A method and system for determining assurance of out-of-band management of a passive optical network

Abstract

This invention discloses a deterministic guarantee method and system for out-of-band management of passive optical networks (PONs), relating to the fields of PON communication and data center management technology. It aims to solve the problems of high latency uncertainty, lack of traffic tiering, and the contradiction between sleep and response capabilities in traditional solutions. This invention distinguishes between critical and routine management traffic using a traffic classifier, employs a self-attention temporal prediction neural network based on a Transformer architecture to dynamically generate transmission time windows, and uses a sleep arbitrator to prematurely wake up the physical layer transceiver. A dual-path dynamic bandwidth allocation engine is then used to achieve tiered scheduling. This invention can provide microsecond-level deterministic latency guarantees for critical management traffic while reducing device power consumption, making it suitable for out-of-band management scenarios in artificial intelligence data centers.

Description

Technical Field

[0001] This invention relates to the fields of passive optical network communication and data center management technology, and in particular to a deterministic guarantee method and system for out-of-band management of passive optical networks. Background Technology

[0002] With the rapid expansion of AI data centers, out-of-band management (OOBM) networks, as the core infrastructure for remote operation and maintenance and fault recovery of servers and storage devices, are facing increasing demands for deterministic transmission and low power consumption. Passive optical networks (PONs), with their advantages of simple cabling, high bandwidth, and strong anti-interference capabilities, are gradually replacing traditional copper Ethernet as the mainstream architecture for out-of-band management in large data centers.

[0003] Current mainstream out-of-band management solutions for passive optical networks (PONs), such as Ciena's DCOM solution and Nokia's Aurelius data center optical management solution, all rely on traditional PON architectures to build out-of-band management channels and depend on Dynamic Bandwidth Allocation (DBA) mechanisms for bandwidth scheduling. However, existing solutions have the following drawbacks:

[0004] First, there is a lack of deterministic latency guarantees. Traditional dynamic bandwidth allocation mechanisms handle latency in the millisecond range, which cannot meet the microsecond-level deterministic latency requirements of critical management operations in artificial intelligence data centers, such as emergency recovery of the Baseboard Management Controller (BMC), emergency stop commands for the computing cluster, and security key rotation.

[0005] Second, there is no tiered processing mechanism for management traffic. The existing solution uses a unified scheduling queue for all management traffic. Critical management traffic shares bandwidth resources with regular heartbeat and log collection traffic. When the network is congested, critical management traffic is easily blocked and delayed.

[0006] Third, there is a contradiction between the hibernation mechanism and deterministic requirements. Existing passive optical network energy-saving solutions reduce power consumption through physical layer hibernation, but the wake-up latency in hibernation is high, making it impossible to respond to sudden critical management traffic in a timely manner; if it is kept running 24 / 7, the power consumption of the device remains high.

[0007] Fourth, there is a technological gap in the integration of Time-Sensitive Networking (TSN) and Passive Optical Networking (PON). Existing TSN scheduling technologies are mostly applied to industrial Ethernet and automotive network scenarios, and have not been deeply integrated with PON architectures for out-of-band management scenarios in data centers; traditional Gate Control Lists (GCLs) are statically configured and cannot adapt to the dynamic fluctuations in data center management traffic.

[0008] Therefore, there is an urgent need in this field for an out-of-band management solution for passive optical networks that can achieve hierarchical scheduling of traffic management, has the ability to generate dynamic deterministic time windows, and simultaneously takes into account low power consumption and fast response. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a deterministic guarantee method and system for out-of-band management of passive optical networks. By dynamically generating transmission time windows through a self-attention timing prediction neural network and combining a dual-path dynamic bandwidth allocation architecture with a sleep linkage mechanism, it achieves microsecond-level deterministic latency guarantee for critical management traffic while reducing device power consumption.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A deterministic guarantee method for out-of-band management of passive optical networks includes the following steps:

[0012] S1. The optical line terminal receives management traffic from the management terminal and identifies the management traffic according to preset classification rules through a traffic classifier, dividing it into critical management traffic and regular management traffic.

[0013] S2. If the traffic is identified as critical management traffic, the pre-trained self-attention timing prediction neural network model is invoked to predict the arrival time of the critical management traffic based on historical traffic data and dynamically generate the corresponding transmission time window; the sleep arbitrator calculates the wake-up time of the physical layer transceiver based on the global clock and the transmission time window, and wakes up the physical layer transceiver in advance before the transmission time window arrives.

[0014] S3. Allocate transmission bandwidth for critical management traffic through the fast path hardware dynamic bandwidth allocation module, and transmit the critical management traffic to the target optical network unit via the passive optical link;

[0015] S4. If identified as regular management traffic, the transmission bandwidth is allocated to it through the slow path software dynamic bandwidth allocation module, and the regular management traffic is transmitted to the target optical network unit via the passive optical link; after the traffic transmission is completed, the physical layer transceiver enters a sleep state during idle periods.

[0016] Preferably, the preset classification rules are formulated based on traffic type, source address, destination address, or virtual LAN tag; wherein emergency stop command, device emergency recovery command, and security key rotation traffic are marked as critical management traffic, and heartbeat message, operation log collection, and configuration synchronization traffic are marked as regular management traffic.

[0017] Preferably, the input of the self-attention temporal prediction neural network model is multi-dimensional management traffic time-series data within a historical period, and the output is the dynamic time window arrangement corresponding to each type of traffic; the time window arrangement of key management traffic has a higher priority than that of regular management traffic, and the time windows of different categories do not overlap.

[0018] Preferably, in step S2, when the hibernation arbitrator calculates the wake-up time, the wake-up advance is not less than the sum of the physical layer transceiver's wake-up delay and the preset safety margin, ensuring that the physical layer transceiver is ready before the arrival of critical management traffic.

