A communication scheduling method and system based on PON physical layer in-order aggregation

Abstract

The application discloses a communication scheduling method and system based on physical layer ordered aggregation of PON, and belongs to the technical field of optical access network communication. The application realizes full-isomorphic mapping of distributed reasoning task logical topology to physical network by constructing an optical distribution network physical network diagram and establishing an algorithm power collaborative cluster. Each network terminal device embeds a real-time state vector in an extended XGEM frame header of an uplink data frame, and an optical line terminal constructs a dynamic panoramic ledger based on the real-time state vector, predicts a logical fragment completion time, generates an authorized window sequence satisfying phase locking and ordered constraints, and broadcasts the authorized window sequence to each device. The device uploads an execution result in a corresponding window, and an aggregation point records and generates an audit report according to an actual arrival order. The application does not need to cache rearrangement, realizes physical layer ordered aggregation, and improves the scheduling efficiency and real-time performance of a distributed reasoning task.

Description

Technical Field

[0001] This invention relates to the field of optical access network communication technology, and specifically to a communication scheduling method and system based on PON physical layer sequential aggregation. Background Technology

[0002] Distributed inference or training in traditional centralized data centers or edge cloud environments typically rely on high-speed interconnects, stable low latency, and predictable resource scheduling. However, this centralized architecture faces high construction and operating costs, high power consumption, and struggles to meet the requirements of large-scale edge computing for data localization and ultra-low latency.

[0003] When extending distributed intelligent computing capabilities to the optical access networks of operators with a massive number of terminals, existing technologies face the following challenges:

[0004] I. The optical access network adopts a time-division multiple access (TDMA) architecture, with uplink bandwidth centrally authorized by the optical line terminal (OLT). Traditional dynamic bandwidth allocation mechanisms authorize bandwidth based on traffic parameters such as queue length, service priority, and service level agreement (SLA) bandwidth, aiming to optimize traffic throughput or ensure quality of service. When the allocation of the authorization window does not match the completion time of the computation task, it can lead to idle computing power or waiting delays. In the existing scheme, the computation results of each network terminal device's NTD are uploaded in a "first-come, first-served" manner, requiring the aggregation point to rearrange out-of-order results.

[0005] Second, the access network contains a massive number of heterogeneous terminals, including home optical modems, fiber-to-the-room master-slave routers, smart set-top boxes, edge boxes with neural network processing units, and industrial gateways, which vary significantly in computing power, memory, and power consumption. The operating status of these terminals is easily affected by user behavior and main business operations, exhibiting nonlinearity and volatility. In the existing scheduling mechanism, the optical line terminal (OLT) has difficulty obtaining the instantaneous computing power status of each network terminal device's NTD in real time.

[0006] Third, traditional telecommunications services such as voice, IPTV, and leased lines have strict service level agreement requirements regarding latency, jitter, and packet loss. When requisitioning idle computing power from network terminal equipment (NTD), existing solutions lack an isolation mechanism between the service assurance plane and the computing power plane.

[0007] Fourth, existing distributed computing solutions mainly coordinate at the application layer through transmission control protocols or Internet protocols, resulting in the separation of logical topology and physical topology. This leads to uncertainty in the actual transmission path of data packets, making it difficult to utilize the deterministic characteristics of the underlying physical network of the optical access network.

[0008] Fifth, existing technologies attempt to obtain the computing power of optical network units through optical line terminals (OLTs) and adjust the transmission bandwidth accordingly. However, such solutions mainly focus on capacity matching and do not involve the timing and sequence control of uplink transmission.

[0009] VI. Distributed scheduling algorithms typically run internally within a system, and their scheduling behavior, resource allocation, and data flow processes lack externally observable audit evidence. Scheduling records can be tampered with and are difficult to detect.

[0010] In summary, under the existing optical access network architecture and distributed intelligent computing paradigm, the optical line terminal (OLT) cannot dynamically adjust the timing and order of the uplink authorization window according to the computing progress of each network terminal device (NTD), nor can it uniformly perceive and schedule the real-time computing power status of each NTD. The aggregation point needs to rely on the cache to rearrange the out-of-order results. Summary of the Invention

[0011] To address the aforementioned shortcomings in existing technologies, this invention provides a communication scheduling method and system for physical layer sequential aggregation based on PON. This solves the problems in existing technologies where the optical line terminal (OLT) cannot dynamically adjust the uplink authorization window timing and order according to the network terminal device's NTD calculation progress, and where the aggregation point needs to cache and rearrange out-of-order results.

[0012] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a communication scheduling method based on PON physical layer sequential aggregation, comprising the following steps:

[0013] S1. Construct the physical network diagram of the optical distribution network (ODN), divide the physical node set in the physical network diagram into computing power collaborative clusters (PON-Cluster), and establish a physical-logical isomorphic mapping from the logical topology diagram of the distributed inference task to the physical network diagram to obtain the target network terminal device (NTD).

[0014] S2. Generate logical fragments for the distributed inference task to be processed, determine the expected aggregation order of the logical fragments, and send each logical fragment to the corresponding target network terminal device (NTD).

[0015] S3. During execution, each target network terminal device (NTD) periodically embeds an in-band real-time state vector in the extended XGEM frame header of its uplink data frame and sends it to the optical line terminal (OLT) to construct a dynamic panoramic ledger.

[0016] S4. On the optical line terminal (OLT), based on the dynamic panoramic ledger, predict the completion time of each logical segment and generate an authorization window sequence that satisfies phase-locked loop and in-order constraints.

[0017] S5. Broadcast the authorization window sequence to all target network terminal devices (NTDs). Each target network terminal device (NTD) uploads its logical fragment execution results within the settings of the corresponding authorization window and records the actual logical fragment order that arrives at the aggregation point.

[0018] S6. Based on the logical fragmentation order of actual arrival at the aggregation point, the optical line terminal (OLT) generates and stores audit records in each scheduling cycle.

[0019] Furthermore, S1 includes the following sub-steps:

[0020] S11. Construct the physical network diagram of the optical distribution network (ODN), and divide the physical node set in the physical network diagram into computing power collaborative clusters (PON-Cluster);

[0021] S12. Establish a physical-logical isomorphic mapping from the set of logical computing units in the logical topology graph of the distributed reasoning task to the set of physical nodes in the physical network graph.

[0022] S13. Set multiple constraints on the physical-logical isomorphic mapping to obtain the target network terminal device NTD;

[0023] The physical network diagram of the optical distribution network (ODN) in S1 includes: a set of physical nodes and a set of physical links. The set of physical nodes includes: optical line terminal (OLT), optical splitter (Splitter), and network terminal equipment (NTD). The set of physical links represents optical fiber connections.

