Optical fiber loop dynamic addressing method, device and electronic equipment

Abstract

This application provides a method, apparatus, and electronic device for dynamic addressing of fiber optic loops. The method includes: responding to a data read request, querying a maintained global dynamic mapping table to obtain the most recent node observation information of the target data packet corresponding to the data read request at the optical monitoring node; determining the expected arrival time window of the target data packet to the optical capture module based on a preset dynamic addressing algorithm and the most recent node observation information; activating the optical capture module within the time window to match and extract the target data packet from the fiber optic loop. This application achieves extremely low-latency data access and efficient loop bandwidth utilization by continuously storing data in the fiber optic loop and combining it with intelligent addressing.

Description

Technical Field

[0001] This disclosure relates to the field of optical storage and data addressing technology, and in particular to a method, apparatus and electronic device for dynamic addressing of optical fiber loops. Background Technology

[0002] With the rapid development of artificial intelligence technology, the efficient storage and rapid access to massive amounts of AI training data, inference intermediate results, and model parameters have become key bottlenecks restricting system performance. Traditional disk-based or solid-state drive (SSD) storage systems suffer from high read / write latency and significant I / O bottlenecks, while memory-based storage solutions, although offering lower latency, are expensive and volatile. For example, in distributed AI training, parameter servers need to frequently read and update model parameters. If NVMe SSDs are used, the average latency is approximately 50 microseconds, which is insufficient to meet the synchronization requirements of large models with hundreds of billions of parameters.

[0003] In recent years, optical storage technology has attracted widespread attention due to its advantages such as high bandwidth, low power consumption, and resistance to electromagnetic interference. An optical fiber loop, as a closed optical signal transmission channel, allows optical signals to propagate continuously within it. If data is injected into the optical fiber loop in the form of optical data packets and continuously circulated, theoretically, "flowing storage" of data can be achieved—data is no longer fixed to a specific physical address but moves continuously within the loop. This storage method eliminates the mechanical seek time of traditional storage media and is expected to significantly reduce data access latency.

[0004] However, the continuous movement of optical data packets within the loop presents new technical challenges: how to quickly and accurately locate and extract specific data packets? Since the packet position changes over time, traditional fixed-address-based addressing methods are no longer applicable. Existing optical buffering schemes based on fiber delay lines (FDLs) (such as US20050271391A1) typically only support first-in-first-out (FIFO) access, cannot achieve random addressing, and the buffer depth is limited by the delay line length, making expansion difficult. Another approach is circular storage of optical packets based on fixed timestamps (such as US20190341995A1). This scheme records a timestamp during data injection, calculates the packet position based on the difference between the current time and the injection time during retrieval, and then activates the capture device at the predicted time. However, this scheme has the following serious drawbacks: it does not consider the minute changes in the speed of light caused by gain saturation of the optical amplifier in the loop, and the accumulated time error can reach several microseconds after multiple cycles; it does not consider the collision retransmission caused by the simultaneous injection of multiple data packets, which will cause the actual position of the data packets to deviate from the theoretically calculated value; and it does not consider the time jitter caused by amplifier spontaneous emission (ASE) noise, resulting in an acquisition failure rate of over 30% in actual measurements under high load (injection rate > 5000 pps). Therefore, an intelligent addressing method is needed that can dynamically sense changes in loop load, adaptively predict the position of optical data packets, and precisely control the acquisition timing. Summary of the Invention

[0005] This disclosure provides a method, apparatus, and electronic device for dynamic addressing of fiber optic loops. By continuously storing data in the fiber optic loop and cooperating with intelligent addressing, it achieves extremely low latency data access and efficient loop bandwidth utilization.

[0006] In a first aspect, a dynamic addressing method for fiber optic loops is applied to a controller of a dynamic addressing system for fiber optic loops. The dynamic addressing system further includes a fiber optic loop, a data injection module, an optical monitoring node, and an optical capture module. At least one optical monitoring node, at least one data injection module, and at least one optical capture module are disposed on the fiber optic loop. The controller is connected to the optical monitoring node, the data injection module, and the optical capture module. The method includes: In response to a data read request, the maintained global dynamic mapping table is queried to obtain the latest observation information of the target data packet corresponding to the data read request at the optical monitoring node. Based on a preset dynamic addressing algorithm, the expected arrival time window of the target data packet to the optical capture module is determined according to the most recent observation information of the node; The optical capture module is activated within the time window to match and extract the target data packet from the optical fiber loop.

[0007] The above-described fiber optic loop dynamic addressing method achieves extremely low-latency data access and efficient loop bandwidth utilization by continuously storing data in the fiber optic loop and combining it with intelligent addressing. Since optical signals propagate at near-light speed in the fiber, mechanical seek time is eliminated, and the average data access latency can be as low as sub-microseconds.

[0008] Optionally, before the step of querying the maintained global dynamic mapping table and obtaining the most recent observation information of the node of the target data packet corresponding to the data read request in response to the data read request, it further includes: Obtain the physical parameters of the optical fiber loop and the data packet injection parameters of the data injection module; Based on the physical parameters and the data packet injection parameters, a state-space model is established to describe the position distribution and time relationship of optical data packets in the optical fiber loop. The state-space model is used to determine the position information of the target data packet at any time.

[0009] Optionally, before the step of querying the maintained global dynamic mapping table and obtaining the most recent observation information of the node of the target data packet corresponding to the data read request in response to the data read request, it further includes: The optical monitoring node continuously detects optical data packets passing through its location to obtain monitoring information of the optical data packets; The controller updates the corresponding entries for optical data packets in the global dynamic mapping table based on the monitoring information; When the status information of any data packet meets the preset deletion status, the controller deletes the corresponding entry of the arbitrary data packet from the global dynamic mapping table.

