A PON distributed intelligent computing task scheduling method

Abstract

The application discloses a PON distributed intelligent computing task scheduling method and belongs to the technical field of optical access network communication technology. The application obtains the link attenuation coefficients of each optical network unit through an optical line terminal and broadcasts; the allocation information of a differential double-wavelength channel and a gradient vector is established and broadcasted; a differential double-wavelength mapping is used to allocate target optical power to gradient components, and each unit is instructed to transmit after compensation in combination with the attenuation coefficients; optical signals are received on the differential double-wavelength channel and are converted into photoelectric current, and a global gradient vector is constructed after dark current is eliminated; the maximum value of the net photoelectric current difference value is counted to adaptively update the mapping coefficients and broadcast, and the global model parameters are updated and issued to each unit to complete synchronization. The application solves the problems of communication bottlenecks in the existing gradient aggregation and a large dynamic range of optical signal power.

Description

Technical Field

[0001] This invention relates to the field of optical access network communication technology, and specifically to a PON distributed intelligent computing task scheduling method. Background Technology

[0002] With the rapid development of artificial intelligence technology, especially the continuous expansion of deep learning model scale, distributed intelligent computing has become an important means to improve model training efficiency. In an optical access network (PON, Passive Optical Network) environment, the optical line terminal (OLT) and multiple optical network units (ONUs) form a point-to-multipoint physical topology, naturally possessing the ability to aggregate and distribute data. Applying PON networks to distributed machine learning training can leverage their high bandwidth, low latency, and natural multi-point aggregation characteristics to achieve rapid aggregation and updating of model parameters.

[0003] Currently, in PON-based distributed machine learning systems, a common gradient aggregation method is a centralized digital domain aggregation approach: "receive first, decode then sum." Specifically, after each optical network unit (ONU) calculates its model gradient locally, it quantizes the gradient vector into a digital bitstream and sequentially transmits it to the optical line terminal (OLT) as a digital optical signal, according to a pre-allocated time window or wavelength channel. The OLT receives the optical signal from each ONU, performs photoelectric conversion, analog-to-digital conversion, demodulation, and decoding to recover the gradient vector values ​​of each node. Then, it performs element-wise summation and averaging in the digital domain to obtain the global gradient vector, which is used to update the global model.

[0004] While the aforementioned existing technologies can achieve gradient aggregation, they present the following technical problems in practical PON environments: As the number of optical network units (ONUs) increases or the dimension of the gradient vector grows, the optical line terminal (OLT) needs to sequentially receive and decode the complete gradient data from all nodes. This leads to a linear increase in uplink bandwidth usage with the number of nodes, significantly increasing communication latency and creating a "communication bottleneck." Furthermore, the transmit power of each ONU is affected by differences in link attenuation, resulting in a large dynamic range of optical signal power reaching the OLT. Direct digital optical communication requires complex power control and dynamic range adaptation mechanisms; otherwise, receiver overload or insufficient signal-to-noise ratio (SNR) issues may easily occur. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a PON distributed intelligent computing task scheduling method that solves the problems of communication bottlenecks and large dynamic range of optical signal power in existing gradient aggregation methods.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a PON distributed intelligent computing task scheduling method, comprising the following steps:

[0007] S1. Based on the fixed transmit power and actual delivered power of the optical network unit, the optical line terminal obtains the link attenuation coefficient of each optical network unit and broadcasts it to all optical network units.

[0008] S2. Establish allocation information for multiple sets of differential dual-wavelength channels and gradient vectors to be aggregated on the optical line terminal, and broadcast the allocation information to each optical network unit.

[0009] S3. Based on the allocation information, a differential dual-wavelength mapping method is used to allocate target optical power values ​​to the gradient components, and compensation is performed based on the link attenuation coefficient to obtain the compensated optical power value.

[0010] S4. Instruct all optical network units to transmit optical signals to compensate for optical power values ​​via the optical line terminal;

[0011] S5. Receive the optical signal on each differential dual-wavelength channel through the optical line terminal, convert it into photocurrent, eliminate dark current, obtain the global gradient component sum, and construct the global gradient vector.

[0012] S6. The maximum value of the net photocurrent difference is statistically analyzed through the optical line terminal to adaptively update the mapping coefficients and broadcast them. At the same time, the global model parameters are updated according to the global gradient vector and sent to each optical network unit to complete local synchronization and confirmation feedback.

[0013] Furthermore, S2 includes the following sub-steps:

[0014] S21. Training data and labels are transmitted via optical link terminal;

[0015] S22. Based on the training data and labels, perform local backpropagation on the optical network unit to obtain the gradient vector to be aggregated, and upload the dimension of the gradient vector to be aggregated to the optical line terminal.

[0016] S23. At the optical line terminal, the number of wavelength channels is preset according to the dimension of the gradient vector to be aggregated.

