Method for distributing data to a client in iterative data processing

Abstract

A method for distributing data to clients in iterative data processing is provided. The method includes establishing a multicast session with a plurality of clients, where the multicast session is provided by one or more data replication components that replicate multicast data; for each of a plurality of iterations of iterative data processing, determining a subset of the plurality of clients for the iteration, determining the data to be distributed to the determined subset of clients, notifying one or more data replication components of the determined subset, providing the determined data to one or more data replication components, controlling the one or more data replication components to replicate the data according to the determined subset and transmit the determined data to each client in the determined subset within the multicast session. Selected drawing relating to the abstract: FIG. 15

Description

Technical Field

[0001] The present disclosure relates to a method of distributing data to clients in iterative data processing.

Background Art

[0002] In iterative data processing where data processing is distributed among multiple clients, such as federated learning (FL) for example, data needs to be distributed to clients in each iteration. To avoid duplicate transmissions via various interfaces within the communication system used for data distribution (which may include an internal core network interface, as well as a core network - radio access network interface and an air interface), multicast may be used for data transmission. However, depending on the data processing method, it may be desirable to select, for each iteration, a subset of clients rather than sending data to all clients registered for data processing in that iteration, where each subset is the destination for which respective data is to be sent. This may be done, for example, for privacy reasons, but in federated learning for example, it results in better training outcomes. Therefore, for each iteration, it may be necessary to establish a new multicast session for each respective subset. This incurs a high signaling overhead, so a more efficient approach for distributing data to clients in iterative data processing is desirable.

Summary of the Invention

[0003] A method for distributing data to clients in iterative data processing is provided. The method includes establishing a multicast session with a plurality of clients, where the multicast session is provided by one or more data replication components that replicate multicast data; for each of a plurality of iterations of the iterative data processing, determining a subset of the plurality of clients for the iteration, determining the data to be distributed to the determined subset of clients, notifying one or more data replication components of the determined subset, providing the determined data to one or more data replication components, controlling one or more data replication components to replicate the data according to the determined subset, and transmitting the determined data to each client in the determined subset within the multicast session.

Brief Description of the Drawings

[0004] In the drawings, like reference numerals generally represent the same parts throughout the various figures. The drawings are not necessarily to scale, but rather emphasis is placed on explaining the principles of the present invention. In the following description, various aspects are described with reference to the following drawings.

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

DETAILED DESCRIPTION OF THE INVENTION

[0005] The following detailed description refers to the accompanying drawings that illustrate specific details and aspects of the present disclosure in which the invention may be practiced. Other aspects may be used and structural, logical, and electrical changes may be made without departing from the scope of the present invention. Since some aspects of the present disclosure can combine with one or more other aspects of the present disclosure to form new aspects, the various aspects of the present disclosure are not necessarily mutually exclusive.

[0006] The following describes various examples corresponding to aspects of the present disclosure:

[0007] Example 1 is a method for distributing data to clients in iterative data processing, which includes the step of establishing a multicast session with a number of clients, where the multicast session is provided by one or more data replication components that replicate multicast data, and for each of a plurality of iterations of the iterative data processing, · determining a subset of the number of clients for the iteration; · determining the data to be distributed to the determined subset of clients; · notifying the determined subset to one or more data replication components; · providing the determined data to one or more data replication components, including the steps of replicating the data according to the determined subset and controlling the one or more data replication components to transmit the determined data to each client of the determined subset within the multicast session.

[0008] Example 2 is the method of Example 1, further including, in each iteration, the step of processing the determined data by each client of the determined subset.

[0009] Example 3 is the method of Example 1 or 2, and includes the steps where the server determines a subset and data, and provides the determined data to one or more data replication components.

[0010] Example 4 is the method of Example 3, and further includes the step where in each iteration, the server receives the processing results of the determined data from each client of the determined subset.

[0011] Example 5 is the method of Example 4, and further includes the step where the server merges (or combines) the processing results.

[0012] Example 6 is the method of any one of Examples 1 to 5, the iterative data processing is federated learning, and in each iteration, the determined data is the specification or details of the current version of the machine learning model for the iteration.

[0013] Example 7 is the method of any one of Examples 1 to 6, and one or more data replication components include one or more core network components of the mobile communication system and / or one or more radio access networks.

[0014] Example 8 is the method of any one of Examples 1 to 7, and one or more data replication components include one or more base stations and / or one or more user plane functions.

[0015] Example 9 is the method of any one of Examples 1 to 7, and one or more data replication components include one or more multicast / broadcast user plane functions.