[0010] Optionally, the step, based on a preset dynamic addressing algorithm, determines the expected arrival time window of the target data packet to the optical capture module according to the most recent observation information of the node, including: The controlling agent observes the current state information based on the most recent observation information of the node; Based on the dynamic addressing algorithm and the current state information, the expected arrival time window for the target data packet to reach the optical capture module is determined.

[0011] Optionally, the step of activating the optical capture module within the time window and matching and extracting the target data packet from the fiber optic loop includes: An activation signal is sent to the light capture module within the time window; The light capture module is controlled to continuously monitor passing light signals within the time window; If the target identifier of the target data packet is matched in the optical signal, the optical capture module is controlled to obtain the target data packet from the optical fiber loop and update the global dynamic mapping table.

[0012] Optionally, after activating the optical capture module within the time window and matching and extracting the target data packet from the fiber optic loop, the step further includes: If the target data packet is not captured within the time window, the global dynamic mapping table is queried again and the dynamic addressing algorithm is called again for prediction until the timeout or the maximum number of retries is reached. Based on the capture result information of the target data by the light capture module, the reward information and the next state information are fed back to the dynamic addressing algorithm.

[0013] Optionally, optical carriers of different wavelengths in the optical fiber loop independently carry data packet streams, and the method further includes: When multiple read requests access different data packets simultaneously, a time-slice round-robin or multi-threaded parallel prediction strategy is used to adjust the time window of each capture device. When the time windows of different target data packets overlap, a preset priority preemption strategy is used to execute them one at a time.

[0014] Optionally, the dynamic addressing algorithm is trained using a deep Q-network or near-end strategy optimization; the controller is implemented using a field-programmable gate array or an embedded processor; the global dynamic mapping table module is implemented using a static random access memory; and the light capture module is an optical switch based on a semiconductor optical amplifier or a micro-ring resonator.

[0015] Secondly, a dynamic addressing device for an optical fiber loop is provided. The device includes a controller, an optical fiber loop, a data injection module, an optical monitoring node, and an optical capture module. At least one optical monitoring node, at least one data injection module, and at least one optical capture module are disposed on the optical fiber loop. The controller is connected to the optical monitoring node, the data injection module, and the optical capture module. The controller includes: The data acquisition module is used to respond to a data read request by querying the maintained global dynamic mapping table and obtaining the most recent observation information of the node of the target data packet corresponding to the data read request. The data analysis module is used to determine the expected arrival time window of the target data packet to the light acquisition module based on the preset dynamic addressing algorithm and the most recent observation information of the node. A data extraction module is used to activate the optical capture module within the time window to match and extract the target data packet from the optical fiber loop.

[0016] Thirdly, an electronic device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the fiber optic loop dynamic addressing method provided in the first aspect above. Attached Figure Description

[0017] The accompanying drawings used in the description of the embodiments of this disclosure are briefly introduced below: Figure 1 A schematic diagram of a fiber optic loop dynamic addressing system provided in some embodiments of this application is shown; Figure 2 A schematic flowchart of a fiber optic loop dynamic addressing method provided in some embodiments of this application is shown; Figure 3 This paper illustrates a schematic diagram of the optical data packet location distribution in a state-space model provided in some embodiments of this application; Figure 4 The diagram illustrates the data structure of the global dynamic mapping table provided in some embodiments of this application; Figure 5 The present application illustrates a framework for a dynamic addressing algorithm based on reinforcement learning, provided in some embodiments thereof. Detailed Implementation

[0018] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the specific implementation methods of this disclosure will be described below with reference to the accompanying drawings. The accompanying drawings described below are merely some embodiments of this disclosure. For those skilled in the art, other drawings or embodiments can be obtained based on these drawings or embodiments without creative effort. Adjustments and improvements made without departing from the concept of this disclosure are all within the protection scope of this disclosure.

[0019] To keep the drawings simple, each figure only schematically shows the parts relevant to the embodiment, and they do not represent the actual structure of the product. In addition, for the sake of simplicity and ease of understanding, only some structures or parts are schematically shown, and there may be more or fewer similar structures or parts in reality.

[0020] like Figure 1 As shown, the fiber optic loop dynamic addressing system includes: a fiber optic loop: a closed fiber optic channel with a length of L meters. The propagation speed of the optical signal in the fiber is v (approximately 2 × 10⁸ m / s). Therefore, the propagation delay of the optical data packet around the loop is... Seconds. Optical data packets continuously circulate in a single direction (e.g., clockwise) within the loop. Data injection module: Located at injection position xinj in the fiber optic loop, it encapsulates the AI ​​data to be stored into optical data packets and injects them into the loop. Each optical data packet contains: a unique identifier ID, an injection timestamp t0, a data payload (AI data fragment), and a loop count c (recording the number of loops). Optical monitoring node: Located at the monitoring position. It continuously detects passing optical data packets, reads their identifiers and timestamps, and reports the information to the controller. Optical capture module: located at the capture position. Activated within a specific time window by controller commands, it matches and extracts target optical data packets from the fiber optic loop. The controller is the core computing and control unit of the system, maintaining a global dynamic mapping table, executing dynamic addressing algorithms, and coordinating the timing of injection, monitoring, and capture of various devices. AI data includes model parameters from distributed AI training, order book data from high-frequency financial trading, or feature frames from real-time video analysis.