[0017] S24. Divide at the optical line terminal Each positive wavelength and For each negative wavelength, pair the positive wavelength with the negative wavelength of the same index to obtain... Group differential dual-wavelength channels, This refers to the number of wavelength channels;

[0018] S25. At the optical line terminal, formulate the allocation rules for the gradient component indices in the gradient vector to be aggregated, and allocate each index to each group of differential dual-wavelength channels to obtain the allocation information.

[0019] S26. The allocation information is broadcast to each optical network unit through the optical line terminal.

[0020] Furthermore, S3 includes the following sub-steps:

[0021] S31. Based on the allocation information issued by the optical line terminal, each gradient component in the gradient vector to be aggregated is assigned to a designated differential dual-wavelength channel, and a target optical power value is assigned to each gradient component in the gradient vector to be aggregated using a differential dual-wavelength mapping method.

[0022] S32. Based on its own link attenuation coefficient, the target optical power value is pre-compensated on the optical network unit to obtain the compensated optical power value.

[0023] Furthermore, the specific process of the differential dual-wavelength mapping method includes:

[0024] When each gradient component in the gradient vector to be aggregated is greater than or equal to 0:

[0025] , ,

[0026] When each gradient component in the gradient vector to be aggregated is less than 0:

[0027] , ,

[0028] in, For optical network units, the gradient component at the positive wavelength is... The target optical power value, For mapping coefficients, For the first The gradient vector to be aggregated in the nth optical network unit One gradient component, For optical network units, the gradient component at the negative wavelength is... The target optical power value, Here, | represents the reference power at the positive and negative wavelengths, and | represents the absolute value. This is an index for an optical network cell. This is the index of the gradient component.

[0029] Furthermore, the specific process of S32 includes: adjusting the optical network unit at the positive wavelength for the gradient component. The ratio of the target optical power value to its own link attenuation coefficient is used as the compensation optical power value corresponding to the positive wavelength;

[0030] The optical network unit is configured with respect to the gradient component at the negative wavelength. The ratio of the target optical power value to its own link attenuation coefficient is used as the compensation optical power value corresponding to the negative wavelength.

[0031] Furthermore, S4 includes the following sub-steps:

[0032] S41. Broadcast an aggregation trigger frame on the optical line terminal by controlling the wavelength. The aggregation trigger frame includes: specifying the transmission time and the time slot length.

[0033] S42. All optical network units simultaneously transmit optical signals with a power equal to the compensated optical power value at a specified transmission time, the transmission time lasting for the duration of the time slot.

[0034] Furthermore, S5 includes the following sub-steps:

[0035] S51. Through the optical line terminal, at the positive and negative wavelengths corresponding to each group of differential dual-wavelength channels, the optical signals sent by all optical network units that belong to the gradient components responsible for that channel are simultaneously received.

[0036] S52. In each group of differential dual-wavelength channels, obtain the total transmit power of the optical signal of each optical network unit, and obtain the positive wavelength received optical power and negative wavelength received optical power of each group.

[0037] S53. The photodetector PD at the optical line terminal converts the received optical power of the positive wavelength and the received optical power of the negative wavelength of each group into photocurrent, so as to obtain the photocurrent of the positive wavelength and the photocurrent of the negative wavelength of each group.

[0038] S54. Subtract the pre-stored dark current from the photocurrent of the positive wavelength and the photocurrent of the negative wavelength of each group to obtain the net photocurrent of the positive wavelength and the net photocurrent of the negative wavelength of each group.

[0039] S55. Based on the difference between the net photocurrent of the positive wavelength and the net photocurrent of the negative wavelength in each group, calculate the global gradient value of each time slot in that group.

[0040] S56. Collect the global gradient values ​​calculated from all channels and all time slots through the optical line terminal, and fill them back into the corresponding positions of the global gradient vector one by one according to the original index of each global gradient value, so as to construct the global gradient vector.

[0041] Furthermore, the formula for calculating the sum of global gradient values ​​for each time slot in this group in S55 is as follows:

[0042] ,

[0043] in, In the first Group differential dual-wavelength channels in the first The global sum of the gradient components output in each time slot. For the first The positive wavelength of the differential dual-wavelength channel is at the first Net photocurrent per time slot For the first The negative wavelength of the group differential dual-wavelength channel is at the first Net photocurrent per time slot For photodetector PD responsivity, For mapping coefficients, For group indexes.

[0044] Furthermore, the process in S6 of adaptively updating the mapping coefficients and broadcasting the maximum value of the net photocurrent difference through the optical line terminal includes:

[0045] A1. Calculate the maximum net photocurrent difference of each differential dual-wavelength channel in the current round of aggregation through the optical line terminal;

[0046] A2. The ratio of the maximum net photocurrent difference to the saturation current threshold of the photodetector PD is taken as the current dynamic range utilization rate.