[0016] Example 10 is any one of the methods of Examples 1 to 9, and includes steps of associating a determined subset with a time slot among a series of time slots, notifying one or more data replication components of the association, and transmitting the determined data to one or more data replication components by notifying the determined subset to the one or more data replication components at the time slot.

[0017] Example 11 is the method of Example 10, and includes a step of notifying one or more data replication components of the relevance of the determined subset with the time slot in the control plane.

[0018] Example 12 is any one of the methods of Examples 1 to 9, and includes steps of associating a determined subset with a subset identifier, notifying one or more data replication components of the association, and transmitting the determined data to one or more data replication components by notifying the determined subset to the one or more data replication components by transmitting the subset identifier to the one or more data replication components.

[0019] Example 13 is the method of Example 12, and includes a step of notifying one or more data replication components of the relevance of the determined subset with the subset identifier in the control plane.

[0020] Example 14 is the method of Example 12 or 13, and includes a step of transmitting the determined data together with the subset identifier to one or more data replication components by notifying the determined subset to the one or more data replication components.

[0021] Example 15 is the method of Example 14, and includes a step of transmitting the determined data together with the subset identifier to one or more data replication components in the user plane by notifying the determined subset to the one or more data replication components.

[0022] Example 16 is one of the methods of Examples 1 to 9, and includes the step of notifying one or more data replication components of the determined subset by sending a list of clients of the determined subset to one or more data replication components.

[0023] Example 17 is the method of Example 16, and includes the step of notifying one or more data replication components of the determined subset by sending a list of clients of the determined subset together with the determined data to one or more data replication components.

[0024] Example 18 is the method of Example 17, and includes the step of notifying one or more data replication components of the determined subset by sending a list of clients of the determined subset together with the determined data to one or more data replication components in the user plane.

[0025] Example 19 is one of the methods of Examples 1 to 18, and includes the step of registering a large number of clients for iterative data processing.

[0026] Example 20 is a communication system configuration configured to execute one of the methods of Examples 1 to 19.

[0027] It should be noted that any one or more features of the above specific examples may be combined with any one of the other specific examples. In particular, the examples described in the context of the device are equally applicable to the method.

[0028] According to yet another embodiment, there is provided a computer-readable medium and a computer program including instructions, which, when executed by a computer, cause the computer to execute any one of the methods of the above examples.

[0029] Hereinafter, various embodiments will be described in more detail.

[0030] Figure 1 shows federated learning.

[0031] Federated learning (FL) is a decentralized machine learning (ML) technique in which, under the control of a central server (denoted as the FL server) 104 (implemented, for example, by a 5G or 6G cloud), each client uses a local training dataset 103 to train an ML model 101 (e.g., a deep neural network (DNN)) across multiple (FL) clients 102.

[0032] This typically involves multiple training iterations (or training cycles), and in each training iteration (or training "cycle") ● The FL server 104 selects several clients 102 to participate in this iteration of the FL training;

[0033] ● The FL server 104 sends the current version of the model 101 to a subset of the clients 102 (starting with the initial version of the model in the first iteration). Note that the subset may change from iteration to iteration. This is referred to as dynamic member selection.

[0034] ● Each FL client 102 (also called an FL node) in each subset locally trains the model 101 (i.e., updates the current version of the model, e.g., updates the weights of the neural network). Each FL client in each subset 101 uses its own training data 103 (also called local training data).

[0035] ● The FL server 104 pools the training results from the clients 102 of each subset and generates a new version of the (global) ML model 101 by aggregating the results.

[0036] The version obtained as a result of the ML model 101 (i.e., the new version of the last iteration) can then be used by a model consumer such as the NWDAF (AnLF) (Network Data Analytics Function (Analytical Logic Function)), and the model consumer can use it to make predictions from the data obtained from the surroundings (such as communication resource requests).

[0037] Dynamic member selection can be used for the following purposes: ● Mitigate client heterogeneity: Clients in federated learning may have various characteristics such as varying computing resources, network conditions, and data distributions. Dynamic client selection helps mitigate client heterogeneity by selecting the clients most suitable for the current iteration based on those characteristics, thereby improving the efficiency and effectiveness of the training process.

[0038] ● Improve privacy: Federated learning involves training models on data distributed across multiple clients. Dynamic client selection may help improve privacy by selecting clients with less sensitive or more randomized data, reducing the risk of exposing confidential information.

[0039] ● Reduce communication overhead: In federated learning, the communication overhead between clients and the central server can become a bottleneck. Dynamic client selection may help reduce communication overhead and thereby the amount of data that needs to be transmitted by selecting a subset of clients that are most relevant to the current iteration.