[0021] A fiber optic loop storage system is constructed and data is injected. The fiber optic loop is initialized, and the position parameters of the optical monitoring node, data injection module, and optical capture module are set. When AI data needs to be stored, the controller divides the data into several fragments, encapsulates each fragment into an optical data packet, assigns a unique identifier (ID), records the current timestamp (t0), and sends the data packet into the fiber optic loop through the data injection module. Each optical data packet also contains a priority identifier (3 bits, supporting 8 priorities) and a time-to-live (TTL) field (16 bits, in milliseconds) to support differentiated services and automatic data expiration management. After injection, the optical data packet continuously circulates along the fiber optic loop at a speed (v). When the number of loops of the optical data packet exceeds the maximum number of loops corresponding to the TTL, the controller instructs the optical capture module to remove it or allow it to decay naturally.

[0022] Figure 2 A schematic flowchart of a fiber optic loop dynamic addressing method provided in some embodiments of this application is shown. The fiber optic loop dynamic addressing method includes at least the following steps: S110: In response to a data read request, query the maintained global dynamic mapping table to obtain the latest observation information of the target data packet corresponding to the data read request at the optical monitoring node. S120: Based on a preset dynamic addressing algorithm, determine the expected arrival time window of the target data packet to the optical capture module according to the most recent observation information of the node; S130: Activate the optical capture module within the time window to match and extract the target data packet from the optical fiber loop.

[0023] Specifically, when the controller receives a data read request, it queries the maintained global dynamic mapping table in response to the request. The global dynamic mapping table includes, but is not limited to, the packet identifier ID and the time of the most recent passage through the optical monitoring node. Cycle counting Logical position Inject timestamp And status flags (such as active, extracted, expired).

[0024] The controller retrieves the most recent observation information of the target data packet corresponding to the data read request at the optical monitoring node from the global dynamic mapping table, and then determines the expected arrival time window of the target data packet to the optical acquisition module by combining it with a preset dynamic addressing algorithm. Within the time window, the controller activates the optical acquisition module to match and extract the target data packet from the fiber optic loop.

[0025] This embodiment achieves extremely low-latency data access and efficient loop bandwidth utilization by continuously storing data in the fiber optic loop and using intelligent addressing. Optical signals propagate at near-light speed in the fiber, eliminating mechanical seek time and reducing average data access latency to sub-microsecond levels. Through reinforcement learning-based adaptive prediction of the time window, the acquisition success rate can reach over 98% under high-load dynamic environments. The intelligent algorithm outputs a compact window, avoiding prolonged activation of the acquisition device and reducing loop bandwidth waste. It supports wavelength division multiplexing extension and parallel processing of multiple read requests, making it suitable for large-scale data centers and edge computing scenarios.

[0026] In some embodiments, S110: In response to a data read request, querying the maintained global dynamic mapping table to obtain the target data packet corresponding to the data read request before the most recent observation information of the optical monitoring node, further includes: S210: Obtain the physical parameters of the optical fiber loop and the data packet injection parameters of the data injection module; S220: Based on the physical parameters and the data packet injection parameters, establish a state-space model describing the position distribution and time relationship of optical data packets in the optical fiber loop. The state-space model is used to determine the position information of the target data packet at any time. S230: The optical monitoring node continuously detects optical data packets passing through its location and obtains monitoring information of the optical data packets; S240: The controller updates the corresponding entry for the optical data packet in the global dynamic mapping table based on the monitoring information; S250: When the status information of any data packet meets the preset deletion status, the controller deletes the corresponding entry of the arbitrary data packet from the global dynamic mapping table.

[0027] Specifically, a state-space model for optical data packets is established. This model is used to determine the location information of the target data packet at any given time. The physical parameters of the optical fiber loop include, but are not limited to, the loop length and the location of the selected reference point along the loop. The data packet injection parameters of the data injection module include, but are not limited to, the data injection location and the data injection speed. By combining the physical parameters and the data packet injection parameters, a state-space model describing the relationship between the location distribution of optical data packets in the optical fiber loop and time can be established.

[0028] Define the position distribution of all optical data packets on the fiber loop at time t as a state vector: ; Where N = N(t) is the total number of data packets in the loop at time t (including all data packets that have not yet been extracted after injection). This represents the arc length distance between the i-th data packet and a selected reference point on the loop (such as the location of an optical monitoring node). For example... Figure 3As shown in the diagram, this schematic diagram uses a reference point (the location of the light monitoring node) as an example. Starting from the origin, the coordinates of different data packets are marked along the loop in a clockwise direction. and injection location j and capture position This intuitively reflects the dynamic distribution of data packets within the loop. Since the optical data packets move unidirectionally along the loop at a constant speed v, the deterministic law governing their position evolution over time is as follows: ; To describe the time jitter caused by the spontaneous emission noise of the optical amplifier in a real system, a Gaussian perturbation term can be introduced into the deterministic model: ; in B represents the optical filter bandwidth, and kASE represents the amplifier noise figure. This extended model serves as a source of stochasticity in the reinforcement learning environment, improving the generalization ability of the addressing algorithm.

[0029] The data injection process is modeled as occurring at the injection node location. At the injection point, new data packets are randomly added according to the injection rate λ(t). Let λ(t) be a time-varying Poisson process, then within a small time interval... The number of newly injected packets follows a Poisson(λ(t)dt) pattern. The initial position of each new packet... The initial timestamp t0 = t, and the initial cycle count c = 0. Data packets may be removed due to exceeding the maximum lifetime (TTL) or being successfully retrieved; the removal process is modeled as the deletion of the corresponding element in the state vector.

[0030] The state-space model further includes a packet cycle count field ci(t), which is updated as follows: ; Used to record the number of loops of a data packet, assisting the addressing algorithm in determining whether a data packet is about to expire.