[0024] The process of dividing the computing power collaboration cluster PON-Cluster in S1 includes: taking each optical splitter in the physical node set of the physical network diagram as the root node, extracting the subtree of the root node, and classifying all physically connected network terminal devices (NTDs) under the subtree of the root node into a computing power collaboration cluster PON-Cluster.

[0025] The logical topology graph in S1 includes a set of logical computing units and a set of logical edges. The set of logical computing units includes data units of task partitioning, operators, and large model LLM. The set of logical edges represents the dependencies or data flow between logical computing units.

[0026] Furthermore, the multiple constraints in S13 include: computing power collaborative cluster PON-Cluster constraint, routing isomorphism constraint, and overflow isomorphism constraint;

[0027] The constraints of the PON-Cluster (Power-Oriented Networking Cluster) include: Each logical computing unit allocated to the PON-Cluster constitutes a subset of logical computing units, and the mapping result satisfies: ,in, To be allocated to the first Each logical computing unit in a PON-Cluster constitutes a subset of logical computing units. It is a physical-logical isomorphic mapping symbol. For the first NTD, a network terminal device For the first The physical subtree of the root node is a single optical splitter. and It is a positive integer;

[0028] Routing isomorphic constraints include: ,in, For a set of logical edges, A set of physical links;

[0029] Overflow isomorphism constraints include: , ,in, For overflow data, For a set of data types that are allowed to overflow, The amount of data that overflows. This represents the maximum capacity of the overflow data.

[0030] Furthermore, S3 includes the following sub-steps:

[0031] S31. During the execution of logical fragmentation, each target network terminal device (NTD) periodically embeds an in-band real-time state vector in the extended XGEM frame header of its uplink data frame.

[0032] S32. Receive the in-band real-time state vector sent by the target network terminal device NTD through the optical line terminal (OLT);

[0033] S33. The out-of-band computing power ledger and the in-band real-time state vector are merged through the optical line terminal (OLT) to obtain the dynamic panoramic ledger of the target network terminal device (NTD).

[0034] The in-band real-time state vector in S31 includes: reporting timestamp, NPU busy / idle status, LLM inference progress percentage, available memory, temperature, and preemption warning;

[0035] The process of acquiring the out-of-band computing power ledger in S33 includes: acquiring the out-of-band computing power ledger for each network terminal device (NTD) belonging to the physical node set through the network management and control channel. The out-of-band computing power ledger includes: collection timestamp, computing power of the network terminal device (NTD), memory capacity of the network terminal device (NTD), hardware health status of the network terminal device (NTD), and link quality of the network terminal device (NTD).

[0036] Furthermore, S4 includes the following sub-steps:

[0037] S41. Based on the dynamic panoramic ledger of the target network terminal device NTD, predict the time when the logical fragment is completed in the calculation of the target network terminal device NTD.

[0038] S42. Based on the completion time of the calculation of the target network terminal device NTD, generate an authorization window sequence that satisfies the phase-locked loop constraint and the sequential constraint for multiple target network terminal devices NTD of the same computing power collaborative cluster PON-Cluster.

[0039] Furthermore, the phase-locked constraint in S42 is:

[0040] ,

[0041] in, For the first The start time of the authorization window of the target network terminal device (NTD). For the first The completion time of the calculation of the target network terminal device NTD. This is the preset phase-locked loop tolerance; | represents the absolute value operation. It is a positive integer;

[0042] The sequential constraint in S42 is:

[0043] ,

[0044] in, For the first The start time of the authorization window of the target network terminal device (NTD). For the first The start time of the authorization window of the target network terminal device (NTD). For the first The duration of the authorization window for each target network terminal device (NTD). For the first The protection interval of the authorization window of a target network terminal device (NTD). For the first The logical shard index corresponding to each logical shard that needs to be aggregated. For the first The logical shard index corresponding to each logical shard that needs to be aggregated. It is a positive integer;

[0045] The authorization windows in the authorization window sequence of S42 are: ,in, For the first PON-Cluster of computing power collaboration The authorization window of the target network terminal device NTD. For the first The duration of the authorization window for each target network terminal device (NTD). For the first The protection interval between the authorization window of a target network terminal device (NTD) and the next authorization window.

[0046] Furthermore, S5 includes the following sub-steps:

[0047] S51. Write each authorized window in the authorized window sequence into the PON bandwidth allocation table and send it to the target network terminal device NTD.

[0048] S52. Determine whether the service assurance plane resources of the target network terminal device NTD need to occupy computing plane resources. If so, set the preemption warning flag to valid and save the micro-checkpoint of the current logical segment. The micro-checkpoint includes KV cache, token cursor and inter-layer activation digest. Send the preemption warning flag to the optical line terminal OLT through the uplink XGEM frame header and jump to step S53. Otherwise, keep the preemption warning flag in an invalid state, the logical segment continues to execute normally, and jump to step S54.

[0049] S53. Through the optical line terminal (OLT), select the target network terminal device (NTD) with the lowest NPU utilization in the same computing power collaborative cluster (PON-Cluster) as the backup target. Move the logical fragments and micro-checkpoints in the target network terminal device (NTD) corresponding to the preemption warning flag to the backup target, update the mapping relationship between the logical fragments and the target network terminal device (NTD), and recalculate the calculation completion time of the logical fragments.

[0050] S54. In the target network terminal device NTD, the actual calculation completion time is sent according to the authorization window setting, and the logical fragmentation order of actual arrival at the aggregation point is recorded.

[0051] Furthermore, S6 includes the following sub-steps:

[0052] S61. Perform a hash operation on the dynamic panoramic ledgers of multiple target network terminal devices (NTDs) in the same computing power collaborative cluster (PON-Cluster) to obtain a ledger snapshot summary;

[0053] S62. Perform a hash operation on the authorization windows of multiple target network terminal devices (NTDs) in the same computing power collaborative cluster (PON-Cluster) to obtain the authorization sequence fingerprint;

[0054] S63. Calculate the order consistency index based on the actual logical fragmentation order and the expected aggregation order of the aggregate points.

[0055] S64. Combine the ledger snapshot summary, authorization sequence fingerprint, and sequence consistency indicator to generate a scheduling audit record, and append the scheduling audit record to the chained audit record set;

[0056] S65. When the aggregation result of logical sharding needs to be transmitted from the current computing power collaborative cluster PON-Cluster to the parent domain, cross-domain overflow processing is performed, an overflow audit record is generated and appended to the chained audit record set.