[0040] ● Improve the convergence speed: Also, dynamic client selection may help improve the convergence speed of the federated learning process by selecting clients with the highest-quality data or those most likely to improve the global model based on their local updates.

[0041] Regarding the implementation in the framework of a 5G communication network, the FL server 104 may be an application function, and the client 102 may be an application in a user equipment (UE).

[0042] Federated learning is a decentralized (distributed) application, and as a result, there is significant communication between the FL server 104 and the FL clients 102. Therefore, it is desirable for the communication to be efficient, especially when the number of clients is large. In each iteration (e.g., the cycle of FL training), the same data (i.e., the current version of the (global version) of the ML model 101) is distributed from the server 104 to multiple clients 101, that is, there may be redundant data transmission via each communication network used for FL data transmission. In the following, FL using dynamic member selection is assumed, that is, the FL server 104 selects a subset of clients 102 for each iteration (i.e., "dynamic selection" of clients is performed).

[0043] When transmitting the same data to multiple devices, the communication network typically has a multicast function.

[0044] FIG. 2 shows a wireless communication system 200 configured according to 5G (fifth generation) as defined by 3GPP (Third Generation Partnership Project) in this example, in relation to multicast and various relevant interfaces.

[0045] The wireless communication system 200 includes mobile wireless terminal devices 202 such as UEs (user equipment) and nano equipment (NE). The mobile wireless terminal device 202, also called a subscriber terminal, forms the terminal side, while the other components of the wireless communication system 200 described below are parts of the mobile wireless communication network side, that is, parts of a mobile wireless communication network (e.g., a Public Land Mobile Network (PLMN)).

[0046] Furthermore, the wireless communication system 200 includes a radio access network 203, which can include a plurality of radio access network nodes, that is, base stations configured to provide radio access according to fifth generation (5G) radio access technology (5G New Radio). Also, the wireless communication system 200 may be configured according to LTE (Long Term Evolution) or other mobile wireless communication standards, but here 5G is used as an example. Each radio access network node can provide wireless communication with the mobile wireless terminal device 202 via an air interface. It should be noted that the radio access network 203 can include any number of radio access network nodes.

[0047] The wireless communication system 200 further includes a core network (CN, here 5GC) including an Access and Mobility Management Function (AMF) 201 connected to the RAN 203 and a Unified Data Management (UDM) 204. Here and in the following examples, the UDM may further be composed of, for example, an actual UE subscription data base known as a Unified Data Repository (UDR). The core network has one or more Policy Control Functions (PCF) 205.

[0048] The core network further includes a Session Management Function (SMF) 206 and a plurality of User Plane Functions (UPF) 207. The SMF 206 is for processing Protocol Data Unit (PDU) sessions, i.e., for creating, updating, removing PDU sessions and managing session contexts with User Plane Functions (UPF).

[0049] The core network further includes an Application Function (AF) 208 (or Application Server (AS)). The AF 208 is connected to other core network components such as the PCF 205 via a Network Exposure Function (NEF), especially when the AF 208 is maintained by a third party (i.e., someone other than the operator of the mobile wireless communication system 200).

[0050] To support 3GPP Multicast / Broadcast Service (MBS), there are Multicast Broadcast (MB) versions of SMF 206 and UPF 207 respectively: MB-SMF 210 and MB-UPF 211. Furthermore, the communication system includes a Multicast / Broadcast Service Function (MBSF) 212 and a Multicast / Broadcast Service Transport Function (MBSTF) 213.

[0051] Figure 3 shows MBS in 3GPP.

[0052] For example, MBS data (i.e., "traffic") from FL 104 is received by the core network 301 of the communication system. The MBS data should be transmitted to a plurality of UEs 302, 303 via the radio access network (RAN) 304 of the communication system. The core network 301 may use separate PDU sessions 305 for a plurality of UEs 302, or may also use shared transport 306 to each base station of the RAN 304 for a plurality of UEs 302 (the plurality of UEs 302 are served by a base station, and the base station transmits data using wireless point-to-point or wireless point-to-multipoint transmission). These two options are referred to as transmission via 5GC individual MBS traffic delivery and transmission via 5GC shared MBS traffic delivery, respectively. In either case (since it can be assumed that not all clients 302 are served by the same base station), there is a data replication unit 307 at one or more locations within the communication system, i.e., the communication system includes one or more data replication components, which replicate the data, resulting in a version of the data for each UE (or at least for each base station (or radio cell)); within the radio cell, the data can be transmitted to a plurality of UEs only once.