[0031] The optical monitoring node is located at It continuously detects passing optical data packets. Whenever a data packet i passes through the optical monitoring node (satisfying...) The optical monitoring node reads its ID, current timestamp (tobs), and cycle count (c), and reports this information to the controller. The controller then updates the global dynamic mapping table accordingly. For example... Figure 4 As shown, each entry includes: packet identifier ID, and the time of the most recent passage through the optical monitoring node. Cycle counting Logical position Inject timestamp And status flags (such as active, extracted, expired). When an optical data packet is extracted or its lifecycle expires, the controller removes the corresponding entry from the global dynamic mapping table.

[0032] The global dynamic mapping table uses a hash table structure, with the key being the packet identifier and the value being a structure containing the aforementioned fields. The controller periodically scans the global dynamic mapping table; if the status information of any packet matches a preset deletion condition (e.g., timeout due to inaccessibility or exceeding the maximum loop count), the corresponding entry is deleted to free up storage resources. To support high concurrency, the mapping table can employ a multi-bucket hash lock mechanism to reduce read / write conflicts.

[0033] In some embodiments, step S120: determining the expected arrival time window of the target data packet to the optical capture module based on a preset dynamic addressing algorithm and the most recent observation information of the node, including: S310: The control agent observes the current state information based on the most recent observation information of the node; S320: Based on the dynamic addressing algorithm and the current state information, determine the expected arrival time window of the target data packet to the optical capture module.

[0034] Step S130: After activating the optical capture module within the time window, matching and extracting the target data packet from the fiber optic loop, the process further includes: S410: Based on the capture result information of the target data by the light capture module, feedback reward information and next state information to the dynamic addressing algorithm.

[0035] Specifically, when a data read request for a target data packet is received, the controller first queries the global dynamic mapping table to obtain the most recent observation information of the data packet. The current moment can be calculated using a deterministic model. Location: L; However, since loop load conditions can affect the actual propagation of data packets (e.g., optical amplifier gain saturation, data packet collision retransmission, etc.), this application introduces an intelligent prediction mechanism. The addressing process is modeled as a Markov decision process (MDP), and solved using the near-end policy optimization (PPO) algorithm. Figure 5 As shown in the diagram, the dynamic addressing algorithm framework based on reinforcement learning in this application illustrates the interaction process between the agent and the fiber optic loop environment. The agent observes the current state information. (Including mapping table information, loop load, last observation time difference, etc.), and output actions through the policy network. This refers to the predicted expected arrival time window. The environment performs a packet capture attempt based on this action and returns reward information. and next state information During training, the agent collects trajectories and updates network parameters to maximize cumulative rewards. Figure 4 The document also includes annotations for the Experience Replay buffer and the PPO pruning loss calculation module.

[0036] The MDP quadruple design is as follows: State space: The state information observed by the agent at decision time t. Represented as: ; Where N(t) is the total number of data packets in the current loop, λ(t) is the injection rate, and history represents the statistical characteristics of historical addressing successes / failures (including the mean and variance of the time errors of the last 10 addressings). To avoid network training instability caused by differences in the dimensions of various features, the continuous state components are normalized online: ,in For the running mean and variance, .

[0037] Action space: The agent's action is to predict the arrival of the target data packet at the light acquisition module. time window , can be represented as ,in The minimum response time is determined by the light capture module (e.g., SOA-based optical switches). The action space is constrained within a reasonable range: It is determined by the minimum propagation time and the maximum survival time.

[0038] Reward function: Let the actual arrival time be... ,like If the capture is successful, a reward message will be received. Otherwise, you will receive a negative reward. α is the penalty coefficient (usually taken as 0.1). For further optimization, an additional time window width penalty term is introduced: , where β is the weighting coefficient (taken as 0.01). The default window width is half-width (e.g., 2 microseconds). This penalty encourages the agent to output a compact window, reducing the power consumption of the light capture module and decreasing the probability of false captures caused by excessively long windows.

[0039] State transition function: After the agent performs an action, the environment is updated according to the following rules: time advances by Δt; the positions of all existing data packets are updated according to... Update; newly injected data packets are randomly added according to the Poisson process; extracted or expired data packets are removed; the optical monitoring node triggers an observation update mapping table when a data packet passes through.

[0040] Training employs the PPO algorithm, where the agent interacts with the environment to collect trajectories. The generalized advantage estimation (GAE) is used to calculate the advantage function, and the policy and value networks are updated by pruning the objective function until the addressing success rate converges. During training, an experience replay buffer and prioritized experience replay are used to improve sample efficiency. Both the policy and value networks use a three-layer fully connected structure with 128 neurons per layer, using ReLU activation function, and the output layer uses Tanh-adapted action space.

[0041] In some embodiments, a deep Q-network (DQN) can be used instead of a PPO. In this case, the action space needs to be discretized: the center moment of the prediction time window is used. If the half-width δ is discretized into M and K discrete values ​​respectively, the total number of actions is M×K. DQN uses empirical replay and target network stabilization training, making it suitable for scenarios with a small action space (e.g., M=50, K=5). Compared to PPO, DQN is simpler to implement, but discretization results in some loss of accuracy. Those skilled in the art can choose a suitable algorithm based on the actual system resources (number of FPGA logic gates).

[0042] Furthermore, the fiber optic loop length L is an adjustable parameter in the system design. For short-distance scenarios (such as within data center racks), The loop latency is short, and data packet updates are frequent, requiring the controller to have a higher processing speed (nanosecond-level response). In this case, the state update interval of reinforcement learning should be reduced to 0.1 μs, and the time error penalty coefficient α in the reward function should be increased accordingly (e.g., to 0.5) to encourage more accurate predictions. For long-distance scenarios (such as metropolitan area edge nodes), The loop time is extended, the number of data packet cycles is reduced, and the requirement for prediction accuracy is relatively lower. However, the pulse broadening caused by fiber dispersion needs to be considered. At this time, the minimum interval of optical data packets should be increased, and the load characteristic N(t) in the state space can reduce the update frequency.