[0019] Preferably, the fast path hardware dynamic bandwidth allocation module implements bandwidth allocation based on hardware logic, with a processing latency in the microsecond range; the slow path software dynamic bandwidth allocation module implements bandwidth allocation based on software algorithms, supporting multi-dimensional scheduling strategies.

[0020] A deterministic guarantee system for out-of-band management of a passive optical network includes an optical line terminal and several optical network units deployed in each rack. The optical line terminal has a built-in traffic classifier, a self-attention timing prediction and scheduling module, a sleep arbitrator, a hardware dynamic bandwidth allocation module for fast paths, a software dynamic bandwidth allocation module for slow paths, and a physical layer transceiver.

[0021] The traffic classifier is used to receive management traffic sent by the management terminal and divide it into critical management traffic and regular management traffic according to preset rules;

[0022] The self-attention timing prediction scheduling module is used to call a pre-trained self-attention timing prediction neural network to predict the arrival time of key management traffic based on historical traffic data and dynamically generate a transmission scheduling schedule for various types of traffic.

[0023] The sleep arbitrator is used to control the sleep and wake-up switching of the physical layer transceiver according to the scheduling information of the transmission scheduling schedule and the global clock;

[0024] The fast path hardware dynamic bandwidth allocation module is used to perform low-latency bandwidth allocation for critical management traffic.

[0025] The slow path software dynamic bandwidth allocation module is used to perform bandwidth allocation scheduling for regular management traffic.

[0026] The physical layer transceiver is used to realize the transmission and reception of optical signals between the optical line terminal and the optical network unit, and supports the switching between sleep and wake-up states.

[0027] Preferably, the traffic classifier supports classification rules based on traffic type, address information, or virtual LAN tags, and allows for customization of identification conditions for critical management traffic and regular management traffic.

[0028] Preferably, in the transmission scheduling timetable generated by the self-attention timing prediction scheduling module, the time window of critical management traffic has a higher priority than that of regular management traffic, the time windows do not overlap within the scheduling period, and the window parameters can be dynamically adjusted according to the real-time traffic status.

[0029] Preferably, the hibernation arbiter is synchronized with the global clock and can calculate the wake-up advance time based on the wake-up delay parameters of the physical layer transceiver and the preset safety margin, so as to complete the wake-up of the physical layer transceiver before the time window of the critical management traffic arrives.

[0030] Preferably, the fast path hardware dynamic bandwidth allocation module is a hardware logic circuit with a processing latency in the microsecond range; the slow path software dynamic bandwidth allocation module runs on the main control processor and supports complex scheduling algorithms.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] This invention dynamically generates deterministic transmission time windows using a self-attention temporal prediction neural network based on a Transformer architecture, and combines this with a fast path hardware dynamic bandwidth allocation mechanism to control the end-to-end latency of critical management traffic at the microsecond level, thus meeting the deterministic requirements of critical management operations in artificial intelligence data centers.

[0033] This invention employs a dual-engine architecture of fast path hardware scheduling and slow path software scheduling to achieve physical isolation scheduling of critical management traffic and regular management traffic, thereby preventing critical traffic from being blocked by regular traffic in congested scenarios.

[0034] This invention links a hibernation arbitrator with a dynamic scheduling schedule to wake up the physical layer transceiver in advance before the critical management traffic time window arrives, and controls the physical layer to enter hibernation during idle periods, thereby reducing the device's operating power consumption while ensuring zero latency penalty.

[0035] Compared to traditional static gating control list schemes, this invention uses a self-attention-based time-series prediction model that can dynamically adjust time window parameters according to historical traffic patterns, adapting to the dynamic fluctuation characteristics of data center management traffic and improving bandwidth resource utilization. Attached Figure Description

[0036] Figure 1 is a schematic diagram of the architecture of the passive optical network out-of-band management deterministic guarantee system according to the present invention;

[0037] Figure 2 is a flowchart illustrating the deterministic guarantee method for out-of-band management of passive optical networks according to the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0039] Example 1

[0040] like Figure 2 As shown, this embodiment is a deterministic guarantee method for out-of-band management of passive optical networks. It fully describes the execution logic of the entire process from management traffic access to transmission completion and sleep control. All steps are executed collaboratively on the optical line terminal (OLT) side, and work with the rack-side optical network unit (ONU) to complete end-to-end transmission.

[0041] S1. Traffic Access Management and Classification Steps:

[0042] This step is the entry point of the method. Its core function is to analyze, identify, and divert all out-of-band management traffic entering the optical line terminal, so as to achieve physical and logical separation between critical management traffic and regular management traffic, and provide a prerequisite for subsequent hierarchical scheduling.

[0043] The uplink management interface of the optical line terminal receives management traffic from the data center control plane. Message types include Ethernet frames, Internet Protocol (IP) messages, and out-of-band management protocol messages. Upon message access, header parsing is performed first to extract the message's five-tuple information (source address, destination address, source port, destination port, and protocol type), Virtual Local Area Network (VLAN) tag, message priority field, message payload type identifier, and message length information. The parsing process is completed by a hardware message parsing engine, with a single message parsing latency of no more than 1 microsecond, supporting line-speed processing and avoiding additional latency introduced during the classification stage.

[0044] After parsing, the message feature information is sent to the classification and matching engine, and the complete message is stored in the corresponding cache queue. The cache queue adopts a dual-queue architecture, corresponding to the critical management traffic cache area and the regular management traffic cache area respectively. The queue depth is configured according to the business scale, and the cache capacity of a single queue is not less than 10,000 messages to avoid packet loss caused by sudden traffic.

[0045] The traffic classifier has a built-in configurable classification rule table that supports multi-dimensional matching. Matching dimensions include six categories: traffic service type, source device type, destination device type, VLAN priority, packet DSCP field, and custom packet identifier bits. The rule table is implemented in hardware using ternary content addressable memory (TCAM), and the matching process is completed in a single cycle with a matching latency of no more than 200 nanoseconds.