[0057] Furthermore, in S65, after the logical fragmentation of the target network terminal device NTD is completed and aggregated in sequence on the backup target, if the aggregation result needs to be transmitted from the current computing power collaborative cluster PON-Cluster to the parent domain, the cross-domain overflow requirement is marked, whitelist conversion and size verification are performed on the aggregation result, an overflow audit record containing the current time, the current optical line terminal OLT identifier, the converted data and the parent domain target address is generated, it is appended to the chained audit record set, and the converted data is transmitted to the parent domain target address.

[0058] A PON-based physical layer sequential aggregation communication scheduling system includes:

[0059] The physical-logical isomorphic mapping module is used to construct the physical network graph of the optical distribution network (ODN), divide the physical node set in the physical network graph into computing power collaborative clusters (PON-Cluster), and establish a physical-logical isomorphic mapping from the logical topology graph of the distributed inference task to the physical network graph to obtain the target network terminal device (NTD).

[0060] The task distribution module is used to generate logical fragments for the distributed inference tasks to be processed, determine the expected aggregation order of the logical fragments, and distribute each logical fragment to the corresponding target network terminal device (NTD).

[0061] The panoramic ledger construction module is used to periodically embed an in-band real-time state vector into the extended XGEM frame header of each target network terminal device (NTD) during execution and send it to the optical line terminal (OLT) to construct a dynamic panoramic ledger.

[0062] The reverse scheduling module is used on the optical line terminal (OLT) to predict the completion time of each logical segment based on the dynamic panoramic ledger, and generate an authorization window sequence that satisfies phase-locked loop and in-order constraints.

[0063] The authorization execution module is used to broadcast the authorization window sequence to all target network terminal devices (NTDs). Each target network terminal device (NTD) uploads its logical fragment execution results within the settings of the corresponding authorization window and records the actual logical fragment order that arrives at the aggregation point.

[0064] The audit module is used to generate and store audit records in each scheduling cycle of the optical line terminal (OLT) based on the logical fragmentation order of actual arrival at the aggregation point.

[0065] The beneficial effects of this invention are as follows:

[0066] 1. This invention, through the authorization window sequence generated in S4 that satisfies phase-locked loop (PLL) constraints and in-order constraints, ensures that each network terminal device (NTD) uploads the calculation results sequentially according to the desired aggregation order, guaranteeing that the logical fragmentation results arrive at the aggregation point in the predetermined order at the physical layer. Compared with the existing "first-come, first-served, then rearranged at the aggregation layer" approach, this invention eliminates the need to maintain a large number of out-of-order rearrangement caches at the aggregation point, reducing tail waiting time and pipeline bubbles, and improving aggregation processing efficiency.

[0067] 2. This invention, through phase-locked loop (PLL) constraints and sequence constraints in S4, enables the licensed window sequence generated by the optical line terminal (OLT) to simultaneously control the microsecond-level timing of uplink transmission and the transmission order of each network terminal device's (NTD). Compared to existing technologies that only focus on bandwidth capacity matching, this invention achieves precise control over both uplink transmission timing and order.

[0068] 3. This invention achieves a match between the start time of the authorization window generated by the optical line terminal (OLT) and the actual computation completion time of the NTD through periodic in-band real-time state vector reporting in S3 and the prediction of the fragmentation completion time based on the dynamic panoramic ledger in S4 (phase-locked constraint). Compared with existing schemes where the authorization window is independent of the computation completion time, this invention reduces the idle time of computing power caused by "computation completion and waiting for authorization," thereby improving the computing power utilization rate of the NTD.

[0069] 4. This invention embeds an in-band real-time state vector into the extended frame header of the uplink data frame of the network terminal device (NTD) in step S3, enabling the optical line terminal (OLT) to obtain the NPU busy / idle status, inference progress, memory pressure, and preemption warnings of each NTD within microseconds. Compared with existing technologies where the optical line terminal cannot perceive the instantaneous computing power status of the terminal, this invention constructs a dynamic panoramic ledger, providing a real-time and accurate data foundation for scheduling decisions.

[0070] 5. This invention employs a dual-plane isolation mechanism, dividing resources within the network terminal device (NTD) into a service assurance plane and a computing power plane, with the service assurance plane given higher priority. Compared to existing solutions lacking isolation mechanisms, this invention ensures zero intrusion into core services such as voice, IPTV, and leased lines, while safeguarding the service level agreements (SLAs) of traditional telecommunications services.

[0071] 6. This invention, through the physical-logical isomorphic mapping established in S1, constrains the edge set of the logical topology graph of the distributed computing task to a subset of the edge set of the physical topology graph of the optical distribution network, and constructs a computing power collaborative cluster with the physical subtree of the optical splitter as the boundary. Compared with the existing schemes that separate logical and physical topologies, this invention achieves data localization constraints by default, reducing the latency uncertainty and bandwidth cost caused by cross-domain transmission.

[0072] 7. This invention generates and stores audit records containing ledger snapshot summaries, authorization sequence fingerprints, and sequence consistency indicators in each scheduling cycle during step S6, enabling externally observable chain-like audit evidence for scheduling behavior, resource allocation, and data flow processes. Compared to existing technologies where scheduling algorithms run internally and records can be tampered with, this invention provides an immutable scheduling behavior verification mechanism, meeting the audit and regulatory requirements of telecommunications operations.

[0073] 8. This invention employs a preemption recovery mechanism to save micro-checkpoints when the network terminal device (NTD) is preempted by the main service. The optical line terminal (OLT) then selects an idle device within the same computing power collaborative cluster (PON-Cluster) to resume computation. Compared to existing technologies where tasks need to be restarted after interruption, this invention achieves millisecond-level rapid recovery, improving the telecom-grade reliability of distributed intelligent computing tasks. Attached Figure Description

[0074] Figure 1 This is a flowchart of a communication scheduling method based on PON physical layer sequential aggregation. Detailed Implementation

[0075] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0076] Example 1, such as Figure 1 As shown, a communication scheduling method based on PON physical layer sequential aggregation includes the following steps:

[0077] S1. Construct the physical network diagram of the optical distribution network (ODN), divide the physical node set in the physical network diagram into computing power collaborative clusters (PON-Cluster), and establish a physical-logical isomorphic mapping from the logical topology diagram of the distributed inference task to the physical network diagram to obtain the target network terminal device (NTD).

[0078] S2. Generate logical fragments for the distributed inference task to be processed, determine the expected aggregation order of the logical fragments, and send each logical fragment to the corresponding target network terminal device (NTD).

