Communication system computing power scheduling method and device based on running complexity prediction

Abstract

This invention discloses a dynamic scheduling method for computing resources in communication systems based on runtime complexity prediction. First, the core operating parameters of the baseband processing unit are collected and preprocessed in real time, and then input into a constructed nonlinear communication computing complexity prediction model. The model outputs a predicted computing complexity value for a preset future time window. Next, based on the predicted value, a computing demand index is calculated and load levels are assigned, implementing a differentiated hierarchical computing power scheduling strategy. Simultaneously, the model is iteratively optimized online at fixed intervals using actual operating data, forming a closed-loop computing power scheduling mechanism. This invention achieves pre-emptive predictive scheduling of computing resources, with prediction errors controlled within 3%, significantly improving computing resource utilization and reducing system power consumption. It perfectly adapts to the low-latency, high-reliability computing power scheduling requirements of 5G / 6G multi-service convergence scenarios.

Description

Technical Field

[0001] This invention relates to the field of wireless communication and computing resource scheduling technology, and in particular to a dynamic scheduling method for computing resources of baseband processing units in 5G / 6G mobile communication systems based on operational complexity prediction. Background Technology

[0002] With the commercialization of 5G-Advanced and the in-depth research and development of 6G technology, mobile communication systems are evolving towards multi-service convergence, high computing power requirements, low latency requirements, and wide scenario adaptability. The baseband processing unit (BBU), as the core of the communication system's computing power, needs to simultaneously support diverse service types. URLLC services require microsecond-level latency and 99.999% reliability, eMBB services pursue high bandwidth and high speed, while mMTC services exhibit characteristics of massive connections, low speed, and low power consumption. The computing power requirements of different services can vary by more than two orders of magnitude. However, current BBU computing power resource configuration and scheduling schemes in communication systems are ill-suited to the development needs of next-generation communication technologies. The core shortcomings are concentrated in four points, and existing technologies lack systematic solutions: Existing solutions mostly adopt a static pre-configuration mode based on peak demand, which is prone to mismatch between computing power supply and demand. Insufficient computing power during peak hours leads to service failures, while a large number of computing units run idle during low-load periods, resulting in both resource waste and high power consumption. Conventional dynamic scheduling is a reactive approach with a scheduling lag window of 100ms-500ms, which cannot meet the microsecond-level low latency requirements of URLLC services. Computing power demand assessment mostly uses linear weighted models, which do not take into account the superlinear characteristics of baseband processing computing power consumption and ignore the nonlinear effects of modulation order and channel quality, resulting in large deviations in demand estimation and insufficient scheduling accuracy. At the same time, the scheduling model parameters are fixed offline and cannot adapt to changes in channel environment, service model, and equipment performance, leading to a decline in prediction accuracy and high manual maintenance costs. Existing industry improvements are only partial optimizations and do not propose a complete solution from the perspectives of pre-prediction of computing power, accurate nonlinear modeling, and full lifecycle adaptation. They cannot fundamentally solve the core problems of inaccurate computing power assessment, scheduling lag, low resource utilization, and excessive power consumption. Developing a computing power scheduling solution adapted to baseband characteristics has become a key technical bottleneck for the deployment of next-generation mobile communications. Summary of the Invention

[0003] The purpose of this invention is to overcome the aforementioned deficiencies of the prior art and provide a dynamic scheduling method for computing resources in communication systems based on runtime complexity prediction. This method is achieved through the construction of a nonlinear computing complexity prediction model, the design of a prediction-driven pre-scheduling mechanism, the formulation of a hierarchical low-power scheduling strategy, and an online iterative optimization mechanism for the model. This fundamentally solves the core problems of the prior art and achieves the following invention objectives: 1. To address the pain points of inaccurate computing power scheduling assessment and scheduling lag in the existing computing power scheduling, a nonlinear computing power assessment system that fits the characteristics of baseband processing is constructed, a forward predictive scheduling mechanism is built, the inherent defects of passive scheduling are overcome, and the fine-grained on-demand allocation of computing power resources is realized simultaneously, taking into account the stability of communication service transmission and resource utilization efficiency, and optimizing the overall operating power consumption of the system. 2. Establish a model lifecycle adaptive optimization mechanism to adapt to various dynamic scenario changes, ensure long-term scheduling accuracy, reduce system operation and maintenance costs without frequent manual intervention, and make the overall technical solution adaptable to multi-service converged communication scenarios, with good versatility and practicality, meeting the long-term stable operation requirements of communication systems.