[0046] The preset classification rules in this embodiment are as follows:

[0047] Critical management traffic determination rules: Packets that meet any of the following conditions are marked as critical management traffic:

[0048] The message service types are: cluster emergency stop command, baseboard management controller emergency reset command, security key forced rotation command, and hardware-level fault circuit interruption command;

[0049] The message source address belongs to the security management zone server network segment, and the destination address belongs to the computing cluster baseboard management controller network segment;

[0050] Configure the VLAN priority field of the packet to the highest level (priority value 7), or mark the DSCP field as the accelerated forwarding class;

[0051] The message header carries a custom emergency management flag.

[0052] Regular management traffic determination rules: All management traffic that does not meet the key management traffic determination conditions is marked as regular management traffic, including baseboard management controller heartbeat messages, device operation log reports, batch configuration synchronization, environmental sensor data acquisition, firmware upgrade messages, etc.

[0053] After matching is completed, the classifier adds an internal traffic category tag to the packet. This tag only circulates within the optical line terminal and does not modify the original packet content, ensuring the compatibility of the management protocol.

[0054] Based on traffic category labels, the classifier performs traffic splitting: packets labeled as critical management traffic are sent to a dedicated critical traffic buffer queue, simultaneously triggering subsequent timing prediction and scheduling processes; packets labeled as regular management traffic are sent to a regular traffic buffer queue, awaiting slow path scheduling. The traffic splitting process is implemented using a hardware switching matrix, eliminating software forwarding overhead and ensuring low latency for critical traffic paths.

[0055] S2. Critical Management Traffic Timing Prediction and Sleep / Wake-Up Scheduling Steps:

[0056] This step is the core innovation of the method. It replaces the traditional static gate control list (GCL) scheme and adopts a self-attention temporal prediction neural network based on the Transformer architecture. It dynamically generates the transmission time window according to the historical traffic pattern and links the sleep arbitrator to complete the early wake-up of the physical layer transceiver, achieving a combination of determinism and low power consumption.

[0057] The input to the self-attention temporal prediction neural network is historical time-series traffic data. The data sampling granularity matches the scheduling cycle. In this embodiment, the default scheduling cycle is configured to be 5000 microseconds, that is, a scheduling table is generated every 5 milliseconds. The historical data window length is configured to be the past 128 scheduling cycles, corresponding to a total duration of 640 milliseconds, which can cover the short- and medium-cycle fluctuation patterns of data center management traffic.

[0058] The input features have eight dimensions, specifically including: the number of critical management traffic packets, total number of bytes, packet arrival time distribution entropy, peak traffic intensity, proportion of emergency stop commands, proportion of emergency reset commands, proportion of safety commands, and window utilization rate of the previous cycle. Before inputting the raw data into the model, normalization is performed, mapping all features to the [0,1] interval. The normalization formula is as follows:

[0059]

[0060] in, These are the original eigenvalues. This is the minimum value of the feature in the training set. This represents the maximum value of the feature in the training set. Normalization improves the model's convergence speed and prediction accuracy.

[0061] The self-attention temporal prediction neural network in this embodiment is built on a Transformer encoder architecture, consisting of 6 encoder layers, each configured with an 8-head self-attention mechanism, and a feedforward network hidden layer dimension of 512. The model has approximately 2 million parameters, making it a lightweight temporal prediction model. It can complete single-cycle inference on a Field Programmable Gate Array (FPGA) accelerator card, with an inference latency of no more than 50 microseconds, far less than the scheduling cycle length, and will not affect the real-time performance of scheduling.

[0062] The calculation process for single-head self-attention satisfies the following formula:

[0063]

[0064] in, The query matrix is ​​obtained by linear transformation of the input features; The key matrix is ​​obtained from the input features through a linear transformation. The value matrix is ​​obtained from the input features through a linear transformation; is the dimension of the key vector, which is 64 in this embodiment. Divided by The purpose of this is to scale the dot product result to prevent the gradient of the softmax function from vanishing due to excessively large dot product values.

[0065] The multi-head self-attention mechanism maps input features to eight different subspaces, calculates the self-attention results for each subspace, and then concatenates them. This allows for the simultaneous capture of multi-dimensional dependent features of traffic flow time series, including long-term trends, short-term bursts, and periodic fluctuations. The multi-head self-attention calculation formula is as follows:

[0066]

[0067]

[0068] in, To focus on the number of heads, in this embodiment ; , , For the first The linear transformation matrix corresponding to each attention head; To output the linear transformation matrix.

[0069] In addition to the self-attention sublayer, the encoder layer also includes a feedforward network sublayer and layer normalization and residual connection structures. The feedforward network adopts a two-layer fully connected structure, using a Gaussian Error Linear Unit (GELU) as the activation function in the middle, calculated as follows:

[0070]

[0071] in, , This is the weight matrix. , This is the bias vector.

[0072] The model outputs predicted features of critical management traffic in the next scheduling cycle, including three core output values: predicted arrival time, predicted total traffic volume, and predicted traffic duration, which serve as the basis for dynamically generating time windows.

[0073] The model employs an offline pre-training + online fine-tuning training mode: In the offline phase, it is trained using three months of out-of-band management traffic data from the data center. The training optimizer uses the Adaptive Moment Estimation (Adam) optimizer, with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. The loss function is the Mean Squared Error (MSE) loss, calculated as follows:

[0074]

[0075] in, For the sample size, For the true value, These are the model's predicted values. During the online phase, the model is fine-tuned every 24 hours using the day's traffic data to ensure continuous optimization of prediction accuracy as business changes occur.

[0076] Based on the predictions output by the model, the scheduler generates the transmission time window configuration for the next cycle. The time window generation follows three principles: "critical priority, no overlap, and allowance for margin," with the specific generation rules as follows:

[0077] Critical management traffic time windows are prioritized at the beginning of the scheduling cycle to ensure that critical traffic receives the highest scheduling priority.