[0047] A3. When the current dynamic range utilization is less than 0.5, the optical line terminal increases the mapping coefficient:

[0048] ,

[0049] in, For the new mapping coefficients, The mapping coefficients for the current round of aggregation. For the current dynamic range utilization, For adjustment coefficients, ;

[0050] A4. When the current dynamic range utilization is greater than 0.9, the optical line terminal reduces the mapping coefficient:

[0051] ,

[0052] A5. The updated version is transmitted via optical line terminal. Broadcast to all optical network units in the next round of aggregation trigger frames.

[0053] Furthermore, the process in S6 of updating the global model parameters based on the global gradient vector and distributing them to each optical network unit to complete local synchronization and confirmation feedback includes:

[0054] B1. Obtain the global model parameters for the current round through the optical line terminal;

[0055] B2. At the optical line terminal, the global model parameters are updated using gradient descent based on the global gradient vector.

[0056] B3. The optical line terminal broadcasts the updated global model parameters to all optical network units;

[0057] B4. Each optical network unit receives the updated global model parameters and sends an acknowledgment frame to the optical line terminal.

[0058] The beneficial effects of this invention are as follows:

[0059] 1. This invention utilizes the inherent broadcasting and aggregation characteristics of PON networks to transform gradient aggregation from the traditional "serial reception and summation in the digital domain" to "parallel superposition and summation in the optical power domain." The gradient components of all optical network units are directly summed at the physical layer through optical power superposition at the optical line terminal, avoiding the serial transmission, demodulation, and decoding processes of massive amounts of data. This prevents uplink communication bandwidth from increasing linearly with the number of nodes, significantly reducing communication latency and solving the "communication bottleneck" problem in distributed training.

[0060] 2. This invention pre-calibrates and broadcasts the link attenuation coefficients of each optical network unit, allowing each node to pre-compensate its transmit power, thus ensuring that the optical signal power reaching the optical line terminal is consistent across all nodes. Simultaneously, a differential dual-wavelength mapping mechanism is introduced to map gradient components onto two sets of wavelengths, positive and negative. The magnitude of the gradient components is reflected by the difference in net photocurrent at the receiving end, and the mapping coefficients are adaptively adjusted based on dynamic range utilization. This ensures that gradient information is transmitted within the linear operating range of the photodetector, effectively suppressing interference caused by link attenuation differences and the dynamic range of the receiving signal, thereby improving aggregation accuracy and system stability.

[0061] 3. This invention proposes a multi-group differential dual-wavelength channel parallel aggregation architecture. The high-dimensional gradient vector is split and distributed into multiple independent differential dual-wavelength channels. Gradient aggregation calculations are performed in parallel on each channel, and then the global gradient vector is reconstructed according to the original indices. This multi-channel parallel processing method significantly improves the aggregation speed of high-dimensional gradients, enabling the system to efficiently support distributed training tasks of large-scale deep learning models, while maintaining good compatibility with existing PON optical access networks. Attached Figure Description

[0062] Figure 1 This is a flowchart of a PON distributed intelligent computing task scheduling method. Detailed Implementation

[0063] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0064] like Figure 1As shown, a PON distributed intelligent computing task scheduling method includes the following steps:

[0065] S1. Based on the fixed transmit power and actual delivered power of the optical network unit, the optical line terminal obtains the link attenuation coefficient of each optical network unit and broadcasts it to all optical network units.

[0066] S2. Establish allocation information for multiple sets of differential dual-wavelength channels and gradient vectors to be aggregated on the optical line terminal, and broadcast the allocation information to each optical network unit.

[0067] S3. Based on the allocation information, a differential dual-wavelength mapping method is used to allocate target optical power values ​​to the gradient components, and compensation is performed based on the link attenuation coefficient to obtain the compensated optical power value.

[0068] S4. Instruct all optical network units to transmit optical signals to compensate for optical power values ​​via the optical line terminal;

[0069] S5. Receive the optical signal on each differential dual-wavelength channel through the optical line terminal, convert it into photocurrent, eliminate dark current, obtain the global gradient component sum, and construct the global gradient vector.

[0070] S6. The maximum value of the net photocurrent difference is statistically analyzed through the optical line terminal to adaptively update the mapping coefficients and broadcast them. At the same time, the global model parameters are updated according to the global gradient vector and sent to each optical network unit to complete local synchronization and confirmation feedback.

[0071] This invention is applied to a passive optical network architecture consisting of an optical line terminal, a splitter, and several optical network units.

[0072] In this embodiment, the formula for calculating the link attenuation coefficient in S1 is: ,in, For the first Link attenuation coefficient of each optical network unit, The fixed transmission power set for the testing process For the first The actual power achieved by the calibration signal of each optical network unit at the optical line terminal.