[0004] Therefore, to achieve the above-mentioned objectives, the present invention adopts the following technical solution, which has the characteristics of clear steps, strong engineering feasibility, and wide adaptability. Specifically, it discloses a communication system computing power scheduling method based on runtime complexity prediction, characterized by including the following steps: S1. Real-time acquisition of core operating parameters of the baseband processing unit of the communication system; preprocessing of the acquired raw parameters to obtain standardized input parameters; the core operating parameters include at least the current number of online users U in the cell and the average channel quality indication of all online users. The current modulation order M of the system, and the rated maximum channel quality of the system. The computational complexity of the system at full load ; S2. Input the standardized input parameters into the pre-trained nonlinear communication computational complexity prediction model, and output the predicted system computational complexity value C within a future preset time window; the mapping function of the nonlinear communication computational complexity prediction model is: ; In the formula, a is the basic weight coefficient, b is the nonlinear power coefficient of the number of users, d is the channel compensation coefficient, and K is the modulation order exponential correction term. The initial values ​​of the above coefficients are obtained through offline training using historical system operation data. S3 Based on the predicted computational complexity C and the computational complexity at full load of the system. The computing system calculates the demand index I, and classifies the load level of the system within a future preset time window based on the value range of the computing demand index I. S4 executes a differentiated hierarchical computing power dynamic scheduling strategy based on the divided load levels, and dynamically adjusts the operating status of the computing units within the baseband processing unit. S5 uses a fixed update cycle to iteratively optimize the coefficients of the nonlinear communication computational complexity prediction model online using actual system operating data. The updated model is then used for the computational complexity prediction in the next cycle.

[0005] Preferably, the preprocessing of the original parameters in step S1 specifically includes: S11 performs 3-point moving average filtering and noise reduction on the same parameter in 3 consecutive acquisition cycles to eliminate random noise interference; S12 calculates the mean μ and standard deviation σ of the historical operating data of each parameter, and removes outlier sampling points that deviate from the historical mean by more than 3 times the standard deviation. S13 linearly normalizes the preprocessed parameters to the 0~1 range to obtain standardized input parameters.

[0006] Preferably, in step S2, the value range of the nonlinear power coefficient b for the number of users is 1.1~1.3; the value rule for the modulation order exponent correction term K is as follows: when the modulation order M < 6, K = 1.0; when the modulation order M ≥ 6, K increases linearly with the increase of M, and the linear growth formula is as follows: .

[0007] Preferably, in step S3, the formula for calculating the demand index I is as follows: The load level is divided into three levels, and the judgment rule is: I > 0.8 is a high load state, 0.3 ≤ I ≤ 0.8 is a medium load state, and I < 0.3 is a low load state.

[0008] Preferably, in step S4, when a high load state is determined, the scheduling strategy executed is as follows: enable all computing units within the baseband processing unit, allocate URLLC and eMBB core services to high-performance DSP core processing, divert mMTC non-core services to backup general-purpose computing core processing, and disable all hibernation mechanisms.

[0009] Preferably, in step S4, when the load is determined to be moderate, the scheduling strategy is as follows: a corresponding number of computing units are activated proportionally according to the computing demand index I, and the activation quantity formula is: ,in This represents the total number of computing units in the baseband processing unit, and the result is rounded to the nearest integer. Unused computing units enter a shallow sleep state, with the clock signal turned off and only power supply retained, and a 10ms fast wake-up mechanism is set.

[0010] Preferably, in step S4, when a low-load state is determined, the scheduling strategy is as follows: only 30% of the core computing units within the baseband processing unit are enabled to maintain the operation of basic communication services; the remaining computing units perform deep power-off hibernation, completely cutting off power supply and clock signals, while retaining the wake-up trigger mechanism.