[0078] The minimum window duration required to calculate the predicted traffic volume is calculated using the following formula:

[0079]

[0080] in, The duration of the critical management traffic time window. For the predicted total number of bytes of critical management traffic, For the effective downlink bandwidth, To protect the interval duration, the default configuration in this embodiment is 20 microseconds, which is used to offset prediction errors and hardware response deviations.

[0081] Regular management traffic time windows are scheduled after critical traffic windows, occupying the remaining available time within the cycle;

[0082] Set a protection interval between all time windows to avoid traffic conflicts at window boundaries.

[0083] The generated time window must meet the following constraints before it can take effect:

[0084]

[0085]

[0086]

[0087] in, This is a traffic category index, with a value range of [value range missing]. , Corresponding to key management traffic, Corresponds to regular management traffic; For the first The start time of the traffic-like time window; This represents the total duration of the scheduling cycle. The first constraint ensures that a single type of traffic window does not exceed the cycle range, the second constraint ensures that critical traffic windows are prioritized, and the third constraint ensures that two types of traffic windows do not overlap.

[0088] If the forecast results show that there is no critical management traffic in the current period, the critical traffic window allocation will be skipped, and all period resources will be allocated to regular management traffic.

[0089] After the time window configuration takes effect, the hibernation arbiter calculates the wake-up time of the physical layer transceiver based on the start time of the critical management traffic time window. The hibernation arbiter maintains real-time synchronization with the global PTP clock, with a clock deviation of no more than 1 microsecond, ensuring the accuracy of the wake-up time.

[0090] The calculation of wake-up lead time follows the formula below:

[0091]

[0092] in, The wake-up lead time is the time difference between the physical layer transceiver wake-up time and the start time of the time window. The inherent wake-up latency of the physical layer transceiver is determined by hardware characteristics. In this embodiment, the typical wake-up latency of the physical layer chip from deep sleep mode is 20 microseconds, and the maximum latency does not exceed 30 microseconds. To provide a safety margin to offset uncertainties such as clock synchronization deviation, hardware response fluctuations, and scheduling jitter, the default configuration in this embodiment is 10 microseconds.

[0093] Based on the above formula, the minimum wake-up advance time in this embodiment is 30 microseconds, while 35 microseconds is used in the engineering configuration to reserve additional redundancy. If the physical layer transceiver is currently in normal operating mode rather than sleep mode, the wake-up step is skipped, and bandwidth allocation is performed directly.

[0094] At the calculated wake-up time, the sleep arbiter sends a wake-up command to the physical layer transceiver. After receiving the command, the physical layer transceiver sequentially completes four wake-up stages: powering on the internal circuits, stabilizing the clock, activating the optical module, and synchronizing the link. After all stages are completed, it enters normal operation and returns a wake-up completion signal to the sleep arbiter.

[0095] If an anomaly occurs during the wake-up process and the physical layer transceiver fails to return a completion signal within the expected time, the hibernation arbitrator triggers the anomaly handling mechanism, forcibly resets the physical layer chip and reports an alarm, and simultaneously switches subsequent critical traffic to the backup transmission path to ensure that management commands are not interrupted.

[0096] S3. Critical Management Traffic Fast Path Bandwidth Allocation and Transmission Steps:

[0097] This step is the transmission execution stage for critical management traffic. It adopts a hardware-level fast path dynamic bandwidth allocation (DBA) engine, with no central processing unit intervention throughout the process, to achieve microsecond-level bandwidth scheduling and ensure end-to-end deterministic latency for critical traffic.

[0098] Packets in the critical management traffic buffer queue are sent to the fast path hardware DBA module for bandwidth requests. The request packet contains four core pieces of information: the target optical network unit number, the requested bandwidth size, the packet priority, and the time window deadline. The hardware DBA module employs a pipelined architecture design, capable of processing bandwidth requests from 32 optical network units simultaneously. The request processing is completed in a single clock cycle, with no queuing delays.

[0099] The Fast Path Hardware DBA uses a fixed time slot allocation algorithm and employs strict priority scheduling for critical management traffic. Within the reserved time window, bandwidth is allocated sequentially according to the order of request arrival, without performing complex fairness scheduling and congestion control, with the lowest latency as the primary objective.

[0100] The bandwidth allocation calculation process consists of only two steps: time slot mapping and bandwidth verification.

[0101] Time slot mapping: Based on the requested bandwidth, the corresponding number of downlink time slots are allocated to the target optical network unit, with a time slot granularity of 1 microsecond;

[0102] Bandwidth verification: Verify whether the total allocated bandwidth exceeds the available bandwidth within the time window. If it does, requests will be fulfilled in descending order of priority, and low-priority requests will be postponed to the next cycle.

[0103] In this embodiment, the single request processing latency for the fast path DBA is no more than 5 microseconds, which is two orders of magnitude higher than the millisecond-level processing latency of the traditional software DBA.

[0104] After bandwidth allocation is completed, the hardware scheduling engine maps packets to allocated time slots, generates Medium Access Control (MAC) layer frames, adds frame headers, checksums, optical network unit address identifiers, and other information, and then sends them to the physical layer transceiver for optical signal modulation. The entire encapsulation process is completed by hardware logic, with an encapsulation delay of no more than 1 microsecond.

[0105] The physical layer transceiver modulates the electrical signal into an optical signal, which is then transmitted to each optical network unit (ONU) via a passive optical distribution network (ODN). Upon receiving the optical signal, the ONU parses the destination address in the frame header; packets matching the local address are then sent to the corresponding baseboard management controller.