[0079] S3. During execution, each target network terminal device (NTD) periodically embeds an in-band real-time state vector in the extended XGEM frame header of its uplink data frame and sends it to the optical line terminal (OLT) to construct a dynamic panoramic ledger.

[0080] S4. On the optical line terminal (OLT), based on the dynamic panoramic ledger, predict the completion time of each logical segment and generate an authorization window sequence that satisfies phase-locked loop and in-order constraints.

[0081] S5. Broadcast the authorization window sequence to all target network terminal devices (NTDs). Each target network terminal device (NTD) uploads its logical fragment execution results within the settings of the corresponding authorization window and records the actual logical fragment order that arrives at the aggregation point.

[0082] S6. Based on the logical fragmentation order of actual arrival at the aggregation point, the optical line terminal (OLT) generates and stores audit records in each scheduling cycle.

[0083] In this embodiment, S1 includes the following sub-steps:

[0084] S11. Construct the physical network diagram of the optical distribution network (ODN), and divide the physical node set in the physical network diagram into computing power collaborative clusters (PON-Cluster);

[0085] S12. Establish a physical-logical isomorphic mapping from the set of logical computing units in the logical topology graph of the distributed reasoning task to the set of physical nodes in the physical network graph.

[0086] S13. Set multiple constraints on the physical-logical isomorphic mapping to obtain the target network terminal device NTD;

[0087] The physical network diagram of the optical distribution network (ODN) in S1 includes: a set of physical nodes and a set of physical links. The set of physical nodes includes: optical line terminal (OLT), optical splitter (Splitter), and network terminal equipment (NTD). The set of physical links represents optical fiber connections.

[0088] The process of dividing the computing power collaboration cluster PON-Cluster in S1 includes: taking each optical splitter in the physical node set of the physical network diagram as the root node, extracting the subtree of the root node, and classifying all physically connected network terminal devices (NTDs) under the subtree of the root node into a computing power collaboration cluster PON-Cluster.

[0089] The logical topology graph in S1 includes a set of logical computation units and a set of logical edges. The set of logical computation units includes task partitioning, operators (basic computation units in deep learning models, representing a mathematical operation performed on data), and data unit tokens of the large model LLM (Large Language Model). The set of logical edges represents the dependencies or data flow between logical computation units.

[0090] In the distributed inference scenario of large model LLM, the data unit of large model LLM refers to the basic data block that can be independently divided, distributed and processed during the LLM inference process.

[0091] An optical line terminal (OLT) is a core central office device in a passive optical network (PON) system, typically deployed in the operator's central office equipment room.

[0092] An Optical Distribution Network (ODN) is a fiber optic transmission network located between the Optical Line Terminal (OLT) and the Network Terminal Equipment (NTD) in a Passive Optical Network (PON) system. It consists of passive components such as optical fibers, optical splitters, and fiber optic connectors.

[0093] The target network terminal device NTD refers to the NTDs that have been allocated logical fragments after being selected from the set of physical nodes through physical-logical isomorphic mapping.

[0094] In this embodiment, Physical network diagram of an optical distribution network (ODN). For a set of physical nodes, For the physical link set; for each optical splitter subtree The corresponding computing power collaborative cluster (PON-Cluster) is defined as: the set of all network terminal devices (NTDs) physically connected under this subtree, denoted as... , For the first The physical subtree of the root node is a single optical splitter. For the first A computing power collaborative cluster, PON-Cluster.

[0095] In this embodiment, This is the logical topology diagram for a distributed reasoning task. This refers to a set of logical computation units (such as task fragments, operators, and LLM token segments). It is a set of logical edges.

[0096] Physical-Logical Isomorphism Mapping : It satisfies the following constraints:

[0097] The constraints of the PON-Cluster (Power-Oriented Networking Cluster) include: Each logical computing unit allocated to the PON-Cluster constitutes a subset of logical computing units, and the mapping result satisfies: ,in, To be allocated to the first Each logical computing unit in a PON-Cluster constitutes a subset of logical computing units. It is a physical-logical isomorphic mapping symbol. For the first NTD, a network terminal device For the first The physical subtree of the root node is a single optical splitter. and It is a positive integer;

[0098] Routing isomorphic constraints include: ,in, For a set of logical edges, A set of physical links;

[0099] Overflow isomorphism constraints include: , ,in, For overflow data, For a set of data types that are allowed to overflow, The amount of data that overflows. This represents the maximum capacity of the overflow data.

[0100] Overflow data refers to data transmitted from the target network terminal device (NTD) within the PON-Cluster to the optical line terminal (OLT) or higher-level networks.

[0101] Physical-logical isomorphic mapping is an allocation rule that assigns logical fragments to physical NTDs, under the premise of satisfying the constraints of computing power collaborative cluster (PON-Cluster), routing isomorphic constraints, and overflow isomorphic constraints.

[0102] This invention, through the PON-Cluster constraint, strictly restricts logical fragmentation to be executed on network terminal devices (NTDs) within the same optical splitter physical subtree, ensuring that the data flow path is consistent with the physical fiber connection path. Intermediate results and final return data of distributed computing tasks flow within the optical splitter physical subtree by default, treating cross-domain transmission as a controlled exception, reducing data transmission bandwidth costs. Furthermore, a minimum privacy domain is constructed with the optical splitter physical subtree as the boundary, providing a physical-level foundation for data residency and privacy protection. Through routing isomorphism constraints, the edge set of the logical topology graph must be a subset of the edge set of the physical topology graph, ensuring that network terminal devices (NTDs) corresponding to dependent logical fragments are physically directly connected, reducing communication distance and intermediate node forwarding overhead. Through overflow isomorphism constraints, overflow is only allowed to the parent domain, and overflow data is subject to whitelist format restrictions, size upper limit control, and audit log generation, achieving controlled execution and complete auditing of cross-domain transmission.

[0103] In S2, a set of logical fragments is generated for the distributed inference task to be processed. And determine the expected aggregation order of logical fragments. ,in, For the first logical partition, For the second logical partition, For the first A logical partition, This is the logical shard index corresponding to the first logical shard that needs to be aggregated. This is the logical shard index corresponding to the second logical shard that needs to be aggregated. For the first The logical shard index corresponding to each logical shard that needs to be aggregated is used to allocate each logical shard in the logical shard set to the target network terminal device (NTD) according to the mapping relationship.

[0104] In this embodiment, S3 includes the following sub-steps:

[0105] S31. During the execution of logical fragmentation, each target network terminal device (NTD) periodically embeds an in-band real-time state vector in the extended XGEM (XG-PON Encapsulation Method) frame header of its uplink data frames.