[0073] In this embodiment, S2 includes the following sub-steps:

[0074] S21. Training data and labels are transmitted via optical link terminal;

[0075] S22. Based on the training data and labels, perform local backpropagation on the optical network unit to obtain the gradient vector to be aggregated, and upload the dimension of the gradient vector to be aggregated to the optical line terminal.

[0076] S23. At the optical line terminal, the number of wavelength channels is preset according to the dimension of the gradient vector to be aggregated: ,in, This represents the maximum number of parallel channels supported by the hardware. For single gradient component propagation time, Aggregate latency for the target. For dimensional scale, To round up;

[0077] S24. Divide at the optical line terminal Positive wavelength and Negative wavelength , will the The positive wavelength and the first By pairing the negative wavelengths, we obtain Group differential dual-wavelength channels, This refers to the number of wavelength channels;

[0078] S25. At the optical line terminal, formulate the allocation rules for the gradient component indices in the gradient vector to be aggregated, and allocate each index to each group of differential dual-wavelength channels to obtain the allocation information.

[0079] S26. The allocation information is broadcast to each optical network unit through the optical line terminal.

[0080] The target aggregation delay is the maximum preset delay allowed for the entire gradient aggregation process, from the moment each optical network unit begins to synchronously transmit gradient component optical signals until the optical line terminal completes the calculation of all dimensional gradient components and constructs a complete global gradient vector.

[0081] The gradient vector to be aggregated is: ,in, For the first The gradient vectors to be aggregated from each optical network unit. For the first The first gradient component in the gradient vector to be aggregated of the optical network units. For the first The gradient vector to be aggregated in the nth optical network unit One gradient component, For the first The gradient vector to be aggregated in the nth optical network unit One gradient component, For the index of the gradient components, For dimensional scale.

[0082] The allocation rule in S25 is: Group differential dual-wavelength channel gradient component index set , , For the first The set of gradient component indices responsible for the group differential dual-wavelength channels. This refers to the number of wavelength channels. For example, the size of the gradient vector to be aggregated. The number of wavelength channels is 10. The group index of the channel is 3. When the value is 1, the first index set is , The corresponding gradient component range is: ; Channel group index When the value is 2, the second index set is , The corresponding gradient component range is: ; Channel group index When the value is 3, the third index set is , The corresponding gradient component range is: , For the first Each optical network unit has 1 to 10 gradient components. Since there are multiple optical network units, each optical network unit has a gradient vector to be aggregated and has the same allocation rule. Therefore, for the first group of differential dual-wavelength channels, in the first time slot, the optical signals corresponding to the gradient components transmitted by multiple optical network units will exist simultaneously on its positive wavelength, and the optical signals corresponding to the same gradient component transmitted by multiple optical network units will exist simultaneously on its negative wavelength.

[0083] The positive and negative wavelengths are two paired independent optical wavelengths that carry the positive and negative information of the gradient component, respectively.

[0084] This invention adaptively determines the number of parallel wavelength channels based on the gradient vector dimension, single-component transmission duration, and target aggregation delay, combined with the hardware maximum channel limit, and constructs multiple sets of differential dual-wavelength channels. The high-dimensional gradient vector is evenly split and distributed to each channel through a fixed index polling method, realizing multi-channel load balancing and controllable delay. At the same time, the channel allocation information is uniformly broadcast to each optical network unit to ensure the synchronization of configuration rules for all network nodes.

[0085] In this embodiment, S3 includes the following sub-steps:

[0086] S31. Based on the allocation information issued by the optical line terminal, each gradient component in the gradient vector to be aggregated is assigned to a designated differential dual-wavelength channel, and a target optical power value is assigned to each gradient component in the gradient vector to be aggregated using a differential dual-wavelength mapping method.

[0087] S32. Based on its own link attenuation coefficient, the target optical power value is pre-compensated on the optical network unit to obtain the compensated optical power value.

[0088] In this embodiment, the specific process of differential dual-wavelength mapping includes:

[0089] When each gradient component in the gradient vector to be aggregated is greater than or equal to 0:

[0090] , ,

[0091] When each gradient component in the gradient vector to be aggregated is less than 0:

[0092] , ,

[0093] in, For optical network units, the gradient component at the positive wavelength is... The target optical power value, For mapping coefficients, For the first The gradient vector to be aggregated in the nth optical network unit One gradient component, For optical network units, the gradient component at the negative wavelength is... The target optical power value, Here, | represents the reference power at the positive and negative wavelengths, and | represents the absolute value. This is an index for an optical network cell. This is the index of the gradient component.

[0094] For example, the gradient component index set handled by the first group of differential dual-wavelength channels is: The corresponding gradient components are These gradient components are transmitted sequentially in different time slots according to their index order: Time Slot send time slot send time slot send time slot send Within each time slot, this gradient component is at the positive wavelength of group 1. and negative wavelength Power mapping is performed simultaneously. This is the time slot number.