[0106] The end-to-end total latency of critical management traffic satisfies the following constraint formula:

[0107]

[0108] in, For critical management of total end-to-end latency; The latency for traffic classification and processing is typically 1 microsecond. This is for timing prediction and scheduling processing latency, typically 50 microseconds. This is the physical layer transceiver wake-up latency, typically 20 microseconds. For fast-path DBA processing latency, the typical value is 5 microseconds; For fiber optic link transmission delay, the typical value for 20km fiber is approximately 100 microseconds; The latency for optical network units is typically 3 microseconds. To determine the maximum permissible end-to-end delay for critical management traffic, in this embodiment, the maximum permissible delay excluding fiber optic transmission is 100 microseconds, and the total delay constraint including fiber optic transmission is 200 microseconds.

[0109] Based on the typical parameters of this embodiment, the total latency is approximately 179 microseconds, which is far less than the constraint threshold of 200 microseconds, thus meeting the deterministic latency requirements for critical management operations in AI data centers.

[0110] S4. Standard Management Traffic Slow Path Bandwidth Allocation and Transmission Steps:

[0111] This step performs scheduling for regular management traffic, using a software-implemented slow path DBA module. With bandwidth utilization and scheduling fairness as core objectives, it supports complex multi-dimensional scheduling strategies to make full use of idle bandwidth resources.

[0112] Regular management traffic is sent to multiple service queues corresponding to the slow path. In this embodiment, four priority queues are configured, corresponding to high-priority regular services, medium-priority regular services, low-priority regular services, and background services, respectively. The queues adopt a first-in-first-out (FIFO) cache structure and are uniformly scheduled and managed by the central processing unit.

[0113] The slow path software DBA supports multiple scheduling algorithms and can be flexibly configured according to business needs. The default algorithm is the Weighted Fair Queueing (WFQ) algorithm, which assigns different bandwidth weights to queues of different priorities to ensure that high-priority regular services get more bandwidth resources, while avoiding starvation for low-priority services.

[0114] The scheduling formula for a weighted fair queue is as follows:

[0115]

[0116] in, For the first The transmission rate obtained by each queue Total available bandwidth For the first The weight values ​​of each queue. This represents the total number of queues. In this embodiment, the weights of the four queues are configured as 8, 4, 2, and 1, respectively, to achieve differentiated bandwidth guarantees.

[0117] In addition to weighted fair queues, the software DBA also supports algorithms such as strict priority scheduling and deficit round-robin scheduling, which can be flexibly switched through the management interface.

[0118] After bandwidth allocation is completed, the software DBA sends the scheduling results to the hardware forwarding engine, which then performs packet encapsulation and optical signal modulation, transmitting the data to the target optical network unit via a passive optical link. Regular management traffic has low latency sensitivity; millisecond-level processing latency fully meets business requirements.

[0119] S5. Idle period sleep control and power consumption optimization steps:

[0120] This step optimizes the power consumption of the physical layer of the passive optical network by using a hierarchical sleep mechanism to minimize the idle power consumption of the device without affecting the response capability of critical traffic.

[0121] The system determines to enter an idle state when both of the following conditions are met:

[0122] All scheduled traffic within the current scheduling period has been transmitted, and the buffer queue is empty;

[0123] The prediction results for the next scheduling cycle show that the interval between the next critical management traffic time window and the sleep threshold is greater than the sleep threshold. In this embodiment, the sleep threshold is set to 100 microseconds.

[0124] If only the first condition is met, but critical traffic is about to arrive in the next cycle, then keep the physical layer in an active state to avoid the additional overhead and latency risks caused by frequent wake-ups.

[0125] In this embodiment, the physical layer transceiver supports two sleep modes: shallow sleep mode and deep sleep mode. The sleep arbitrator automatically selects the sleep level based on the idle time.

[0126] When the idle time is between 100 microseconds and 1000 microseconds, it enters a shallow sleep mode: shutting down some non-core circuits, retaining the clock and fast wake-up circuit, with a wake-up delay of about 5 microseconds and a power consumption reduction of about 20%;

[0127] When the idle time exceeds 1000 microseconds, it enters deep sleep mode: shutting down most of the internal circuits and the optical module transmitter, retaining only the wake-up detection circuit, with a wake-up delay of about 20 microseconds and a power consumption reduction of about 40%.

[0128] The hierarchical sleep mechanism achieves a balance between wake-up speed and power consumption optimization. Short periods of idle time use shallow sleep to ensure fast response, while long periods of idle time use deep sleep to maximize energy saving.

[0129] In sleep mode, the sleep arbitrator continuously tracks the global clock. When the time window before the next critical traffic arrives equals the wake-up lead time, it automatically triggers the wake-up process, ensuring that the physical layer is ready before the traffic arrives. If a sudden emergency traffic arrives during sleep, a fast wake-up can be forcibly triggered through the emergency wake-up channel. The emergency wake-up delay is no more than 10 microseconds, ensuring management accessibility in emergency scenarios.

[0130] Example 2

[0131] like Figure 1 As shown, this embodiment is a system-level embodiment, which fully describes the overall architecture of the system, the hardware and software implementation of each module, the interface interaction logic, and the deployment and operation mode. This system adopts a three-level architecture of "core end - optical distribution network - terminal side". The core end is an optical line terminal deployed in the core computer room of the data center; the optical distribution network is a point-to-multipoint optical link composed of passive optical fiber and optical splitter; the terminal side is a rack-mounted optical network unit deployed at the top of each server rack.

[0132] System overall architecture overview:

[0133] The overall topology of this system is a typical tree-shaped passive optical network structure. A single optical line terminal (OLT) can support up to 64 optical network units (ONUs), corresponding to 64 server racks. By deploying multiple OLTs in a cluster, it can support the out-of-band management needs of ultra-large-scale data centers with thousands of racks. The system management plane and data plane are physically separated, and management traffic is transmitted only in the out-of-band optical links, without intersecting with the business data plane, thus possessing extremely high security and reliability.