[0053] In the context of multicast functions such as 3GPP's MBS, typically, the following concepts exist:

[0054] ● Multicast subscription group: A group of UEs, where a UE (e.g., each user) subscribes to a certain multicast service (e.g., MBS in multicast mode), and thus is authorized to join the multicast group and receive data related to that multicast group via the multicast service.

[0055] ● Multicast Group: Members of a multicast joining group can join a multicast group (associated with the multicast joining group). It has an active multicast service (of MBS in multicast mode), and thus is a group of users who are receiving data transmitted by the multicast service. Therefore, when a UE joins a multicast service, it can be assigned to a multicast group, and the multicast group is a group of UEs that will receive the same multicast data stream. A multicast group is identified by a multicast address, and multicast data is sent by the sender to that address.

[0056] ● Multicast Session: It is the continuous and time-limited reception of multicast service data by a UE. Each UE within a multicast group receives data within the corresponding multicast session.

[0057] In the following example, federated learning is used as a use case, but it should be noted that the embodiments may also be applicable to other applications that repeatedly transfer the same data to a dynamic group of UEs, i.e., other types of iterative data processing (such as distributed graph processing or MapReduce applications).

[0058] FIG. 4 shows a flow diagram 400 showing traffic in a more general application example.

[0059] Server 401 and a plurality of clients 402 (numbered 1 to N) are included in the flow.

[0060] In each iteration 403, the same data (e.g., in the case of FL, the current version of the global model (GM-i)) is sent from the server 401 to the client (subset) (where i is the iteration index). Thus, in each iteration, the data can be multicast.

[0061] In each iteration 403, a subset of the clients 402 is required to receive the data. Thus, according to various embodiments, when multicasting, in each iteration 403, target client (e.g., target UE) filtering is performed (e.g., filtering of UEs in a multicast group including all clients 102).

[0062] The overall data transmission includes periodic data transmission but is not a multimedia stream, which is the case when the size of the data in all iterations is the same, i.e., size(GM-i)=size(GM-j). Therefore, the data volume (per iteration) can be shared with the 5GC to derive service quality (QoS) parameters such as the Maximum Data Burst Volume.

[0063] Hereinafter, embodiments in the context of federated learning will be described.

[0064] FIG. 5 shows a flowchart 500 illustrating the FL procedure, and particularly relates to data transmission using multicast via a filtering mechanism provided according to various embodiments.

[0065] The FL server 501 (e.g., corresponding to the FL server 104) and the FL client 505 (e.g., the UE numbered from 1 to N, e.g., corresponding to the FL client 102) are included in the flow. In this example, the communication between the FL server 501 and the FL client 505 is performed via the 5G communication system, particularly, the MF-SMF 502, the MB-UPF 503, and the data replication component 504 (which may be the UPF, the MB-UPF, or the RAN) (see FIGS. 2 and 3 showing two options for data replication).

[0066] As described above, according to various embodiments, data is filtered in the multicast session, i.e., in each training iteration, the data is transmitted to the client subset selected (or defined) for that training iteration. Thus, at 506, the FL server 501 establishes a "filtered" multicast session (i.e., a multicast session having this filtering function) by the core network of the communication system.

[0067] The filtered multicast indicator (i.e., an indicator indicating that the multicast session is a filtered multicast session) may be included, for example, in the Nmbsmf_MBSSession_CreateRequest message sent from the NEF or the MBSF to the MB-SMF (e.g., in accordance with clause 7.1.1.2, Figure 7.1.1.2-1 of 3GPP TS 23.247 version 17.3.0 release 17). The data volume of the periodic traffic (i.e., the size of the data of the global model) may be defined as part of the MBS service information in the Nmbsmf_MBSSession_CreateRequest message.

[0068] At 507, the FL client 505 registers with the FL server 501 (to participate in the training), and the FL server 501 initializes the FL client 505 for training, which includes notifying the identification of the multicast session (in this example, the MBS session ID). At 508, the FL client 505 participates in the multicast session.

[0069] Then, the training loop is executed, which, in each iteration 509 ● Selecting members by the FL server 501 at 510 (i.e., selecting a subset of FL clients) ● Transmitting the global ML model to the selected subset of FL clients via the core network (initiated by the FL server 501 at 511) ○ This includes performing data replication at 512 by the MB - UPF, UPF or RAN, and transmitting the corresponding data (e.g., the weights of the neural network) to the selected subset of clients 505 at 513. ● Performing local training at each FL client 505 in the subset at 514 ● Transmitting the model updates of the subset of FL clients 505 at 515 ● Aggregating the model updates into a new (updated) version of the (global) ML model at 516.