[0106] S32. Receive the in-band real-time state vector sent by the target network terminal device NTD through the optical line terminal (OLT);

[0107] S33. The out-of-band computing power ledger and the in-band real-time state vector are merged through the optical line terminal (OLT) to obtain the dynamic panoramic ledger of the target network terminal device (NTD).

[0108] The in-band real-time state vector in S31 includes: reporting timestamp, NPU busy / idle status, LLM inference progress percentage, available memory, temperature, and preemption warning;

[0109] The process of acquiring the out-of-band computing power ledger in S33 includes: acquiring the out-of-band computing power ledger for each network terminal device (NTD) belonging to the physical node set through the network management and control channel. The out-of-band computing power ledger includes: collection timestamp, computing power of the network terminal device (NTD), memory capacity of the network terminal device (NTD), hardware health status of the network terminal device (NTD), and link quality of the network terminal device (NTD).

[0110] The network management and control channels include the OMCI channel (ONU Management Control Interface) and the PLOAM channel (Physical Layer Operations, Administration and Maintenance).

[0111] In this embodiment, S4 includes the following sub-steps:

[0112] S41. Based on the dynamic panoramic ledger of the target network terminal device NTD, predict the time when the logical fragment is completed in the calculation of the target network terminal device NTD.

[0113] S42. Based on the completion time of the calculation of the target network terminal device NTD, generate an authorization window sequence that satisfies the phase-locked loop constraint and the sequential constraint for multiple target network terminal devices NTD of the same computing power collaborative cluster PON-Cluster.

[0114] In this embodiment, in step S41, an existing RNN / Transformer machine learning prediction model is invoked, and the dynamic panoramic ledger of the target network terminal device NTD is input into the machine learning prediction model to obtain the prediction completion time.

[0115] This invention utilizes a dynamic panoramic ledger constructed in S33, integrating the NTD benchmark computing power provided by the out-of-band computing power ledger with microsecond-level NPU busy / idle status and inference progress information provided by the in-band real-time state vector. This enables S41 to accurately predict the computation completion time of each logical shard. Based on this prediction, S42 generates an authorization window sequence that satisfies phase-locked loop constraints, ensuring that the start time of the authorization window is closely aligned with the actual computation completion time of the target network terminal device's NTD. This reduces idle computing time caused by "computation completion and waiting for authorization," improving the computing power utilization rate of the target network terminal device's NTD. Simultaneously, S42 generates an authorization window sequence that satisfies in-order constraints based on the computation completion time of each target network terminal device's NTD. This ensures that each NTD uploads computation results sequentially according to the expected aggregation order of the logical shards. The shard results received by the aggregation point arrive at the physical layer in a predetermined order, eliminating the need to maintain a large amount of out-of-order reordering cache, reducing tail waiting time and pipeline bubbles, and improving aggregation processing efficiency.

[0116] In this embodiment, the phase-locked constraint in S42 is:

[0117] ,

[0118] in, For the first The start time of the authorization window of the target network terminal device (NTD). For the first The completion time of the calculation of the target network terminal device NTD. This is the preset phase-locked loop tolerance; | represents the absolute value operation. It is a positive integer;

[0119] The sequential constraint in S42 is:

[0120] ,

[0121] in, For the first The start time of the authorization window of the target network terminal device (NTD). For the first The start time of the authorization window of the target network terminal device (NTD). For the first The duration of the authorization window for each target network terminal device (NTD). For the first The protection interval of the authorization window of a target network terminal device (NTD). For the first The logical shard index corresponding to each logical shard that needs to be aggregated. For the first The logical shard index corresponding to each logical shard that needs to be aggregated. It is a positive integer;

[0122] The authorization windows in the authorization window sequence of S42 are: ,in, For the first PON-Cluster of computing power collaboration The authorization window of the target network terminal device NTD. For the first The duration of the authorization window for each target network terminal device (NTD). For the first The protection interval between the authorization window of a target network terminal device (NTD) and the next authorization window. It specifies when the target network terminal device (NTD) begins sending data. It specifies the duration for which the target network terminal device (NTD) can send data. It is the waiting time from when the current window closes until the next window begins.

[0123] Aggregation refers to the process in distributed inference tasks where multiple logical slices, after execution, need to be aggregated, spliced, or combined to form the final computation result.

[0124] This invention achieves microsecond-level phase-locked looping between the authorization window and the computation completion time of the NTD by using phase-locked constraints to ensure that the start time of the authorization window is closely aligned with the computation completion time of the NTD. By using sequential constraints, the authorization window of subsequent fragments is forced to open after the authorization window of the preceding fragment has completely ended, so that each target network terminal device NTD sends the computation results in the logical aggregation order, ensuring that the fragmentation results arrive at the aggregation point in a predetermined order at the physical layer. The authorization window sequence specifies the authorization start time, duration and guard interval of each target network terminal device NTD, so that the NTD is sent in an orderly manner within the authorization window, achieving timing determinism and order controllability of uplink transmission.

[0125] In this embodiment, the total resource set of the target network terminal device (NTD) is divided into service assurance plane resources and computing power plane resources. Service assurance plane resources have a higher priority than computing power plane resources. Service assurance plane resources are used to ensure the SLA of main services such as broadband / voice / IPTV / leased lines, while computing power plane resources are used for logical sharding to achieve resource isolation between services and computing power tasks.

[0126] In this embodiment, S5 includes the following sub-steps:

[0127] S51. Write each authorized window in the authorized window sequence into the PON bandwidth allocation table and send it to the target network terminal device NTD.

[0128] S52. Determine whether the service assurance plane resources of the target network terminal device NTD need to occupy computing plane resources. If so, set the preemption warning flag to valid and save the micro-checkpoint of the current logical segment. The micro-checkpoint includes KV cache, token cursor and inter-layer activation digest. Send the preemption warning flag to the optical line terminal OLT through the uplink XGEM frame header and jump to step S53. Otherwise, keep the preemption warning flag in an invalid state, the logical segment continues to execute normally, and jump to step S54.

[0129] S53. Through the optical line terminal (OLT), select the target network terminal device (NTD) with the lowest NPU (Neural Processing Unit) utilization rate in the same computing power collaborative cluster (PON-Cluster) as the backup target. Move the logical fragments and micro-checkpoints in the target network terminal device (NTD) corresponding to the preemption warning flag to the backup target, update the mapping relationship between the logical fragments and the target network terminal device (NTD), and recalculate the calculation completion time of the logical fragments.

[0130] S54. In the target network terminal device NTD, the actual calculation completion time is sent according to the authorization window setting, and the logical fragmentation order of actual arrival at the aggregation point is recorded.

