus, TSReq iis expressed as TSReq i=X k∈0 l0 i,k di rk .(12) Based on ωi, the maximum available time slot for service ican be obtained by (ωiT). If the required TSReq ican be accommodated within this limit, it is allocated as requested; otherwise, it is limited to (ωiT). The total time slot allocated to service i, i.e., TSSer i, is expressed as TSSer i=TSReq i,if TSReq i≤ωiT ωiT,otherwise. (13) Following this, TSSer ifor each service is further divided and incorporated into multiple FS frames of its associated ONU groups. The time slot allocated to group kfor service i, denoted as TSGro i,k, can be calculated as TSGro i,k=σi,k·TSSer i,(14) where σi,krepresents the bandwidth weight assigned to group kfor service i. For a multicast service, the data encapsulated in Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on October 06,2025 at 09:44:35 UTC from IEEE Xplore. Restrictions apply. Research Article Vol. 17, No. 9 / September 2025 / Journal of Optical Communications and Networking 765 Algorithm 2. Grouping-Based Downstream Time Slot Allocation (GBDTA) Algorithm Input:E,TPSBd,TFS,TP,3,,0,M,GC,D,R Output: TSGr o i,k 1. T←E−TPSBd −TFS −TP 2. Update L0by Eqs. (4) and (5) 3. fori∈3do 4. Calculate TSReq iand ωiby Eqs. (11) and (12) 5. if TSReq i≤ωiT 6. TSSer i←TSReq i 7. else 8. TSSer i←ωiT 9. end if 10. for k∈0do 11. Calculate σi,kand TSGro i,kby Eqs. (14) and (15) 12. end for 13. end for 14. Generate downstream FS frames based on TSGro i,k each FS frame are identical in size, while the time slot length varies depending on the data rate of the group. Therefore, σi,k is expressed as σi,k=l0 i,k rkPn∈0 l0 i,n rn .(15) Since each ONU in the group can independently identify its own data frames by recognizing the XGEM header, TSGro i,kcan be shared among the ONU members in the group, and there is no need for further division at the ONU level. Finally, the traffic scheduler generates the FS frame for each group based on TSGro i,k. The pseudo-code of the GBDTA algorithm is detailed in Algorithm 2. Step (1) updates the parameters. Step (2) calculates the maximum available time slot for each service. In steps (3–10), the algorithm allocates time slots to services based on their demands. Subsequently, the time slots allocated to each group are calculated in step (11). Finally, in step (14), FS frames are generated based on the allocated time slots for each group. The computational complexity of TAOG-GBDTA is analyzed as follows. The complexity of the TAOG algorithm depends on the number of iterations. For the GBDTA algorithm, its complexity is determined by the number of services and their corresponding groups, expressed as O(|3| · |0|). Given that conventional bandwidth allocation algorithms in flexible-rate PONs also assign time slots to each group associated with a service, the time complexity remains unchanged. Therefore, the primary source of complexity increase in the proposed scheme mainly stems from the optimization of ONU grouping. 4. SIMULATION AND PERFORMANCE EVALUATION In this section, we evaluate the performances of our proposed TAOG-GBDTA scheme compared to two benchmarks. In Benchmark 1, each ONU operates at its maximum available peak data rate, while in Benchmark 2, all ONUs are configured to operate at the lowest available data rate (i.e., 25 Gb/s) throughout the simulation. For bandwidth allocation, both benchmarks use a classic multicast scheduling mechanism, i.e., the multicast sharing-based weighted fair queuing (MSWFQ) mechanism [32]. For this, we develop a simulator in MATLAB, which incorporates our enhanced downstream scheduling protocol to simulate the downstream scheduling procedures in a flexible-rate PON. A. Simulation Conditions Our simulation considers a flexible-rate PON system with a varying number of ONUs: 16, 32, and 64. The ONUs are initially divided into four groups based on their available peak data rates, which correspond to 25, 50, 75, and 100 Gb/s, respectively. This rate combination can be achieved by the fine-granularity rate control demonstrated in [16]. The corresponding peak data rates are assigned based on typical received optical power distributions and rate thresholds reported in existing studies [16,18]. Specifically, the proportion of ONUs operating at 25, 50, 75, and 100 Gb/s is set to approximately 1%, 31%, 58%, and 8%, respectively. At least one ONU is included in each group. The physical distance of each ONU follows a Gaussian distribution, where the mean is 10 km and the variance is 2 km. The downstream traffic consists of 8 unicast and 8 multicast services, each modeled as a typical Internet Protocol television stream. The packet arrival times follow a Poisson distribution, with all services sharing the same average packet arrival rate. Each service has a minimum traffic rate of 10 Mbit/s, where the XGEM frame size is 1324 bytes, including an 8-byte XGEM header [5,34]. The buffer size of each service queue is set to 20 MB. Furthermore, ONU join and leave events (triggered by multicast JOIN and LEAVE requests) for multicast memberships are considered as a Poisson process. Accordingly, the OMCI messages sent by the OLT to assign or withdraw multicast XGEM Port-IDs also follow a Poisson distribution [31–33]. Meanwhile, the number of multicast requests for each multicast service follows a Zipf distribution. The Zipf distribution is commonly used to model multimedia access patterns and reflects the dynamic changes in multicast memberships [31,32]. The request probability for multicast service i, denoted by RPi, is expressed as RPi=C iα,i∈3, (16) where αdetermines the long-tail behavior of the distribution. In our simulation, αis set to 1 to closely reflect typical multicast access patterns [31,32]. The normalization constant Cis defined as C=1 Pi∈3 1 iα .(17) Additionally, another lightweight OMCI service is also considered, which includes periodic Management Information Base (MIB) Upload and performance monitoring (PM) reports. Specifically, the OLT broadcasts a 56-byte MIB Upload request (including an 8-byte header) every 30 s [35]. Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on October 06,2025 at 09:44:35 UTC from IEEE Xplore. Restrictions apply.