[0070] The FL server 501 (which corresponds to AF 208) can share, in each iteration, the list of UEs of each FL client subset (i.e., the list of target UEs), before or during the transmission of multicast data, with replication entities (i.e., replication components), such as MB-UPF, UP, RAN (specifically, a base station, e.g., gNB) (e.g., during or before the steps of multicast data transmission in TS 23.247, clause 7.2.1.3, Figure 7.2.1.3-1; see steps 619, 621, and 624 in Figure 6).

[0071] Figure 6 shows a flow diagram 600 in which the FL server instructs a replication component (found by control plane signaling) of an FL client subset with a target UE list in the user plane.

[0072] UE 601 (representing an FL client), RAN 602, AMF 603, SMF 604, UPF 605, NRF 606, PCF 607, MB-SMF 608, and MB-UPF 609, AF 610 (corresponding to the FL server) are included in the flow.

[0073] In 611 to 618, UE 601 participates in a multicast session.

[0074] In 619, AF 610 provides multicast data (e.g., the details (specification) of the current version of the ML model to be trained). Using the multicast data, AF 610 provides a target UE list (i.e., the list of UEs corresponding to the FL clients selected for the current iteration).

[0075] As described with reference to Figure 3, there are two options: 1) Transmission by 5GC-shared MBS traffic delivery 2) Transmission by 5GC individual MBS traffic distribution

[0076] In the first option, at 620, the core network control plane determines the target base station (i.e., the base station (serving base station) that responds to the target UE). At 621, the MB-UPF 609 transmits the multicast data together with the target UE list to the determined base station of the RAN 602, and that base station transmits the data to each target UE 601 (e.g., using the MBS radio bearer) at 622.

[0077] In the second option, at 623, the core network control plane determines the target UPF (i.e., the UPF that responds to the target UE). At 624, the MB-UPF 609 transmits the multicast data together with the target UE list to the determined UPF, and the UPF transmits the data to each target UE 601 via the RAN 602 using the PDU session for each UE 601 at 625.

[0078] FIG. 7 shows a flow diagram 700 in which the FL server instructs a subset of FL clients (via control plane signaling) before transmitting the multicast data to the replication components. The multicast data is implicitly associated with the list (associated with the list) via the time slot in which it is transmitted, or explicitly (i.e., together with the multicast data) via an ID tag (associated with the list) included in the user plane transmission.

[0079] As shown in FIG. 7, the UE 701 (representing the FL client), RAN 702, AMF 703, SMF 704, UPF 705, NRF 706, PCF 707, MB-SMF 708, MB-UPF 709, AF 710 (corresponding to the FL server) are included in the flow.

[0080] At 711 to 718, the UE 701 participates in the multicast session.

[0081] At 719, AF 710 transmits to MB-SMF 708 a target UE list for the current iteration and an indication of a time slot or ID tag associated with the target UE list.

[0082] Similar to that in FIG. 6, there are differences between the two options. 1) Transmission by 5GC shared MBS traffic delivery 2) Transmission by 5GC individual MBS traffic delivery

[0083] In the first option, at 720, the core network control plane determines a target base station (i.e., the base station that responds to the target UE (serving base station)). Then, at 721, MB-SMF 701 notifies the target base station to MB-UPF, and at 722, transmits the target UE list and the associated time slot or ID tag to the target base station, respectively.

[0084] In the second option, at 723, the core network control plane determines a target UPF (i.e., the UPF that responds to the target UE). Then, at 724, MB-SMF 701 notifies the target UPF to MB-UPF, and at 725, transmits the target UE list and the associated time slot or ID tag to the target UPF, respectively.

[0085] At 726, AF 710 provides multicast data (e.g., details of the current version of the ML model to be trained). AF 710 does this either at a specific time slot associated with the target UE list or provides the ID tag of the target UE list along with the multicast data (the latter case is shown in FIG. 7).

[0086] In the first option, at 727, the MB-UPF 709 transfers the multicast data to the target base station (indicating the ID tag of the target UE list in some cases), and the target base station, at 728, transmits the data to each target UE 701 (e.g., using the MBS radio bearer).

[0087] In the second option, at 729, the MB-UPF 709 transfers the multicast data to the target UPF (indicating the ID tag of the target UE list in some cases), and the target UPF, at 730, transmits the data to each target UE 701 via the RAN 702 using the PDU session for each UE 701.