[0131] In S54, the aggregation point aggregates the results of each piece sequentially to obtain the aggregation result.

[0132] The aggregation point is a network node in this invention that is responsible for receiving and sequentially aggregating the execution results of each NTD fragment. It is usually deployed on the optical line terminal (OLT) or its controlled bypass edge node.

[0133] This invention, through S51, writes the authorized window sequence into the PON bandwidth allocation table and distributes it to each NTD, enabling each target network terminal device NTD to obtain a clear uplink transmission opportunity; through S52, when service assurance plane resources need to occupy computing power plane resources, it sets the preemption warning flag to be valid and saves the micro-checkpoint containing KV cache, token cursor, and inter-layer activation digest, and reports the preemption warning flag in real time through the uplink XGEM frame header; when there is no preemption, the preemption warning flag is kept in an invalid state and normal execution continues; through S53, the OLT selects the target network terminal device NTD with the lowest NPU utilization rate as a backup target within the same computing power collaborative cluster, migrates the logical fragments and micro-checkpoints to the backup target and updates the mapping relationship, and recalculates the calculation completion time; through S54, the target network terminal device NTD sends the calculation results within the authorized window and records the actual arrival order, realizing resource isolation between the service assurance plane and the computing power plane, low-overhead state preservation and fast migration during preemption, automatic selection and task takeover of idle devices within the cluster, and orderly uploading and sequential recording of calculation results.

[0134] Micro-checkpoints are a lightweight state preservation mechanism defined in this invention. They are used to save the intermediate state of distributed computing tasks with minimal overhead when the target network terminal device (NTD) is preempted by the main business, so that execution can be quickly resumed on other target network terminal devices (NTDs) in the future.

[0135] The KV cache is a cache of the key and value matrices of processed tokens in the LLM attention mechanism, used to avoid repeatedly calculating the attention results of historical tokens when generating each new token. The token cursor is a pointer to the position of the currently processed token, recording the index of the next token to be generated. The inter-layer activation summary is a compressed representation of the activation values ​​of the output of a certain layer of the neural network, used as input to the next layer in pipeline parallelism. These three are stored together in micro-checkpoints, enabling the relay target network terminal device NTD to read the attention data of historical tokens from the KV cache, determine the next processing position from the token cursor, and obtain the output state of the previous layer from the inter-layer activation summary, thereby achieving precise resumption of execution from the breakpoint without having to start reasoning from scratch.

[0136] In this embodiment, S6 includes the following sub-steps:

[0137] S61. Perform a hash operation on the dynamic panoramic ledgers of multiple target network terminal devices (NTDs) in the same computing power collaborative cluster (PON-Cluster) to obtain a ledger snapshot summary;

[0138] S62. Perform a hash operation on the authorization windows of multiple target network terminal devices (NTDs) in the same computing power collaborative cluster (PON-Cluster) to obtain the authorization sequence fingerprint;

[0139] S63. Calculate the order consistency index based on the actual logical fragmentation order and the expected aggregation order of the aggregate points.

[0140] S64. Combine the ledger snapshot summary, authorization sequence fingerprint, and sequence consistency indicator to generate a scheduling audit record, and append the scheduling audit record to the chained audit record set;

[0141] S65. When the aggregation result of logical sharding needs to be transmitted from the current computing power collaborative cluster PON-Cluster to the parent domain, cross-domain overflow processing is performed, an overflow audit record is generated and appended to the chained audit record set.

[0142] In S6, the scheduling cycle refers to the time interval during which the Optical Line Terminal (OLT) "generates a new authorization window sequence, hashes and calculates the key metadata (dynamic panoramic ledger, authorization window sequence, actual arrival order) for that round of scheduling, generates scheduling audit records, and appends them to the chained audit record set."

[0143] The expression for obtaining the ledger snapshot summary in S61 is:

[0144] ,

[0145] in, For the first A summary of the ledger snapshot at any given moment. For the first target network terminal device NTD in the 1st A dynamic panoramic ledger in real time. For the second target network terminal device NTD in the A dynamic panoramic ledger in real time. For the first The target network terminal device NTD is in the first A dynamic, panoramic ledger in real time; || is the splicing symbol. For hash functions, This refers to the number of target network terminal devices (NTDs) under a computing power collaborative cluster (PON-Cluster).

[0146] The expression for obtaining the authorized sequence fingerprint in S62 is:

[0147] ,

[0148] in, For the first Authorization sequence fingerprint at any moment This is the authorization window for the first target network terminal device (NTD) under the PON-Cluster computing power collaboration cluster. This is the authorization window for the second target network terminal device (NTD) under the PON-Cluster computing power collaboration cluster. For the first PON-Cluster of computing power collaboration The authorization window of the target network terminal device NTD;

[0149] The expression for calculating the order consistency index in S63 is as follows:

[0150] ,

[0151] in, For the first Time-order consistency index For the first The logical shard index corresponding to each logical shard that needs to be aggregated. The number of times the actual arrival at the aggregation point A logical sharded index, This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise.

[0152] The expression for the audit record is:

[0153] ,

[0154] in, For the first Real-time audit records, This refers to the current moment.

[0155] Cross-domain transmission allows for "overflow to the parent domain"—that is, data is transmitted from the target network terminal device (NTD) within the PON-Cluster to the optical line terminal (OLT), or from the OLT to a higher-level network.

[0156] In this embodiment, in S65, after the logical fragmentation of the target network terminal device NTD is completed and aggregated in sequence on the backup target, if the aggregation result needs to be transmitted from the current optical line terminal (OLT) to the parent domain, a cross-domain overflow requirement is marked, a whitelist conversion and size check are performed on the aggregation result, an overflow audit record containing the current time, the current OLT identifier, the converted data and the parent domain target address is generated, the record is appended to the chained audit record set, and the converted data is transmitted to the parent domain target address.

[0157] The specific process of S65 includes: after logical fragmentation is completed and aggregated in sequence on the backup target, if the aggregation result needs to be transmitted from the current optical line terminal (OLT) to the parent domain, a whitelist conversion is performed on the aggregation result to obtain the converted overflow data; it is determined whether the size of the converted overflow data is less than or equal to the maximum capacity of the overflow data. If so, an overflow audit record is generated and appended to the chained audit record set, and then the converted overflow data is transmitted to the target address of the parent domain; otherwise, the overflow transmission is rejected.

[0158] Overflow audit logs include: current time, current optical line terminal (OLT), converted overflow data, and parent domain target address.