[0134] The overall functional goal of the system is to enable remote out-of-band management of all IT equipment in the data center, with the end-to-end latency of critical management commands not exceeding 200 microseconds (including 20km fiber optic transmission), latency jitter not exceeding 10 microseconds, and 99.999% availability; at the same time, through the intelligent sleep mechanism, the power consumption of the optical line terminal physical layer is reduced by 30% to 45% compared with the all-time operation mode.

[0135] Optical line terminal hardware and software architecture:

[0136] The optical line terminal is the core control node of the system. It adopts a modular hardware architecture of "main control unit + service processing unit + optical interface unit". The software system adopts a layered architecture design, which consists of hardware driver layer, operating system layer, service function layer and management and maintenance layer from bottom to top.

[0137] Main control processing unit:

[0138] The main control processing unit is the control core of the optical line terminal. It adopts a high-performance multi-core central processing unit and a large-capacity memory configuration, and is responsible for running the operating system, managing plane protocols, slow path DBA scheduling, system configuration management, alarm reporting, online model training, and other functions. The main control unit is connected to the service processing unit and the optical interface unit through a high-speed serial bus to realize the interaction of control commands and data.

[0139] The main control unit has a built-in model storage area for storing model weight files of self-attention temporal prediction neural networks. It supports model version management and online upgrades, and can complete model iteration without restarting the device.

[0140] Traffic classifier module:

[0141] The traffic classifier module is implemented using hardware logic and integrated into the FPGA chip of the service processing unit. Its core components include four parts: packet parsing engine, TCAM rule table, classification decision engine, and traffic splitting matrix.

[0142] Message parsing engine: Supports deep parsing of Layer 2 to Layer 4 messages, can extract up to 32 message feature fields, and has a parsing capability of line speed 10Gbps;

[0143] TCAM rule table: Supports up to 1024 category rules, supports dynamic addition, deletion and modification of rules, and the rule update delay is no more than 1 millisecond;

[0144] Classification decision engine: Executes multi-rule combination matching logic, supports "AND", "OR" and "NOT" logic combinations, and the matching precision is at the single-byte level;

[0145] The traffic splitting and switching matrix is ​​based on a cross-connector architecture. It sends the matched packets to the corresponding path's buffer queue, with a switching delay of no more than 500 nanoseconds.

[0146] The traffic classifier module provides a configuration interface, allowing rules to be configured via command line or network management platform, and supports batch import and export of rules.

[0147] Self-attention temporal prediction scheduling module:

[0148] The self-attention timing prediction scheduling module adopts an implementation scheme of "FPGA accelerator card + main control software collaboration". The computationally intensive tasks such as model inference and time window generation are completed by the FPGA accelerator card, while the control tasks such as model training, data preprocessing and result verification are completed by the main control processor.

[0149] The FPGA accelerator card features a specially optimized Transformer encoder inference engine, with hardware-level pipeline optimization for self-attention computation, matrix operations, and activation function operations. It can complete parallel computation of 8 attention heads in a single cycle, offering inference performance more than 10 times higher than general-purpose processors. The accelerator card also includes a dedicated on-chip memory for caching historical traffic data and model weights, with access latency not exceeding 10 nanoseconds, ensuring high-speed execution of the inference process.

[0150] The scheduling module has a built-in time window verification unit to verify the validity of the generated scheduling table and avoid abnormal configurations such as window overlap and over-cycle. The scheduling table that fails the verification will be automatically recalculated. If the verification fails three times in a row, it will be downgraded to the static default scheduling table to ensure system reliability.

[0151] Hibernation Arbitrator Module:

[0152] The sleep arbitrator module is implemented entirely in hardware logic and integrated into the control circuit of the optical interface unit. It boasts extremely high real-time performance and reliability, unaffected by the scheduling delay of the main control processor. The core components of the module include a global clock synchronization unit, a wake-up time calculation unit, a state machine control unit, and an anomaly detection unit.

[0153] Global clock synchronization unit: Supports PTP v2 protocol, keeps synchronized with the data center master clock server, and has a synchronization accuracy better than 1 microsecond;

[0154] Wake-up time calculation unit: The built-in hardware arithmetic logic unit can calculate the wake-up time in a single cycle based on the wake-up delay parameters and safety margin;

[0155] State machine control unit: manages the working state of the physical layer transceiver, including four states: normal operation, shallow sleep, deep sleep, and wake-up. State switching strictly follows timing requirements to avoid hardware damage.

[0156] Anomaly detection unit: Real-time monitoring of physical layer status feedback. If wake-up timeout, status abnormality, or other issues occur, it automatically triggers a reset and alarm mechanism.

[0157] The sleep arbitrator module supports online parameter configuration. Parameters such as wake-up delay, safety margin, and sleep threshold can be modified through the management interface to adapt to different physical layer chips.

[0158] Fast path hardware dynamic bandwidth allocation module:

[0159] The Fast Path Hardware DBA module, also implemented using FPGA hardware, is the core guarantee for low-latency transmission of critical management traffic. The module's core components include four parts: a bandwidth request queue, a time slot allocation engine, a time slot mapping table, and a packet encapsulation engine.

[0160] Bandwidth request queue: 256 entries deep, supporting parallel caching of requests from 32 optical network units;

[0161] Time slot allocation engine: It adopts a strict priority scheduling algorithm, is implemented with pure combinational logic, and the computation latency does not exceed 1 microsecond;

[0162] Time slot mapping table: Stores the time slot information allocated to each optical network unit, adopts a double-buffered architecture, and updates without service interruption;

[0163] Message encapsulation engine: Automatically completes MAC layer frame encapsulation, cyclic redundancy check calculation, preamble insertion and other operations, with an encapsulation rate of 10Gbps line speed.

[0164] The fast path module is connected to the physical layer transceiver via a dedicated high-speed parallel interface with a data width of 64 bits, an operating clock of 156.25MHz, and an interface transmission delay of no more than 10 nanoseconds.