[0088] Below, examples for the transmission of the target UE list are described in relation to FIGS. 8 and 9 for the two possible options described above (as in the example of FIG. 6).

[0089] FIG. 8 shows the transmission of the target UE list in the extended header (e.g., GTP-U or IPv6) from the AF 801 via the MB-UPF 802 to the target base station 803 that caters to the target UEs 806 (in this example, "UE-1" and "UE-2") for the case of shared MBS traffic delivery (Option 1 above).

[0090] The MB-UPF 802 can determine the target base station 803 by consulting (or querying) the MB-SMF 804, and the MB-SMF 805 can consult the AMF 128.

[0091] FIG. 9 shows the transmission of the target UE list in the extended header (e.g., GTP-U or IPv6) from the AF 901 via the MB-UPF 902 to the target UPF 903 that caters to the target UEs 904 (in this example, "UE-1", "UE-2", and "UE-3") for the case of individual MBS traffic delivery (Option 2 above).

[0092] Below, an example of the use of time slots for specifying a target UE, also referred to as the "CP assisted" case (see the description of FIG. 7), is described in relation to FIGS. 10 and 11 for the two possible options described above.

[0093] FIG. 10 shows the transmission of multicast data for the case of shared MBS traffic distribution in the case of CP assistance.

[0094] AF 1002 first sends information to MB-SMF 1004 (via the control plane) that a certain collection of UEs (in this example, "UE-1" and "UE-2") is associated with a certain time slot. Thus, the time slot implicitly associates the multicast data (sent in the time slot) with the target UE list.

[0095] After this is done, AF 1001 can send multicast data to MB-UPF 1002 within each time slot. MB-UPF 1002 receives a notification about the target base station 1003 by MB-SMF 1004 that determines the target UE list (with respect to the time slot) and consults AMF 1005 about the target base station 1003. MB-UPF 1002 can then send the multicast data to the target base station 1003 that serves the target UE 1006 during a given time slot. MB-SMF 1004 can also notify the RAN of the association of the target UE 1006 with the time slot. It should be noted that the complete control plane signaling (indicated by hatching) can be completed before the user plane signaling (i.e., the transmission of multicast data).

[0096] FIG. 11 shows the transmission of multicast data for the case of individual MBS traffic distribution in the case of CP assistance.

[0097] AF 1101 first sends, via the MB-SMF 1102 (over the control plane), to each target UPF 1103 the information that a certain set of UEs 1105 (in this example, "UE-1", "UE-2", "UE-3") that the UPF is to handle is associated with a certain time slot. Thus, the time slot implicitly associates the multicast data (sent in that time slot) with the target UEs handled by that UPF 1103.

[0098] After this is done, AF 1101 sends the multicast data within each time slot to the MB-UPF 1104, and the MB-UPF 1104 forwards it to each target UPF 1103. The target UPF 1103 then sends the multicast data to the target UEs 1105 (using individual PDF sessions) during the time slot. Here too, it should be noted that the complete control plane signaling (shown hatched) can be completed before the user plane signaling (i.e., the transmission of the multicast data).

[0099] Below, an example of the use of ID tags for specifying target UEs, also called the "CP-UP combined" case (see the description of FIG. 7), is described in relation to FIGS. 12 and 13 for the two possible options mentioned above.

[0100] FIG. 12 shows the transmission of multicast data for the case of shared MBS traffic delivery in the case of CP-UP combination.

[0101] AF 1202 first sends, via the control plane, to the MB-SMF 1204 the information that a certain collection of UEs (in this example, "UE-1" and "UE-2") is associated with a certain ID tag.

[0102] After this is done, AF 1001 can send multicast data containing the tag ID to MB-UPF 1202. The tag ID is included, for example, in the extended header (GTP-U, IPv4, IPv6) of the multicast data, and explicitly associates the multicast data with the target UE list. MB-UPF 1202 determines the target UE list (from the tag ID) and receives a notification about target base station 1203 from MB-SMF 1204, which consults AMF 1205 about target base station 1003. MB-UPF 1202 can then send the multicast data to target base station 1203 that serves target UE 1206. MB-SMF 1204 can also notify the RAN of the association with the tag ID of target UE 1206. It should be noted that the complete control plane signaling (indicated by hatching) can be completed before the user plane signaling (i.e., the transmission of multicast data).

[0103] Figure 13 shows the transmission of multicast data for the case of individual MBS traffic delivery in the case of CP-UP binding.

[0104] AF 1302 first sends, via MB-SMF 1302 (through the control plane), information to UPF 1303 that the UEs 1305 (in this example, "UE-1", "UE-2", "UE-3") served by the UPF are associated with a certain ID tag.