[0011] Preferably, in step S5, the specific steps for online iterative optimization of the model are as follows: S51 extracts all actual operating data within a fixed update cycle, including standardized input parameters and the actual computational complexity of the system. ; S52 Using the model's predicted value C and The mean squared error is the loss function, and the formula for the loss function is: In the formula, N is the number of sampling points in the update period; This is the computational complexity of predicting the i-th sampling point; This represents the actual measured value of the system at the i-th sampling point; S53 employs the gradient descent algorithm, aiming to minimize the loss function, and fine-tunes the coefficients of a, b, d, and K of the model with small steps, the fine-tuning step size being 0.001~0.01; S54 updates the fine-tuned coefficients into the nonlinear prediction model, completing the online iterative optimization of the model.

[0012] Preferably, the core operating parameter acquisition cycle in step S1 is 100ms, the future preset time window in step S2 is 500ms, and the model fixed update cycle in step S5 is 1 hour.

[0013] Another aspect of the present invention is to provide a scheduling apparatus for the above-described method, characterized in that the apparatus is deployed in a baseband processing unit of a communication system, comprising: The parameter acquisition and preprocessing module is used to acquire the core operating parameters of the baseband processing unit in real time and preprocess the raw parameters to obtain the standardized input parameters. The nonlinear complexity prediction module has a built-in pre-trained nonlinear communication computation complexity prediction model, which is used to receive standardized input parameters and output the predicted value C of the system computation complexity within a future preset time window. The load level determination module is used to calculate the system's demand index I based on the predicted computational complexity value C, and to classify the load level according to the range of values ​​of the computational demand index I. The computing power dynamic scheduling module is used to execute differentiated hierarchical computing power dynamic scheduling strategies according to the divided load levels, and dynamically adjust the operating status of computing units within the baseband processing unit. The online iterative optimization module is used to perform online iterative optimization of the coefficients of the nonlinear communication computational complexity prediction model using actual system operating data at a fixed update cycle.

[0014] The core inventive concept of this invention is: taking the nonlinear computing power consumption characteristics of baseband processing in communication systems as the core, constructing a nonlinear computational complexity prediction model based on multi-dimensional operating parameters, realizing advance prediction of computing power demand through pre-prediction, and then formulating a hierarchical computing power scheduling strategy in combination with load level classification. At the same time, ensuring the prediction accuracy throughout the entire life cycle through online iterative optimization of the model, and finally realizing the closed-loop management of computing power resources of "prediction-scheduling-optimization".