[0159] "Overflow to the parent domain" can also be performed within the PON-Cluster. The specific process includes: after the logical fragment is executed on the target network terminal device (NTD), if the fragment execution result needs to be transmitted from the current PON-Cluster to the parent domain (OLT), then the fragment execution result is whitelisted to obtain the converted overflow data; it is determined whether the size of the converted overflow data is less than or equal to the maximum capacity of the overflow data. If so, an overflow audit record is generated and appended to the chained audit record set, and then the converted overflow data is transmitted to the target address of the parent domain; otherwise, the overflow transmission is rejected.

[0160] The sliced ​​execution results include: KV cache, inter-layer activation summary (the quantized and compressed representation of the activation values ​​output by the neural network layer), quantized logits fragment (the quantized and compressed fragment of the probability logs before the model output layer), and gradient block summary (a gradient compressed representation used for lightweight LLM training, distillation, or fine-tuning scenarios).

[0161] This invention achieves immutable and chained storage of key metadata in the scheduling process; when the aggregation result needs to be transmitted across domains, S65 performs whitelist conversion and size verification, generates an overflow audit record containing the current time, the current optical line terminal (OLT) identifier, the converted data, and the parent domain target address, and appends it to the chained audit record set, and transmits the converted data to the parent domain target address, thus realizing controlled execution and complete audit traceability of cross-domain overflow.

[0162] A whitelist converts raw data into a predefined allowed data format.

[0163] Example 2: A communication scheduling system based on PON physical layer sequential aggregation, comprising:

[0164] The physical-logical isomorphic mapping module is used to construct the physical network graph of the optical distribution network (ODN), divide the physical node set in the physical network graph into computing power collaborative clusters (PON-Cluster), and establish a physical-logical isomorphic mapping from the logical topology graph of the distributed inference task to the physical network graph to obtain the target network terminal device (NTD).

[0165] The task distribution module is used to generate logical fragments for the distributed inference tasks to be processed, determine the expected aggregation order of the logical fragments, and distribute each logical fragment to the corresponding target network terminal device (NTD).

[0166] The panoramic ledger construction module is used to periodically embed an in-band real-time state vector into the extended XGEM frame header of each target network terminal device (NTD) during execution and send it to the optical line terminal (OLT) to construct a dynamic panoramic ledger.

[0167] The reverse scheduling module is used on the optical line terminal (OLT) to predict the completion time of each logical segment based on the dynamic panoramic ledger, and generate an authorization window sequence that satisfies phase-locked loop and in-order constraints.

[0168] The authorization execution module is used to broadcast the authorization window sequence to all target network terminal devices (NTDs). Each target network terminal device (NTD) uploads its logical fragment execution results within the settings of the corresponding authorization window and records the actual logical fragment order that arrives at the aggregation point.

[0169] The audit module is used to generate and store audit records in each scheduling cycle of the optical line terminal (OLT) based on the logical fragmentation order of actual arrival at the aggregation point.

[0170] The specific implementation process of Example 2 is the same as that of Example 1.

[0171] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A communication scheduling method based on PON physical layer sequential aggregation, characterized in that, Includes the following steps: S1. Construct the physical network diagram of the optical distribution network (ODN), divide the physical node set in the physical network diagram into computing power collaborative clusters (PON-Cluster), and establish a physical-logical isomorphic mapping from the logical topology diagram of the distributed inference task to the physical network diagram to obtain the target network terminal device (NTD). S2. Generate logical fragments for the distributed inference task to be processed, determine the expected aggregation order of the logical fragments, and send each logical fragment to the corresponding target network terminal device (NTD). S3. During execution, each target network terminal device (NTD) periodically embeds an in-band real-time state vector in the extended XGEM frame header of its uplink data frame and sends it to the optical line terminal (OLT) to construct a dynamic panoramic ledger. S4. On the optical line terminal (OLT), based on the dynamic panoramic ledger, predict the completion time of each logical segment and generate an authorization window sequence that satisfies phase-locked loop and in-order constraints. S5. Broadcast the authorization window sequence to all target network terminal devices (NTDs). Each target network terminal device (NTD) uploads its logical fragment execution results within the settings of the corresponding authorization window and records the actual logical fragment order that arrives at the aggregation point. S6. Based on the logical fragmentation order of actual arrival at the aggregation point, the optical line terminal (OLT) generates and stores audit records in each scheduling cycle. S1 includes the following steps: S11. Construct the physical network diagram of the optical distribution network (ODN), and divide the physical node set in the physical network diagram into computing power collaborative clusters (PON-Cluster); S12. Establish a physical-logical isomorphic mapping from the set of logical computing units in the logical topology graph of the distributed reasoning task to the set of physical nodes in the physical network graph. S13. Set multiple constraints on the physical-logical isomorphic mapping to obtain the target network terminal device NTD; The physical network diagram of the optical distribution network (ODN) in S1 includes: a set of physical nodes and a set of physical links. The set of physical nodes includes: optical line terminal (OLT), optical splitter (Splitter), and network terminal equipment (NTD). The set of physical links represents optical fiber connections. The process of dividing the computing power collaboration cluster PON-Cluster in S1 includes: taking each optical splitter in the physical node set of the physical network diagram as the root node, extracting the subtree of the root node, and classifying all physically connected network terminal devices (NTDs) under the subtree of the root node into a computing power collaboration cluster PON-Cluster. The logical topology graph in S1 includes: a set of logical computing units and a set of logical edges. The set of logical computing units includes: data units of task partitioning, operators and large model LLM. The set of logical edges represents the dependencies or data flow between logical computing units. The multiple constraints in S13 include: computing power collaborative cluster PON-Cluster constraint, routing isomorphism constraint, and overflow isomorphism constraint; The constraints of the PON-Cluster (Power-Oriented Networking Cluster) include: Each logical computing unit allocated to the PON-Cluster constitutes a subset of logical computing units, and the mapping result satisfies: ,in, To be allocated to the first Each logical computing unit in a PON-Cluster constitutes a subset of logical computing units. It is a physical-logical isomorphic mapping symbol. For the first NTD, a network terminal device For the first The physical subtree of the root node is a single optical splitter. and It is a positive integer; Routing isomorphic constraints include: ,in, For a set of logical edges, A set of physical links; Overflow isomorphism constraints include: , ,in, For overflow data, For a set of data types that are allowed to overflow, The amount of data that overflows. Maximum capacity for overflow data; S3 includes the following steps: S31. During the execution of logical fragmentation, each target network terminal device (NTD) periodically embeds an in-band real-time state vector in the extended XGEM frame header of its uplink data frame. S32. Receive the in-band real-time state vector sent by the target network terminal device NTD through the optical line terminal (OLT); S33. The out-of-band computing power ledger and the in-band real-time state vector are merged through the optical line terminal (OLT) to obtain the dynamic panoramic ledger of the target network terminal device (NTD). The in-band real-time state vector in S3 includes: reporting timestamp, NPU busy / idle status, LLM inference progress percentage, available memory, temperature, and preemption alert; The process of acquiring the out-of-band computing power ledger in S33 includes: acquiring the out-of-band computing power ledger for each network terminal device (NTD) belonging to the physical node set through the network management and control channel. The out-of-band computing power ledger includes: collection timestamp, computing power of the network terminal device (NTD), memory capacity of the network terminal device (NTD), hardware health status of the network terminal device (NTD), and link quality of the network terminal device (NTD). S4 includes the following steps: S41. Based on the dynamic panoramic ledger of the target network terminal device NTD, predict the time when the logical fragment is completed in the calculation of the target network terminal device NTD. S42. Based on the calculation completion time of the target network terminal device NTD, generate an authorization window sequence that satisfies phase-locked loop constraints and sequential constraints for multiple target network terminal devices NTD of the same computing power collaborative cluster PON-Cluster. The phase-locked constraint in S42 is: ， in, For the first The start time of the authorization window of the target network terminal device (NTD). For the first The completion time of the calculation of the target network terminal device NTD. This is the preset phase-locked loop tolerance; | represents the absolute value operation. It is a positive integer; The sequential constraint in S42 is: ， in, For the first The start time of the authorization window of the target network terminal device (NTD). For the first The start time of the authorization window of the target network terminal device (NTD). For the first The duration of the authorization window for each target network terminal device (NTD). For the first The protection interval of the authorization window of a target network terminal device (NTD). For the first The logical shard index corresponding to each logical shard that needs to be aggregated. For the first The logical shard index corresponding to each logical shard that needs to be aggregated. It is a positive integer; The authorization windows in the authorization window sequence of S42 are: ,in, For the first PON-Cluster of computing power collaboration The authorization window of the target network terminal device NTD. For the first The duration of the authorization window for each target network terminal device (NTD). For the first The protection interval between the authorization window of a target network terminal device (NTD) and the next authorization window.