[0095] In this embodiment, the specific process of S32 includes: adjusting the optical network unit at the positive wavelength for the gradient component. The ratio of the target optical power value to its own link attenuation coefficient is used as the compensation optical power value corresponding to the positive wavelength. ;

[0096] The optical network unit is configured with respect to the gradient component at the negative wavelength. The ratio of the target optical power value to its own link attenuation coefficient is used as the compensation optical power value corresponding to the negative wavelength. .

[0097] This invention precisely maps each gradient component to a designated differential dual-wavelength channel by combining channel allocation information issued by the optical line terminal, ensuring a one-to-one match between gradient components and transmission channels. It employs a differential dual-wavelength mapping method to convert the magnitude of gradient components into optical power differences, achieving accurate optical domain representation of gradient information, effectively distinguishing between positive and negative gradient components, and avoiding gradient information confusion. Simultaneously, it pre-compensates the target optical power based on the link attenuation coefficient of the optical network unit itself, offsetting the optical power loss caused by differences in link attenuation at different nodes. This ensures that the optical signals emitted by each optical network unit reach the optical line terminal with consistent power, avoiding overload or insufficient signal-to-noise ratio at the receiving end, guaranteeing the accuracy and stability of gradient aggregation, and further improving the reliability and efficiency of distributed intelligent computing task scheduling.

[0098] In this embodiment, S4 includes the following sub-steps:

[0099] S41. Broadcast an aggregation trigger frame on the optical line terminal by controlling the wavelength. The aggregation trigger frame includes: specifying the transmission time and the time slot length.

[0100] S42. All optical network units simultaneously transmit optical signals with a power equal to the compensated optical power value at a specified transmission time, the transmission time lasting for the duration of the time slot.

[0101] Control wavelength: A dedicated wavelength used specifically for issuing commands and scheduling signals. Time slot length is the duration of each time slot, that is, the time from the start of transmission in that time slot to the end of transmission in that time slot.

[0102] In this embodiment, S5 includes the following sub-steps:

[0103] S51. Through the optical line terminal, at the positive and negative wavelengths corresponding to each group of differential dual-wavelength channels, the optical signals sent by all optical network units that belong to the gradient components responsible for that channel are simultaneously received.

[0104] S52. Within each group of differential dual-wavelength channels, obtain the total transmit power of the optical signal of each optical network unit to obtain the positive wavelength receive optical power of each group. and negative wavelength received optical power ;

[0105] S53. The photodetector PD at the optical line terminal receives the positive wavelength optical power of each group. and negative wavelength received optical power Converted into photocurrent, the photocurrent of the positive wavelength and the photocurrent of the negative wavelength of each group are obtained;

[0106] S54. Subtract the pre-stored dark current from the photocurrent of the positive wavelength and the photocurrent of the negative wavelength of each group to obtain the net photocurrent of the positive wavelength and the net photocurrent of the negative wavelength of each group.

[0107] S55. Based on the difference between the net photocurrent at the positive wavelength and the net photocurrent at the negative wavelength in each group, calculate the global gradient value for each time slot in that group. ;

[0108] S56. Collect the global gradient values ​​calculated from all channels and all time slots through the optical line terminal. According to each Corresponding raw index Fill it back into the global gradient vector. After traversing all channels and all time slots at each location, a global gradient vector is constructed. ,in, For the global gradient vector, The sum of the global gradient values ​​of the first gradient component. The sum of the global gradient values ​​of the second gradient component. For the first The sum of the global gradient values ​​of the gradient components.

[0109] S52 specifically includes: obtaining the total transmit power of the optical signal of each optical network unit at the positive wavelength of each group of differential dual-wavelength channels, and obtaining the received optical power at the positive wavelength; obtaining the total transmit power of the optical signal of each optical network unit at the negative wavelength of each group of differential dual-wavelength channels, and obtaining the received optical power at the negative wavelength.

[0110] The gradient components on all optical network units (ONUs) have the same dimensionality. On each ONU, multiple gradient components belonging to the same differential dual-wavelength channel are transmitted sequentially in index order in different time slots. In each time slot, one gradient component is transmitted simultaneously through the positive and negative wavelengths of that channel. All ONUs transmit synchronously in the same time slot and on the same wavelength, and the optical line termination (OLT) achieves global aggregation of the gradient components through optical domain superposition.

[0111] For example, for the first group of differential dual-wavelength channels:

[0112] At the positive wavelength Above, optical network unit 1, optical network unit 2, ..., optical network unit All at the same time and on the same wavelength The optical network unit emits optical signals. Each optical network unit... The emitted optical power is encoded according to the following rules: when the gradient component of the cell... At that time, the transmission power was ;when At that time, the transmission power was These optical signals are naturally superimposed in the optical fiber and arrive at the optical line terminal (OPT) simultaneously. The photodetector (PD) at the OPT then... Only one total optical power can be measured, which includes the reference power of all optical network units and the contribution of all positive gradient components;

[0113] In the negative wavelength Above, optical network unit 1, optical network unit 2, ..., optical network unit At the same time and on the same wavelength It emits optical signals. The emission rule is opposite to that of the positive wavelength: when the gradient component of the unit... At that time, the transmission power was ;when At that time, the transmission power was These optical signals are also naturally superimposed in the optical fiber, and the photodetector (PD) at the optical line terminal... Another total optical power is measured, which includes the reference power of all optical network units and the contribution of the absolute values ​​of all negative gradient components.