[0043] The reinforcement learning-based dynamic addressing algorithm is trained using a deep Q-network or near-end policy optimization. During training, the agent learns the optimal time window prediction strategy by interacting with the environment. The state transitions of the environment are jointly determined by the fiber optic loop deterministic motion model and the random injection process. An experience replay buffer and prioritized experience replay are employed during training to improve sample efficiency.

[0044] In some embodiments, the optical capture module does not employ an SOA electro-optic switch, but instead uses a microring resonator (MRR) to achieve all-optical capture. The MRR tunes the resonant wavelength through thermo-optic effects or carrier injection. When the resonant wavelength matches the optical carrier wavelength of the target data packet, the optical signal is coupled to the download port. This approach requires no electrically driven delay, achieving sub-nanosecond switching times, but necessitates an additional wavelength tuning control loop. The controller sends a tuning voltage signal within a prediction time window to align the MRR resonant peak with the target wavelength, reverting to normal operation after the window ends. Due to the relatively slow tuning speed of the MRR (approximately 1 microsecond), this approach is more suitable for long-distance loops with large data packet intervals.

[0045] To support quality of service assurance, in another embodiment, the global dynamic mapping table is extended to a multi-queue structure: high-priority data packet mapping entries are stored in a high-bandwidth, low-latency SRAM area, while low-priority entries can be stored in slower DRAM. When processing read requests, the controller prioritizes querying the high-priority queues. Furthermore, each priority queue undergoes independent aging scans, with longer timeouts (e.g., 10 seconds) set for high-priority entries and shorter timeouts (e.g., 1 second) for low-priority entries, ensuring that important data is less likely to be evicted.

[0046] In some embodiments, step S130: activating the optical capture module within the time window, matching and extracting the target data packet from the fiber optic loop, includes: S510: Send an activation signal to the light capture module within the time window; S520: Control the light capture module to continuously monitor the passing light signal within the time window; S530: If the target identifier of the target data packet is matched in the optical signal, the optical capture module is controlled to obtain the target data packet from the optical fiber loop and update the global dynamic mapping table.

[0047] Step S130: After activating the optical capture module within the time window, matching and extracting the target data packet from the fiber optic loop, the process further includes: S610: If the target data packet is not captured within the time window, the global dynamic mapping table is queried again and the dynamic addressing algorithm is called again for prediction until the timeout or the maximum number of retries is reached.

[0048] Specifically, after obtaining the expected arrival time window, the controller determines the activation time based on the hardware switching delay of the light capture module (such as an optical switch based on a semiconductor optical amplifier, whose activation time includes an electrical drive delay). SOA carrier establishment time ,generally The activation signal is sent in advance, and the actual sending time is... During this window, the optical capture module continuously monitors the passing optical signal and parses the identifier of each data packet. When a target identifier of a target data packet is matched, the optical capture module extracts the target data packet from the loop (using optical coupler splitting or optical switch routing), sends it to the receive buffer, and returns an acknowledgment message to the controller. The controller then updates the global dynamic mapping table to mark the data packet as extracted. To ensure time synchronization accuracy, the controller and the optical monitoring node use the IEEE 1588 precision time protocol to achieve sub-microsecond synchronization and periodically send synchronization correction frames.

[0049] If the target data packet is not captured within the window period, the controller can re-query the global dynamic mapping table (waiting for possible new observations from the optical monitoring node) and call the dynamic addressing algorithm again for a new round of prediction until a timeout occurs or the maximum number of retries (e.g., 3 times) is reached. If the retries still fail, a "data unavailable" error is returned to the requester.

[0050] Multiple read requests and conflict handling: When multiple data read requests simultaneously target the same capture location, the controller maintains a request queue and schedules them according to the order of their time windows. If the time windows of different target data packets overlap, a priority preemption strategy is adopted: higher priority requests are executed first, and lower priority requests have their windows recalculated. For wavelength division multiplexing (WDM) systems, data packet streams of different wavelengths can be processed in parallel. The controller maintains multiple parallel mapping tables and performs wavelength-level load balancing—when the load on a certain wavelength exceeds a threshold, newly injected data packets are preferentially allocated to idle wavelengths, and read requests also preferentially query wavelengths with lower loads.

[0051] Consider a typical distributed AI training cluster: the parameter server uses the fiber optic loop storage system described in this application as a cache layer for model parameters. Training nodes (workers) frequently read and update model parameters. The specific workflow is as follows: Initialization: The parameter server fragments the initial model parameters into multiple optical data packets. Each data packet carries a parameter block ID (as an ID) and a timestamp, and injects it into the fiber optic loop. The loop length is 10km, and the initial injection rate is 500 data packets / second.

[0052] Training node reads parameters: When a training node requests a parameter block, the parameter server controller receives the read request and queries the mapping table to obtain the time when the parameter block last passed through the monitoring node. And location. The controller uses the current loop load (total number of packets, injection rate) and mapping table information to form the state. The trained PPO agent is invoked to output the prediction time window. The controller sends an activation signal to the light capture module before the window starts. After successfully extracting the light data packet, it converts the parameter block data into an electrical signal and returns it to the training node. The entire reading process takes an average of 1.2 microseconds, which is much shorter than that of an SSD-based parameter server (approximately 50 microseconds).

[0053] Training node parameter update: After completing gradient calculation, the training node sends the updated parameter block back to the parameter server. The controller encapsulates the updated data into a new optical data packet and injects it into the loop, then deletes the original data packet (or allows it to expire naturally). To ensure consistency, the controller uses a write-marking technique: it records the version number of the parameter block in the mapping table, and must match the latest version when reading.