[0165] Slow path software dynamic bandwidth allocation module:

[0166] The slow path software DBA module runs in the user space of the main control processor and adopts a modular software design, which includes four parts: queue management submodule, scheduling algorithm submodule, bandwidth calculation submodule, and result distribution submodule.

[0167] Queue Management Submodule: Responsible for maintaining multi-priority queues, including functions such as enqueue, dequeue, queue length statistics, and packet loss statistics;

[0168] The scheduling algorithm submodule includes three built-in scheduling algorithms: weighted fair queue, strict priority, and deficit round-robin, which can be dynamically switched.

[0169] Bandwidth calculation submodule: Calculates the allocated bandwidth for each optical network unit based on parameters such as available link bandwidth, queue weight, and queue length;

[0170] The result delivery submodule sends the scheduling results to the hardware forwarding engine via the bus, with an execution cycle of 1 millisecond.

[0171] The slow path module supports a wealth of statistical functions, which can count metrics such as bandwidth utilization, average latency, and packet loss rate for each queue, providing data support for operation and maintenance optimization.

[0172] Physical layer transceiver module:

[0173] The physical layer transceiver module uses a commercially available high-performance passive optical network physical layer chip, supports a 10Gbps symmetrical transmission rate, complies with the XGS-PON standard, and features low power consumption and high integration. The core components of the module include four parts: optical module, modem circuit, clock circuit, and power management circuit.

[0174] Optical module: It adopts a pluggable SFP+ package, with a transmit optical power of 0~+5dBm, a receive sensitivity of -28dBm, and supports a transmission distance of 20km;

[0175] Modulation and demodulation circuit: realizes the mutual conversion between electrical signals and optical signals, and supports non-return-to-zero code modulation mode;

[0176] Clock circuit: Provides a high-precision reference clock and supports clock synchronization and recovery;

[0177] Power management circuit: Supports hierarchical power control, and can independently shut down the power supply of different functional modules to optimize power consumption.

[0178] The physical layer chip supports a standard management interface and can receive control commands from the sleep arbitrator to switch between sleep and wake-up states.

[0179] Global clock synchronization module:

[0180] The global clock synchronization module provides a unified time base for the entire system, forming the foundation for deterministic scheduling. The module supports PTP v2 boundary clock functionality and Synchronous Ethernet (SyncE) frequency synchronization, achieving dual synchronization of frequency and time. The module incorporates a high-precision temperature-controlled crystal oscillator, maintaining an accuracy better than 0.01ppm, ensuring time accuracy even after external clock loss.

[0181] Optical Network Unit Terminal Architecture:

[0182] The optical network unit is deployed at the top of the server rack, with each device providing 8 management Ethernet interfaces, capable of connecting to the baseboard management controllers of 8 servers. The device adopts a fanless, low-power design, with a mean time between failures (MTBF) of over 1 million hours, meeting the 24 / 7 operation requirements of data centers.

[0183] The core components of an optical network unit include:

[0184] Optical interface unit: integrates optical module and physical layer chip to realize the reception and transmission of optical signals, and supports downlink rate of 10Gbps;

[0185] Media Access Control Unit: Implements PON layer protocol processing, including functions such as frame parsing, bandwidth allocation, and queue scheduling;

[0186] Management Service Interface Unit: Provides 8 Gigabit Ethernet electrical interfaces, supports functions such as auto-negotiation and PoE power supply, and can be directly connected to the server baseboard management controller;

[0187] Local power management unit: Supports local sleep control, and can enter low power mode when there is no traffic, further reducing terminal power consumption.

[0188] The optical network unit supports remote firmware upgrades and remote configuration management, enabling functional iterations and fault repairs to be completed without on-site operation.

[0189] Management module:

[0190] The management console serves as the system's human-computer interaction and business entry point. Deployed in the data center's management area, it connects to the optical fiber terminal via an uplink. The management console comprises three core functional modules:

[0191] The baseboard management controller management platform enables basic out-of-band management functions such as server hardware status monitoring, remote power on / off, remote control, and firmware upgrades, and is the primary source of management traffic.

[0192] Data center management platform: Enables unified scheduling of network resources, capacity management, fault alarms, automated operation and maintenance, and manages traffic such as batch configuration and cluster control;

[0193] Security Management Module: Responsible for functions such as key management, security auditing, access control, and emergency security command issuance. It is the source of high-priority security management traffic.

[0194] All management devices support PTP clock synchronization, ensuring that management commands carry accurate timestamps and providing a precise data source for time series prediction.

[0195] System workflow and interaction logic:

[0196] After the system powers on, it first performs a hardware self-test, sequentially initializing the main control unit, service unit, and optical interface unit. After the self-test passes, it loads the operating system and drivers for each functional module. Then, it loads a pre-trained self-attention temporal prediction neural network model from local storage to the FPGA accelerator card, completing model parameter verification and inference engine warm-up. Finally, it establishes a link connection with the optical network unit, synchronizes the global clock, and enters normal operating mode. The entire initialization process takes no more than 3 minutes.

[0197] In daily operation and maintenance, routine management traffic accounts for over 90%. Its workflow is as follows: the management terminal issues routine management commands → the optical line terminal traffic classifier identifies it as routine traffic → it is sent to the slow path queue → the software DBA periodically schedules and allocates bandwidth → it is transmitted to the optical network unit via the optical link → it is delivered to the managed device. This process prioritizes bandwidth utilization and fairness, and operates stably and reliably.