[0105] After this is done, AF 1301 sends the multicast data containing the tag ID to MB-UPF 1304, and MB-UPF 1304 forwards it to each target UPF 1303. The tag ID is included, for example, in the extension header (GTP-U, IPv4, IPv6) of the multicast data, and explicitly associates the multicast data with the target UE list. The target UPF 1303 then sends the multicast data to the target UE 1305 (using individual PDF sessions). Here too, it should be noted that the complete control plane signaling (indicated by hatching) can be completed before the user plane signaling (i.e., the transmission of multicast data).

[0106] For communication between MB-UPF and MB-SMF, N4 (specifically, the N4mb interface) is extended by PFCP message types to query the location of the UE (to determine the serving base station of that UE).

[0107] Regarding communication between AF and MB-SMF (especially regarding specifying the currently selected subset), it is possible to introduce MB-SMF service extensions. For example, the Nmbsmf_MBSSession_ContextUpdate message can be extended to enable suspension or resumption of multicast data reception.

[0108] Furthermore, N4 (specifically, the N4mb interface) is extended with an extension to update the Forwarding Action Rule Information Element (FAR IE) within the PFCP session modification request to include only the current target UE list.

[0109] Figure 14 shows the multicast of data from AF 1401 to UE 1404 via MB-UPF 1402 and base station 1403.

[0110] Therefore, multicast data travels via the N6 (specifically, N6mb) interface and N3 (specifically, N3mb interface), as well as a dedicated radio bearer.

[0111] The approach described herein enables the use of a multicast session to distribute the global model from the FL server to the FL clients, thereby reducing the number of transmissions of multiple replicas of the global model from the AF to the core network (e.g., 5GC). This includes the transmission and resource consumption efficiency (e.g., in terms of the number of packets) on the N3 interface and the N6 interface when shared MBS traffic distribution is deployed.

[0112] The use of a target UE list for multicast data eliminates the unnecessary duplication of data for clients not selected for this iteration of FL training, thus improving transmission and resource consumption efficiency and eliminating the creation and participation procedures of the MBS session for each FL cycle (thus increasing signaling efficiency).

[0113] The achievable transmission and resource efficiency is particularly high when a small number of (MB-)UPFs and base stations respond to the selected clients and the UE does not hand over during the iteration.

[0114] In short, according to various embodiments, a method as shown in FIG. 15 is provided.

[0115] FIG. 15 shows a flowchart 1500 showing a method of distributing data to clients in iterative data processing.

[0116] At 1501, a multicast session with a number of clients (also referred to as a "filtered" multicast session) is established, where the multicast session is provided by one or more data replication components for replicating multicast data.

[0117] At 1502, for each of a plurality of iterations of iterative data processing · At 1503, a subset of the number of clients (a subset that may be different from the subset in other iterations) is determined for that iteration. · At 1504, the data to be distributed to the determined subset of clients is determined. · At 1505, one or more data replication components receive a notification for the determined subset. · At 1506, the determined data is provided to one or more data replication components, and the one or more data replication components are controlled to replicate the data according to the determined subset and send the determined data to each client of the determined subset within the multicast session.

[0118] It should be noted that the steps in FIG. 15 do not necessarily have to be executed in the order shown (for example, one or more data replication components may receive a notification for the determined subset before the data is determined, or the data may be determined before the subset is determined).

[0119] That is, various embodiments provide an approach (at the architecture level) for efficiently distributing (exchanging) data to clients. This is achieved by controlling the subset of clients to which data is sent within a multicast session, while using the same multicast session (using the same one over multiple iterations, i.e., avoiding establishing a multicast session for each determined subset of clients). Thus, data is sent for multiple operations within the same multicast context. In other words, each iteration has an associated (determined) data set to be sent in that iteration, and the data sets for multiple operations are sent within the same multicast session (i.e., the same multicast (session) context).

[0120] A client can, for example, correspond to a mobile terminal in a mobile communication system (e.g., each client may correspond to a respective mobile terminal).

[0121] Sending data within the same multicast context (i.e., maintaining the multicast context over multiple iterations) can, for example, include the following: ● Sending each piece of data using the same session ID, the same type of devices involved, the same protocol used, and / or any other relevant details specific to the multicast session for all of the multiple iterations.

[0122] ● The server that sends data for all of the multiple iterations includes the same multicast group address and port number in the header of the data packet that transmits the data.