2. The communication scheduling method based on PON physical layer sequential aggregation according to claim 1, characterized in that, S5 includes the following steps: S51. Write each authorized window in the authorized window sequence into the PON bandwidth allocation table and send it to the target network terminal device NTD. S52. Determine whether the service assurance plane resources of the target network terminal device NTD need to occupy the computing power plane resources. If so, set the preemption warning flag to be valid, save the micro-checkpoint of the current logical fragment. The micro-checkpoint includes KV cache, token cursor and inter-layer activation digest. Send the preemption warning flag to the optical line terminal OLT through the uplink XGEM frame header, and jump to step S53. Otherwise, the preemption warning flag remains invalid, the logical fragmentation continues to execute normally, and the process jumps to step S54; S53. Through the optical line terminal (OLT), select the target network terminal device (NTD) with the lowest NPU utilization in the same computing power collaborative cluster (PON-Cluster) as the backup target. Move the logical fragments and micro-checkpoints in the target network terminal device (NTD) corresponding to the preemption warning flag to the backup target, update the mapping relationship between the logical fragments and the target network terminal device (NTD), and recalculate the calculation completion time of the logical fragments. S54. In the target network terminal device NTD, the actual calculation completion time is sent according to the authorization window setting, and the logical fragmentation order of actual arrival at the aggregation point is recorded.

3. The communication scheduling method based on PON physical layer sequential aggregation according to claim 1, characterized in that, S6 includes the following steps: S61. Perform a hash operation on the dynamic panoramic ledgers of multiple target network terminal devices (NTDs) in the same computing power collaborative cluster (PON-Cluster) to obtain a ledger snapshot summary; S62. Perform a hash operation on the authorization windows of multiple target network terminal devices (NTDs) in the same computing power collaborative cluster (PON-Cluster) to obtain the authorization sequence fingerprint; S63. Calculate the order consistency index based on the actual logical fragmentation order and the expected aggregation order of the aggregate points. S64. Combine the ledger snapshot summary, authorization sequence fingerprint, and sequence consistency indicator to generate a scheduling audit record, and append the scheduling audit record to the chained audit record set; S65. When the aggregation result of logical sharding needs to be transmitted from the current computing power collaborative cluster PON-Cluster to the parent domain, cross-domain overflow processing is performed, an overflow audit record is generated and appended to the chained audit record set.

4. The communication scheduling method based on PON physical layer sequential aggregation according to claim 3, characterized in that, In S65, after the logical fragmentation of the target network terminal device (NTD) is completed and aggregated sequentially on the backup target, if the aggregation result needs to be transmitted from the current optical line terminal (OLT) to the parent domain, the cross-domain overflow requirement is marked, whitelist conversion and size verification are performed on the aggregation result, and an overflow audit record containing the current time, the current OLT identifier, the converted data and the parent domain target address is generated. This record is then appended to the chained audit record set, and the converted data is transmitted to the parent domain target address.

5. A PON-based physical layer sequential aggregation communication scheduling system, implemented based on the PON-based physical layer sequential aggregation communication scheduling method according to any one of claims 1 to 4, characterized in that, include: The physical-logical isomorphic mapping module is used to construct the physical network graph of the optical distribution network (ODN), divide the physical node set in the physical network graph into computing power collaborative clusters (PON-Cluster), and establish a physical-logical isomorphic mapping from the logical topology graph of the distributed inference task to the physical network graph to obtain the target network terminal device (NTD). The task distribution module is used to generate logical fragments for the distributed inference tasks to be processed, determine the expected aggregation order of the logical fragments, and distribute each logical fragment to the corresponding target network terminal device (NTD). The panoramic ledger construction module is used to periodically embed an in-band real-time state vector into the extended XGEM frame header of each target network terminal device (NTD) during execution and send it to the optical line terminal (OLT) to construct a dynamic panoramic ledger. The reverse scheduling module is used on the optical line terminal (OLT) to predict the completion time of each logical segment based on the dynamic panoramic ledger, and generate an authorization window sequence that satisfies phase-locked loop and in-order constraints. The authorization execution module is used to broadcast the authorization window sequence to all target network terminal devices (NTDs). Each target network terminal device (NTD) uploads its logical fragment execution results within the settings of the corresponding authorization window and records the actual logical fragment order that arrives at the aggregation point. The audit module is used to generate and store audit records in each scheduling cycle of the optical line terminal (OLT) based on the logical fragmentation order of actual arrival at the aggregation point.

Images