[0054] Load Adaptation: When the training cluster load increases (injection rate exceeds 5000 pps), the controller detects an increase in the number of data packets N(t) in the loop. The reinforcement learning agent automatically adjusts the half-width of the time window δ from 0.5 microseconds to 1.0 microseconds to ensure that the capture success rate remains above 98%. At the same time, the controller initiates wavelength-level load balancing, migrating some data packets to backup wavelengths to avoid congestion on a single wavelength.

[0055] To facilitate understanding, a specific example parameter is given: fiber optic loop length L = 10 km, refractive index n = 1.47, and speed of light. ,but The monitoring node is located at The capture device is located Injection location Packet injection rate (Steady state), each data packet is 500 bytes in size. Training is performed using the PPO algorithm with a learning rate of 3×10⁻⁶. 4. Discount factor γ=0.99, GAE parameter λGAE=0.95, K_epochs=10, minibatch size=64. 5000 episodes were trained in a simulated environment (each episode simulating 10 seconds of system time). The first 1000 episodes used epsilon-greedy exploration, followed by gradual annealing. The final addressing success rate reached 98.7%, with an average time error of [missing value]. The window half-width δ converged to 0.5 μs (initial value 2 μs). Compared with existing methods based on fixed timestamp estimation, the capture success rate was improved from 72% to 98.5% under high load (injection rate > 5000pps), and loop bandwidth waste was reduced by 85%.

[0056] To more clearly illustrate the specific implementation logic of the above methods, the following description is further illustrated with pseudocode. Algorithm 1 presents the main flow of fiber optic loop storage and dynamic addressing, including event loop, data injection, monitoring triggering, and read request processing; Algorithm 2 describes the update function of the state space model, including packet position update, loop count update, new packet injection, and expired packet removal; Algorithm 3 presents the PPO-based addressing agent training process, including the updates of the policy network, value network, and advantage estimation.

[0057] Algorithm 1: Main Process of Fiber Loop Storage and Dynamic Addressing Input: Fiber optic loop parameters L, v, x_mon, x_cap, x_inj WriteReqs is the data write request sequence, and ReadReqs is the data read request sequence. Output: The data read 1. Initialization: Create an empty mapping table (Map), initialize the reinforcement learning agent, and start the fiber optic loop environment. 2. for each event in time order: 3. If the event is WriteReq(data): 4. ID = generateUniqueID() 5.t0 = currentTime() 6.packet = (ID, t0, data, cycle=0) 7. The injection device injects the packet into the loop at x_inj. 8.Map.add(ID, t_last=t0, s_last=x_inj, c_last=0, status=active) 9.if event is MonitorTrigger at x_mon: 10.packet = Data packets read by the monitoring node 11.Map.update(ID, t_last=currentTime(), s_last=x_mon, c_last+=1) 12. if event is ReadReq(ID): 13. If ID is not in Map: 14. Return "Data does not exist or has expired" 15.state = getState(Map[ID], currentTime(), N(t), λ(t), L, v, x_cap) 16. retry_count = 0 17. while retry_count <max_retries: 18.(t_arrival, delta) = Agent.predict(state) 19.# Consider hardware switch delay and activate in advance. 20.activate_capturer(x_cap, t_arrival - delta - t_driver - t_SOA, t_arrival + delta) 21.result = wait_for_capture(timeout=2 delta) 22. if result.success: 23.data = result.packet.data 24. Map.remove(ID) 25. Return data 26.else: 27. sleep(min(0.1) T_cycle, 10e-6))# Waiting for the mapping table to be updated 28.state = getState(updated_info) 29. retry_count += 1 30. Return "Addressing failed" Algorithm 2: State-space model update function 1.Function updateState(S, N, λ, L, v, Δt, injection_events, removal_events): 2. For each packet i in S: 3. s_i = (s_i + v Δt) mod L 4.c_i = c_i + floor((s_i_old + v Δt) / L)# Update loop count 5.# Inject new data packets 6.new_packets = generatePoisson(λ Δt) 7.for each new_packet in new_packets: 8.S.add( id=new_id, s=x_inj, t0=currentTime, c=0 ) 9.# Remove extracted or expired data packets 10.for each removal in removal_events: 11.S.remove(removal.id) 12. return S Algorithm 3: PPO-based addressing agent training 1. Initialize the policy network π_θ, the value network V_φ, and the experience buffer. 2. For episode = 1 to M: 3. Reset environment state S0 4. for step = 1 to T: 5. Observation state O_t 6. Calculate the action probability distribution π_θ(·|O_t), and sample the action a_t = (t_arrival, δ). 7. Execute the action, obtain the actual arrival time t_actual, and calculate the reward R_t. 8. State transition to O_{t+1} 9. Store (O_t, a_t, log π_θ(a_t|O_t), R_t, V_φ(O_t), O_{t+1}) to Buffer 10.if Buffer size>= batch_size: 11. for epochs in 1..K_epochs: 12. Sample minibatch from Buffer 13. Calculate the GAE dominance function Â_t 14. Calculate the policy loss L_clip and the value loss L_value. 15. Update θ ← θ + α_actor _θ L_clip 16. Update φ ← φ - α_critic _φ L_value 17. Clear Buffer Through the above training, the agent can learn the optimal time window prediction strategy, thereby accurately capturing target optical data packets during actual addressing.