[0123] ● The client participates in a multicast session identified by a unique ID notified by the server only once. The list of clients for each session is maintained by the network, and for each iteration, the replication entity (i.e., the replicated component) is instructed to replicate packets for the clients within a determined subset of the group. Packets sent to the network associated with the unique ID by the server are replicated within the network and received by all the clients in the determined subset participating in the session.

[0124] It should be noted that for at least some iterations, the subset changes for each iteration. The size of the data may be the same for all iterations.

[0125] As described above, one or more data replication components may receive notifications for their respective subsets via an id tag, a time slot, a control plane, or a user plane, i.e., the data replication entity is instructed to apply target client (e.g., UE) list filtering as follows: ● Transmit the list simultaneously with the multicast data via an extended header of the user plane protocol. ● Configure the data replication entity to apply target UE list filtering to a given time slot before transmitting the multicast data via an extension of the control plane protocol and / or service. ● Transmit the target UE list with an ID tag to the data replication entity via an extension of the control plane protocol / service, and embed the ID tag in the extended header of the user plane protocol.

[0126] According to various embodiments, a communication system configuration (e.g., a server or server system) configured to execute the method of FIG. 15 is provided.

[0127] The communication system configuration includes components such as one or more communication interfaces (e.g., for communicating with core network components, radio access network components, and / or mobile terminals), one or more processors, and memory.

[0128] The components of the communication system configuration may be implemented, for example, by one or more circuits. A "circuit" may be understood as any kind of logical implementation entity that may be processor-executed software or a dedicated circuit stored in memory, firmware, or any combination thereof. Thus, a "circuit" may be a hard-wired logic circuit or a programmable logic circuit, which may be a programmable processor, such as a microprocessor (e.g., a microprocessor). Also, a "circuit" may be a processor that executes software, e.g., any kind of computer program. Any other kind of implementation of each of the functions described above may also be understood as a "circuit".

[0129] Although specific embodiments have been described, those skilled in the art should understand that various changes in form and detail may be made therein without departing from the spirit and scope of the embodiments of the present disclosure as defined by the appended claims. Accordingly, the scope of this application is indicated by the appended claims and is intended to include all modifications within the meaning and scope of equivalence of the claims.

Claims

1. 1. A method for distributing data to clients in an iterative data processing, comprising: establishing a multicast session with a plurality of clients, the multicast session being provided by one or more data replication components that replicate multicast data; and For each of a plurality of iterations of the iterative data processing, determining a subset of the plurality of clients for the iteration; determining data to be distributed to the determined subset of clients; notifying the one or more data replication components of the determined subset; providing the determined data to the one or more data replication components, where the one or more data replication components are controlled to replicate data according to the determined subset and transmit the determined data to each client of the determined subset in the multicast session; The method includes:

2. 2. The method of claim 1, further comprising the step of processing the determined data by each client of the determined subset in each iteration.

3. 2. The method of claim 1, comprising the step of a server determining the subset and data and providing the determined data to the one or more data replication components.

4. 4. The method of claim 3, further comprising the step of: in each iteration, the server receiving results of processing the determined data from each client of the determined subset.

5. The method of claim 4 further comprising the step of the server merging the processing results.

6. 2. The method of claim 1, wherein the iterative data processing is federated learning, and in each iteration, the determined data is details of a current version of a machine learning model for that iteration.

7. The method of claim 1, wherein the one or more data replication components include one or more core network components and / or one or more radio access networks of a mobile communication system.

8. 2. The method of claim 1, wherein the one or more data replication components include one or more base stations and / or one or more user plane functions.

9. 10. The method of claim 1, wherein the one or more data replication components include one or more multicast broadcast user plane functions.

10. 2. The method of claim 1, comprising the steps of associating the determined subset with a time slot of a series of time slots, notifying the one or more data replication components of the association, and notifying the one or more data replication components of the determined subset by transmitting the determined data to the one or more data replication components in the time slot.

11. 11. The method of claim 10, further comprising the step of notifying, in a control plane, the one or more data replication components of the association of the determined subset with the time slot.

12. 2. The method of claim 1, comprising associating the determined subset with a subset identifier, notifying the one or more data replication components of the association, and notifying the one or more data replication components of the determined subset by transmitting the subset identifier to the one or more data replication components.

13. 13. The method of claim 12, comprising informing, in a control plane, the one or more data replication components of an association of the determined subset with the subset identifier.

14. 13. The method of claim 12, comprising notifying the one or more data replication components of the determined subset by transmitting the subset identifier together with the determined data to the one or more data replication components.

15. A communication system configured to carry out a method according to any one of claims 1 to 14.

Images