[0114] S53 specifically includes: converting the received optical power of the positive wavelength and the received optical power of the negative wavelength of each group into photocurrent through the photodetector PD at the optical line terminal, and obtaining the photocurrent measurement value of the positive wavelength and the photocurrent measurement value of the negative wavelength of each group through the analog-to-digital converter ADC.

[0115] Dark current: refers to the inherent current generated inside a photodetector (PD) when there is no light and only the operating voltage is applied.

[0116] In this embodiment, the formula for calculating the sum of global gradient values ​​for each time slot in S55 is as follows:

[0117] ,

[0118] in, In the first Group differential dual-wavelength channels in the first The global sum of the gradient components output in each time slot. For the first The positive wavelength of the differential dual-wavelength channel is at the first Net photocurrent per time slot For the first The negative wavelength of the group differential dual-wavelength channel is at the first Net photocurrent per time slot For photodetector PD responsivity, For mapping coefficients, For group indexes.

[0119] This invention achieves precise correspondence between gradient components and transmission channels by receiving gradient component optical signals from each optical network unit via differential dual-wavelength channels, avoiding signal crosstalk between different channels and ensuring the orderliness of multi-channel parallel aggregation. By acquiring the total received optical power of each group of positive and negative wavelengths and completing photoelectric conversion, the gradient information superimposed in the optical domain is efficiently converted into an electrical domain signal. By subtracting the pre-stored dark current, the inherent noise interference of the photodetector (PD) is effectively eliminated, improving the purity and accuracy of the net photocurrent and ensuring the accuracy of gradient calculation. Based on the difference between the positive and negative net photocurrents, the global gradient sum of each time slot in each group is calculated. Relying on the inverse process of differential dual-wavelength mapping, the gradient information is accurately restored. At the same time, by combining the correspondence between time slots and channels and backfilling the combination according to the original index, a complete global gradient vector is efficiently constructed.

[0120] In this embodiment, the process in S6 of statistically analyzing the maximum net photocurrent difference through the optical line terminal to adaptively update the mapping coefficients and broadcast the result includes:

[0121] A1. By statistically analyzing the maximum net photocurrent difference of each differential dual-wavelength channel in the current aggregation round using the optical line terminal, the absolute value of the difference between the net photocurrent of the positive wavelength and the net photocurrent of the negative wavelength is calculated for each group of differential dual-wavelength channels in each time slot, thus obtaining the absolute difference of net photocurrent. The maximum value of the absolute difference of net photocurrent in all channels and all time slots during the current round of aggregation is taken as the maximum value of net photocurrent difference.

[0122] A2. The ratio of the maximum net photocurrent difference to the saturation current threshold of the photodetector PD is taken as the current dynamic range utilization rate.

[0123] A3. When the current dynamic range utilization is less than 0.5, the optical line terminal increases the mapping coefficient:

[0124] ,

[0125] in, For the new mapping coefficients, The mapping coefficients for the current round of aggregation. For the current dynamic range utilization, For adjustment coefficients, ;

[0126] A4. When the current dynamic range utilization is greater than 0.9, the optical line terminal reduces the mapping coefficient:

[0127] ,

[0128] A5. The updated version is transmitted via optical line terminal. Broadcast to all optical network units in the next round of aggregation trigger frames.

[0129] The current dynamic range utilization is within the range of 0.5 to 0.9, and the mapping coefficient remains unchanged. The adjustment coefficient is an empirical preset value, with a typical value of 0.7. The specific value can be adjusted according to the characteristics of the photodetector, the gradient distribution range, and the system stability requirements of the training task.

[0130] During the first round of training, The initial value is set as the initial value. The initial value can be calculated by dividing the difference between the maximum safe transmit power and the minimum resolvable transmit power by the difference between the upper limit and the lower limit of gradient clipping. The maximum safe transmit power is determined by the performance of the laser, the minimum resolvable transmit power is determined by the sensitivity of the photodetector (PD), and the upper and lower limits of gradient clipping are preset according to the statistical distribution range of the gradient in the deep learning model.