[0058] Some embodiments of this application provide a dynamic addressing device for an optical fiber loop. The device includes a controller, an optical fiber loop, a data injection module, an optical monitoring node, and an optical capture module. At least one optical monitoring node, at least one data injection module, and at least one optical capture module are arranged on the optical fiber loop. The controller is connected to the optical monitoring node, the data injection module, and the optical capture module. The controller includes: The data acquisition module is used to respond to a data read request by querying the maintained global dynamic mapping table and obtaining the most recent observation information of the node of the target data packet corresponding to the data read request. The data analysis module is used to determine the expected arrival time window of the target data packet to the light acquisition module based on the preset dynamic addressing algorithm and the most recent observation information of the node. A data extraction module is used to activate the optical capture module within the time window to match and extract the target data packet from the optical fiber loop.

[0059] In some embodiments, the data analysis module is further configured to acquire the physical parameters of the optical fiber loop and the data packet injection parameters of the data injection module; and based on the physical parameters and the data packet injection parameters, to establish a state-space model describing the position distribution and time relationship of optical data packets in the optical fiber loop, wherein the state-space model is used to determine the position information of the target data packet at any given time.

[0060] Optionally, the optical monitoring node continuously detects optical data packets passing through its location to obtain monitoring information of the optical data packets; the data analysis module is also used to update the corresponding entries of the optical data packets in the global dynamic mapping table according to the monitoring information; when the status information of any data packet meets the preset deletion status, the controller deletes the corresponding entry of the arbitrary data packet from the global dynamic mapping table.

[0061] Optionally, the data analysis module is also used to control the agent to observe the current state information based on the most recent observation information of the node; and to determine the expected arrival time window of the target data packet to the light acquisition module based on the dynamic addressing algorithm and the current state information.

[0062] Optionally, the data analysis module is also used to send an activation signal to the optical capture module within the time window; control the optical capture module to continuously monitor the passing optical signals within the time window; if the target identifier of the target data packet is matched in the optical signal, control the optical capture module to obtain the target data packet from the optical fiber loop and update the global dynamic mapping table.

[0063] Optionally, the data analysis module is also used to query the global dynamic mapping table again and call the dynamic addressing algorithm again for prediction if the target data packet is not captured within the time window, until timeout or the maximum number of retries is reached; and to feed back reward information and next state information to the dynamic addressing algorithm based on the capture result information of the target data by the light capture module.

[0064] Optionally, the data analysis module is also used to adjust the time window of each capture device by using a time-slice round-robin or multi-threaded parallel prediction strategy when multiple read requests access different data packets at the same time; and to execute once by using a preset priority preemption strategy when the time windows of different target data packets overlap.

[0065] Optionally, the dynamic addressing algorithm is trained using a deep Q-network or near-end strategy optimization; the controller is implemented using a field-programmable gate array or an embedded processor; the global dynamic mapping table module is implemented using a static random access memory; and the light capture module is an optical switch based on a semiconductor optical amplifier or a micro-ring resonator.

[0066] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0067] Accordingly, embodiments of this application also provide an electronic device, which includes one or more processors; a memory; and one or more application programs, wherein the one or more application programs are stored in the memory and configured to be executed by the processors using the steps of the above-described fiber optic loop dynamic addressing method.

[0068] This application does not limit the type of processor. A processor is a circuit with signal processing capabilities. For example, a processor includes circuits with instruction read and execute capabilities, such as a CPU, microprocessor, graphics processing unit (GPU), or digital signal processor (DSP). Alternatively, a processor can be implemented through the logical relationships of hardware circuits, where these relationships are fixed or reconfigurable. For instance, a processor may include hardware circuits implemented using an ASIC or PLD, such as a field-programmable gate array (FPGA). In reconfigurable hardware circuits, the process of the processor loading a configuration document and configuring the hardware circuits can be understood as the processor loading instructions to implement the functions of some or all of the aforementioned units. Furthermore, a processor may include hardware circuits designed for artificial intelligence, which can be understood as an ASIC, such as one or more of a neural network processing unit (NPU), tensor processing unit (TPU), or deep learning processing unit (DPU).

[0069] As can be seen, each unit in the above fiber optic loop dynamic addressing device can be one or more processors (or processing circuits) configured to implement the above method, such as: CPU, GPU, NPU, TPU, DPU, microprocessor, DSP, ASIC, FPGA, or a combination of at least two of these processor types.

[0070] Furthermore, the units in the above-mentioned fiber optic loop dynamic addressing device can be integrated in whole or in part, or they can be implemented independently. In one implementation, these units are integrated together as a system-on-a-chip (SOC). The SOC may include at least one processor for implementing any of the above-mentioned fiber optic loop dynamic addressing methods or implementing the functions of each unit in the above-mentioned chip testing device.

[0071] This application does not limit the type of memory. For example, the memory may include read-only memory (ROM) or random access memory (RAM). It may include non-volatile or volatile memory. It may include on-chip or off-chip storage resources. ROM includes, for example, mask ROM, programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash ROM; RAM includes, for example, static random access memory (SRAM) or dynamic random access memory (DRAM). On-chip storage resources refer to the storage circuitry integrated within the chip. This circuitry is integrated with other circuitry (such as processing circuitry) to provide storage space for the chip. Examples include on-chip static random access memory (SRAM) resources or on-chip configuration register space resources. Off-chip storage resources refer to independent storage circuitry or chips, such as off-chip double data rate synchronous dynamic RAM (DDRSDRAM, or DDR for short).

[0072] In this disclosure, unless otherwise expressly specified and limited, ordinal numbers, such as "first," "second," etc., are used only to distinguish and describe related objects, and should not be construed as indicating or implying the relative importance or order between related objects. Furthermore, ordinal numbers do not represent the quantity of related objects.

[0073] "Multiple" includes two or more, and other classifiers are similar.