[0198] When an emergency occurs, such as a cluster failure or security vulnerability, requiring the issuance of critical management commands, the workflow is as follows: the management terminal issues an emergency command, with the message carrying a highest priority marker → the traffic classifier quickly identifies it as critical traffic and sends it to the fast path → the self-attention scheduling module generates a dedicated time window → the dormant arbitrator wakes up the physical layer in advance → the fast path hardware DBA allocates bandwidth → the command is transmitted at high speed to the target device → emergency operations are executed. The entire process, from the message entering the optical line terminal to its issuance by the optical network unit, has a processing latency of no more than 80 microseconds, achieving extremely fast response.

[0199] During the daily off-peak hours in the early morning, the system automatically performs online model fine-tuning: it extracts traffic data from the past 24 hours as training samples, completes model fine-tuning training in the main control processor, verifies the model's prediction accuracy after training, and updates the model weights in the FPGA if the accuracy meets the standard; otherwise, it retains the original model to ensure scheduling stability. The entire update process is uninterrupted and has no impact on normal traffic management.

[0200] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application.

Claims

1. A deterministic guarantee method for out-of-band management of passive optical networks, characterized in that, Includes the following steps: S1. The optical line terminal receives management traffic from the management terminal and identifies the management traffic according to preset classification rules through a traffic classifier, dividing it into critical management traffic and regular management traffic. S2. If the traffic is identified as critical management traffic, the pre-trained self-attention timing prediction neural network model is invoked to predict the arrival time of the critical management traffic based on historical traffic data and dynamically generate the corresponding transmission time window; the sleep arbitrator calculates the wake-up time of the physical layer transceiver based on the global clock and the transmission time window, and wakes up the physical layer transceiver in advance before the transmission time window arrives. S3. Allocate transmission bandwidth for critical management traffic through the fast path hardware dynamic bandwidth allocation module, and transmit the critical management traffic to the target optical network unit via the passive optical link; S4. If identified as regular management traffic, the transmission bandwidth is allocated to it through the slow path software dynamic bandwidth allocation module, and the regular management traffic is transmitted to the target optical network unit via the passive optical link; after the traffic transmission is completed, the physical layer transceiver enters a sleep state during idle periods.

2. The deterministic guarantee method for out-of-band management of passive optical networks according to claim 1, characterized in that, The preset classification rules are based on traffic type, source address, destination address, or virtual LAN label; among them, emergency stop command, device emergency recovery command, and security key rotation traffic are marked as critical management traffic, while heartbeat message, operation log collection, and configuration synchronization traffic are marked as regular management traffic.

3. The deterministic guarantee method for out-of-band management of passive optical networks according to claim 1, characterized in that, The input of the self-attention time-series prediction neural network model is multi-dimensional management traffic time-series data within a historical period, and the output is the dynamic time window arrangement corresponding to each category of traffic; the time window arrangement of key management traffic has a higher priority than that of regular management traffic, and the time windows of different categories do not overlap.

4. The deterministic guarantee method for out-of-band management of passive optical networks according to claim 1, characterized in that, In step S2, when the hibernation arbiter calculates the wake-up time, the wake-up advance is not less than the sum of the physical layer transceiver's wake-up delay and the preset safety margin, ensuring that the physical layer transceiver is ready before the arrival of critical management traffic.

5. The method for deterministic guarantee of out-of-band management of passive optical networks according to claim 1, characterized in that, The fast path hardware dynamic bandwidth allocation module implements bandwidth allocation based on hardware logic, with a processing latency in the microsecond range; the slow path software dynamic bandwidth allocation module implements bandwidth allocation based on software algorithms, supporting multi-dimensional scheduling strategies.

6. A deterministic guarantee system for out-of-band management of passive optical networks, characterized in that, It includes an optical line terminal and several optical network units deployed in each rack. The optical line terminal has a built-in traffic classifier, a self-attention timing prediction and scheduling module, a sleep arbitrator, a fast path hardware dynamic bandwidth allocation module, a slow path software dynamic bandwidth allocation module, and a physical layer transceiver. The traffic classifier is used to receive management traffic sent by the management terminal and divide it into critical management traffic and regular management traffic according to preset rules; The self-attention timing prediction scheduling module is used to call a pre-trained self-attention timing prediction neural network to predict the arrival time of key management traffic based on historical traffic data and dynamically generate a transmission scheduling schedule for various types of traffic. The sleep arbitrator is used to control the sleep and wake-up switching of the physical layer transceiver according to the scheduling information of the transmission scheduling schedule and the global clock; The fast path hardware dynamic bandwidth allocation module is used to perform low-latency bandwidth allocation for critical management traffic. The slow path software dynamic bandwidth allocation module is used to perform bandwidth allocation scheduling for regular management traffic. The physical layer transceiver is used to realize the transmission and reception of optical signals between the optical line terminal and the optical network unit, and supports the switching between sleep and wake-up states.

7. The deterministic guarantee system for out-of-band management of passive optical networks according to claim 6, characterized in that, The traffic classifier supports classification rules based on traffic type, address information, or virtual LAN tags, and allows for customization of the identification conditions between critical management traffic and regular management traffic.

8. The deterministic guarantee system for out-of-band management of passive optical networks according to claim 6, characterized in that, In the transmission scheduling timetable generated by the self-attention timing prediction scheduling module, the time window of critical management traffic has a higher priority than that of regular management traffic. The time windows do not overlap within the scheduling period, and the window parameters can be dynamically adjusted according to the real-time traffic status.

9. The deterministic guarantee system for out-of-band management of passive optical networks according to claim 6, characterized in that, The hibernation arbiter is synchronized with the global clock and can calculate the wake-up advance time based on the wake-up delay parameters of the physical layer transceiver and the preset safety margin, so as to complete the wake-up of the physical layer transceiver before the time window of critical management traffic arrives.

10. The deterministic guarantee system for out-of-band management of passive optical networks according to claim 6, characterized in that, The fast path hardware dynamic bandwidth allocation module is a hardware logic circuit with a processing latency in the microsecond range; the slow path software dynamic bandwidth allocation module runs on the main control processor and supports complex scheduling algorithms.

Images