[0131] This invention uses real-time statistics of the maximum net photocurrent difference between each differential dual-wavelength channel at the optical line terminal, combined with the saturation current threshold of the photodetector (PD) to define the dynamic range utilization rate, characterizing the actual occupancy of the PD's working range. Based on the dynamic range utilization rate, a reasonable threshold range is set. When the utilization rate is too low, the mapping coefficient is adaptively increased to enhance the gradient optical signal modulation amplitude, fully utilizing the PD's linear response range and avoiding quantization errors and decreased detection accuracy caused by weak signals. When the utilization rate is too high, the mapping coefficient is adaptively decreased to effectively avoid signal clipping and distortion caused by the photocurrent exceeding the saturation threshold, ensuring the linearity of photoelectric conversion and gradient calculation. This ensures that the PD always operates within its optimal linear working range.

[0132] In this embodiment, the process in S6 of updating the global model parameters based on the global gradient vector and sending them to each optical network unit to complete local synchronization and confirmation feedback includes:

[0133] B1. Obtain the global model parameters for the current round via the optical line terminal. ;

[0134] B2. At the optical line terminal, update the global model parameters using gradient descent based on the global gradient vector: ,in, For the first Global model parameters for each round, For the first Global model parameters for each round (all trainable parameters of the neural network model, such as the weight matrix and bias vector of each layer). For learning rate, This is the global gradient vector;

[0135] B3. The optical line terminal will update the global model parameters. Broadcast to all optical network units;

[0136] B4. Each optical network unit receives the updated global model parameters. And send an acknowledgment frame to the optical line terminal.

[0137] This invention addresses the "communication bottleneck" problem by employing optical domain simulation overlay technology. All optical network units (ONUs) simultaneously transmit optical signals carrying gradient information at the same wavelength. The optical line terminal (OLT) directly receives the total optical power after overlay and calculates the sum of gradient components of all ONUs in a single operation using the net photocurrent difference. The time required for gradient aggregation is related to the transmission duration of a single gradient component and does not increase with the number of ONUs, thus eliminating the bottleneck of linear growth in uplink bandwidth with the number of nodes.

[0138] This invention addresses the problem of "excessive dynamic range caused by link attenuation differences." It obtains the link attenuation coefficient of each optical network unit through pre-calibration and performs power pre-compensation at the transmitting end, ensuring that optical signals from different nodes reach the optical line terminal with the same power mapping relationship. Simultaneously, a differential dual-wavelength mapping mechanism is employed to encode positive and negative gradient information onto two independent wavelengths. The reference power term is eliminated through the net photocurrent difference, and the dynamic range utilization rate is adaptively adjusted to adjust the mapping coefficient, ensuring that the photocurrent always operates within the linear region of the photodetector. This avoids complex receiver power control and effectively suppresses interference caused by link attenuation differences and dynamic range fluctuations.

[0139] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A PON distributed intelligent computing task scheduling method, characterized in that, Includes the following steps: S1. Based on the fixed transmit power and actual delivered power of the optical network unit, the optical line terminal obtains the link attenuation coefficient of each optical network unit and broadcasts it to all optical network units. S2. Establish allocation information for multiple sets of differential dual-wavelength channels and gradient vectors to be aggregated on the optical line terminal, and broadcast the allocation information to each optical network unit. At the optical line terminal, the number of wavelength channels is preset according to the dimension of the gradient vector to be aggregated; Division at optical line terminal Each positive wavelength and By pairing the positive and negative wavelengths of the same index, we obtain... Group differential dual-wavelength channels, This refers to the number of wavelength channels; At the optical line terminal, the allocation rules for the gradient component indices in the gradient vector to be aggregated are formulated, and each index is assigned to each group of differential dual-wavelength channels to obtain the allocation information. S3. Based on the allocation information, a differential dual-wavelength mapping method is used to allocate target optical power values ​​to the gradient components, and compensation is performed based on the link attenuation coefficient to obtain the compensated optical power value. When each gradient component in the gradient vector to be aggregated is greater than or equal to 0: ， When each gradient component in the gradient vector to be aggregated is less than 0: ， in, For optical network units, the gradient component at the positive wavelength is... The target optical power value, For mapping coefficients, For the first The gradient vector to be aggregated in the nth optical network unit One gradient component, For optical network units, the gradient component at the negative wavelength is... The target optical power value, Here, | represents the reference power at the positive and negative wavelengths, and | represents the absolute value. This is an index for an optical network cell. For the index of the gradient component; S4. Instruct all optical network units to transmit optical signals to compensate for optical power values ​​via the optical line terminal; S5. Receive the optical signal on each differential dual-wavelength channel through the optical line terminal, convert it into photocurrent, eliminate dark current, obtain the global gradient component sum, and construct the global gradient vector. Subtract the pre-stored dark current from the photocurrent of the positive wavelength and the photocurrent of the negative wavelength of each group to obtain the net photocurrent of the positive wavelength and the net photocurrent of the negative wavelength of each group. Based on the difference between the net photocurrent at the positive wavelength and the net photocurrent at the negative wavelength in each group, calculate the global gradient value for each time slot in that group: ， in, In the first Group differential dual-wavelength channels in the first The global sum of the gradient components output in each time slot. For the first The positive wavelength of the differential dual-wavelength channel is at the first Net photocurrent per time slot For the first The negative wavelength of the group differential dual-wavelength channel is at the first Net photocurrent per time slot For photodetector PD responsivity, For mapping coefficients, For group index, For time slot numbering; The global gradient values ​​calculated from all channels and all time slots are collected through the optical line terminal. Based on each global gradient value and its corresponding original index, they are backfilled into the corresponding positions of the global gradient vector to construct the global gradient vector. S6. The maximum value of the net photocurrent difference is statistically analyzed through the optical line terminal to adaptively update the mapping coefficients and broadcast them. At the same time, the global model parameters are updated according to the global gradient vector and sent to each optical network unit to complete local synchronization and confirmation feedback.