[0074] The terms "or" and "and / or" in this disclosure are used to describe relationships between related objects, indicating a non-exclusive inclusion. For example, "A and / or B" and "A or B" can both include: "A alone," "B alone," or "A and B," where "A" and "B" can include a single object or multiple objects. Similarly, "A, B and / or C," "A, B or C," and "A, B and C" can both include: "A alone," "B alone," "C alone," "A and B," "A and C," "B and C," or "A, B and C," where "A," "B," and "C" can include a single object or multiple objects. Additionally, the " / " in this disclosure is used to indicate an "or" relationship between preceding and following related objects. The meanings of "at least one of A or B" and "one or more of A and B" in this disclosure are the same as the meaning of "A or B" above. The meanings of "one or more of A, B, and C" and "at least one of A, B, or C" are the same as the meaning of "A, B, or C" above. The meaning of "one or more of A, B, and C" is the same as the meaning of "A, B, or C" above.

[0075] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not described in detail or in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Furthermore, the above embodiments can be freely combined as needed.

Claims

1. A dynamic addressing method for fiber optic loops, applied to a controller of a dynamic addressing system for fiber optic loops, wherein the dynamic addressing system for fiber optic loops further includes a fiber optic loop, a data injection module, an optical monitoring node, and an optical capture module, wherein at least one optical monitoring node, at least one data injection module, and at least one optical capture module are disposed on the fiber optic loop, and the controller is connected to the optical monitoring node, the data injection module, and the optical capture module, characterized in that... The method includes: In response to a data read request, the maintained global dynamic mapping table is queried to obtain the latest observation information of the target data packet corresponding to the data read request at the optical monitoring node. Based on a preset dynamic addressing algorithm, the expected arrival time window of the target data packet to the optical capture module is determined according to the most recent observation information of the node; The optical capture module is activated within the time window to match and extract the target data packet from the optical fiber loop.

2. The fiber optic loop dynamic addressing method according to claim 1, characterized in that, Before responding to a data read request by querying the maintained global dynamic mapping table and obtaining the most recent observation information of the node corresponding to the target data packet of the data read request, the steps also include: Obtain the physical parameters of the optical fiber loop and the data packet injection parameters of the data injection module; Based on the physical parameters and the data packet injection parameters, a state-space model is established to describe the position distribution and time relationship of optical data packets in the optical fiber loop. The state-space model is used to determine the position information of the target data packet at any time.

3. The fiber optic loop dynamic addressing method according to claim 1, characterized in that, Before responding to a data read request by querying the maintained global dynamic mapping table and obtaining the most recent observation information of the node corresponding to the target data packet of the data read request, the steps also include: The optical monitoring node continuously detects optical data packets passing through its location to obtain monitoring information of the optical data packets; The controller updates the corresponding entries for optical data packets in the global dynamic mapping table based on the monitoring information; When the status information of any data packet meets the preset deletion status, the controller deletes the corresponding entry of the arbitrary data packet from the global dynamic mapping table.

4. The fiber optic loop dynamic addressing method according to claim 1, characterized in that, The steps, based on a preset dynamic addressing algorithm, determine the expected arrival time window of the target data packet to the optical capture module according to the most recent observation information of the node, including: The controlling agent observes the current state information based on the most recent observation information of the node; Based on the dynamic addressing algorithm and the current state information, the expected arrival time window for the target data packet to reach the optical capture module is determined.

5. The fiber optic loop dynamic addressing method according to claim 4, characterized in that, The step of activating the optical capture module within the time window and matching and extracting the target data packet from the fiber optic loop includes: An activation signal is sent to the light capture module within the time window; The light capture module is controlled to continuously monitor passing light signals within the time window; If the target identifier of the target data packet is matched in the optical signal, the optical capture module is controlled to obtain the target data packet from the optical fiber loop and update the global dynamic mapping table.

6. The fiber optic loop dynamic addressing method according to claim 4, characterized in that, After activating the optical capture module within the time window and matching and extracting the target data packet from the fiber optic loop, the step further includes: If the target data packet is not captured within the time window, the global dynamic mapping table is queried again and the dynamic addressing algorithm is called again for prediction until the timeout or the maximum number of retries is reached. Based on the capture result information of the target data by the light capture module, the reward information and the next state information are fed back to the dynamic addressing algorithm.

7. The fiber optic loop dynamic addressing method according to claim 1, characterized in that, In the optical fiber loop, optical carriers of different wavelengths independently carry data packet streams, and the method further includes: When multiple read requests access different data packets simultaneously, a time-slice round-robin or multi-threaded parallel prediction strategy is used to adjust the time window of each capture device. When the time windows of different target data packets overlap, a preset priority preemption strategy is used to execute them one at a time.

8. The fiber optic loop dynamic addressing method according to claim 1, characterized in that, The dynamic addressing algorithm is trained using a deep Q-network or near-end strategy optimization; the controller is implemented using a field-programmable gate array or an embedded processor; the global dynamic mapping table module is implemented using a static random access memory; and the light capture module is an optical switch based on a semiconductor optical amplifier or a micro-ring resonator.

9. A fiber optic loop dynamic addressing device, characterized in that, The device includes a controller, an optical fiber loop, a data injection module, an optical monitoring node, and an optical capture module. At least one optical monitoring node, at least one data injection module, and at least one optical capture module are disposed on the optical fiber loop. The controller is connected to the optical monitoring node, the data injection module, and the optical capture module. The controller includes: The data acquisition module is used to respond to a data read request by querying the maintained global dynamic mapping table and obtaining the most recent observation information of the node of the target data packet corresponding to the data read request. The data analysis module is used to determine the expected arrival time window of the target data packet to the light acquisition module based on the preset dynamic addressing algorithm and the most recent observation information of the node. A data extraction module is used to activate the optical capture module within the time window to match and extract the target data packet from the optical fiber loop.

10. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the fiber optic loop dynamic addressing method according to any one of claims 1 to 8.

Images