2. The PON distributed intelligent computing task scheduling method according to claim 1, characterized in that, S2 includes the following steps: S21. Training data and labels are transmitted via optical link terminal; S22. Based on the training data and labels, perform local backpropagation on the optical network unit to obtain the gradient vector to be aggregated, and upload the dimension of the gradient vector to be aggregated to the optical line terminal.

3. The PON distributed intelligent computing task scheduling method according to claim 1, characterized in that, S3 includes the following steps: S31. Based on the allocation information issued by the optical line terminal, each gradient component in the gradient vector to be aggregated is assigned to a designated differential dual-wavelength channel, and a target optical power value is assigned to each gradient component in the gradient vector to be aggregated using a differential dual-wavelength mapping method. S32. Based on its own link attenuation coefficient, the target optical power value is pre-compensated on the optical network unit to obtain the compensated optical power value.

4. The PON distributed intelligent computing task scheduling method according to claim 3, characterized in that, The specific process of S32 includes: adjusting the optical network unit for the gradient component at the positive wavelength. The ratio of the target optical power value to its own link attenuation coefficient is used as the compensation optical power value corresponding to the positive wavelength; The optical network unit is configured with respect to the gradient component at the negative wavelength. The ratio of the target optical power value to its own link attenuation coefficient is used as the compensation optical power value corresponding to the negative wavelength.

5. The PON distributed intelligent computing task scheduling method according to claim 1, characterized in that, S4 includes the following steps: S41. Broadcast an aggregation trigger frame on the optical line terminal by controlling the wavelength. The aggregation trigger frame includes: specifying the transmission time and the time slot length. S42. All optical network units simultaneously transmit optical signals with a power equal to the compensated optical power value at a specified transmission time, the transmission time lasting for the duration of the time slot.

6. The PON distributed intelligent computing task scheduling method according to claim 1, characterized in that, S5 includes the following steps: S51. Through the optical line terminal, at the positive and negative wavelengths corresponding to each group of differential dual-wavelength channels, the optical signals sent by all optical network units that belong to the gradient components responsible for that channel are simultaneously received. S52. In each group of differential dual-wavelength channels, obtain the total transmit power of the optical signal of each optical network unit, and obtain the positive wavelength received optical power and negative wavelength received optical power of each group. S53. The photodetector PD at the optical line terminal converts the received optical power of the positive wavelength and the received optical power of the negative wavelength of each group into photocurrent, thereby obtaining the photocurrent of the positive wavelength and the photocurrent of the negative wavelength of each group.

7. The PON distributed intelligent computing task scheduling method according to claim 1, characterized in that, The process in S6 of statistically analyzing the maximum net photocurrent difference through the optical line terminal to adaptively update the mapping coefficients and broadcasting the result includes: A1. Calculate the maximum net photocurrent difference of each differential dual-wavelength channel in the current round of aggregation through the optical line terminal; A2. The ratio of the maximum net photocurrent difference to the saturation current threshold of the photodetector PD is taken as the current dynamic range utilization rate. A3. When the current dynamic range utilization is less than 0.5, the optical line terminal increases the mapping coefficient: ， in, For the new mapping coefficients, The mapping coefficients for the current round of aggregation. For the current dynamic range utilization, For adjustment coefficients, ; A4. When the current dynamic range utilization is greater than 0.9, the optical line terminal reduces the mapping coefficient: ， A5. The updated version is transmitted via optical line terminal. Broadcast to all optical network units in the next round of aggregation trigger frames.

8. The PON distributed intelligent computing task scheduling method according to claim 1, characterized in that, The process in S6 of updating the global model parameters based on the global gradient vector and sending them to each optical network unit to complete local synchronization and confirmation feedback includes: B1. Obtain the global model parameters for the current round through the optical line terminal; B2. At the optical line terminal, the global model parameters are updated using gradient descent based on the global gradient vector. B3. The optical line terminal broadcasts the updated global model parameters to all optical network units; B4. Each optical network unit receives the updated global model parameters and sends an acknowledgment frame to the optical line terminal.

Images