[0015] The core innovation that distinguishes it from existing technologies lies in: 1. An innovative nonlinear computational complexity prediction model adapted to baseband processing characteristics was constructed: For the first time, the superlinear power term of the number of users, the exponential correction term of the modulation order, and the square root compensation term of the channel quality were integrated into the model, which accurately matched the nonlinear computing power consumption law under multi-user concurrency, high-order modulation, and weak coverage scenarios, and solved the problem of low evaluation accuracy of traditional linear models. 2. A prediction-driven, proactive dynamic scheduling mechanism for computing power is proposed: based on the predicted value of computing power complexity in future time windows, the expansion or reduction of computing power resources is completed in advance, changing from "post-event response" to "pre-event prediction", completely eliminating scheduling lag and adapting to the low latency requirements of URLLC. 3. A hierarchical hibernation and service diversion scheduling strategy based on load level was designed: the load status is divided into three levels and matched with differentiated computing power scheduling schemes. Combined with the "core / non-core service diversion" and "shallow / deep hibernation" mechanisms, the dual optimization of computing power utilization and system power consumption is achieved. 4. An online iterative optimization mechanism based on gradient descent was designed: the model coefficients are fine-tuned using actual operating data in a 1-hour cycle, so as to achieve adaptive matching of the model to the channel environment, service model and equipment aging, ensuring the prediction accuracy throughout the entire life cycle without manual intervention. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments or related technologies of this specification, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the computing power scheduling method for communication systems based on runtime complexity prediction according to the present invention. Figure 2 This is a flowchart illustrating the preprocessing of the collected raw parameters according to the present invention. Figure 3 The following is a flowchart illustrating the specific process of online iterative optimization and updating of the prediction model of this invention; Figure 4 This is a functional diagram of the dynamic scheduling structure for computing resources in a communication system based on runtime complexity prediction.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. Those skilled in the art should understand that the following embodiments are only used to illustrate the implementation process and application effects of the present invention, and are not intended to limit the scope of protection of the present invention. Minor adjustments and optimizations made to the steps, calculation methods, and algorithms of the present invention without departing from the core technical concept of the present invention are all within the scope of protection of the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that the terms "comprising" and "having," and any variations thereof, in the embodiments and drawings of this specification are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0021] Example 1: like Figure 1 As shown, to achieve the above-mentioned objectives, this invention adopts the following technical solution, which features clear steps, strong engineering feasibility, and wide adaptability. Specifically, it includes five core steps: real-time acquisition and preprocessing of system operating parameters, prediction of nonlinear communication computational complexity, calculation of computational demand index and determination of load level, dynamic scheduling of computing units based on load level, and online iterative updating of the prediction model. These steps form a closed loop, enabling dynamic and refined scheduling of computing resources, as detailed below: S1: Real-time acquisition and preprocessing of system operating parameters With a fixed acquisition period of 100ms, the main control module of the communication system's baseband processing unit (BBU) collects core operating parameters reflecting the system's operating status and service load in real time. This parameter set provides multi-dimensional input for computing complexity prediction and covers key influencing factors from the user side, channel side, and system side, specifically including: The current number of online users in the community, U, reflects the service load characteristics of massive connections. Channel Quality Indicator (CQI) for each online user: Reflects the propagation characteristics of the wireless channel; the average CQI of all users is subsequently calculated. ; The modulation order M currently used in the system reflects the intensity of the transmission rate and the computing power required for baseband processing, corresponding to mainstream modulation methods such as QPSK (M=2), 16QAM (M=4), 64QAM (M=6), and 256QAM (M=8); System rated maximum channel quality The optimal channel quality reference value for the system design is fixed at 15 (compliant with 3GPP protocol standards). System full load computational complexity The computing power threshold when all computing units of the BBU are running at full load is determined by the hardware configuration (such as the number of DSP cores and the computing power per core).

[0022] The collected raw parameters undergo three levels of preprocessing to eliminate the influence of noise and outliers, ensuring the validity and standardization of the parameters input to the model. Figure 2 As shown, the specific processing flow is as follows: S11 performs 3-point moving average filtering and noise reduction on the same parameter in 3 consecutive acquisition cycles to eliminate random noise interference; The moving average of the same parameter over three consecutive acquisition cycles is calculated using the following formula: , in The preprocessed parameters for the nth cycle are used to eliminate wireless channel fluctuations and random noise from the acquisition module. S12 Calculates the mean μ and standard deviation σ of the historical operating data for each parameter, and removes outlier sampling points that deviate from the historical mean by more than 3 times the standard deviation. Calculate the mean μ and standard deviation σ of the historical data for each parameter, and remove parameter points that deviate from the historical mean by more than 3 times the standard deviation (i.e., |X-μ|>3σ) to avoid interference from sudden abnormal data on model prediction; S13 linearly normalizes the preprocessed parameters to the 0~1 range, eliminating the differences in the dimensions of different parameters and obtaining standardized input parameters.

[0023] All preprocessed parameters are linearly normalized to the 0~1 interval, as shown in the formula. ,in The minimum and maximum values ​​of the parameters are used to eliminate the influence of differences in the dimensions of different parameters on the model, thus obtaining standardized input parameters.

[0024] S2: Prediction of computational complexity in nonlinear communication The standardized input parameters obtained in step S1 are input into the pre-trained nonlinear communication computational complexity prediction model. The model outputs the predicted system computational complexity value C (unit: GOPS) within a future 500ms time window. This time window matches the acquisition cycle and scheduling execution delay in step 1, realizing the advance prediction of computing power demand and meeting the low latency requirements of URLLC.

[0025] One of the core innovations of this invention is the construction of this nonlinear model, whose core mapping function integrates a user number superlinear term, a modulation order correction term, and a channel quality compensation term, accurately matching the nonlinear computing power consumption characteristics of baseband processing. The specific formula is as follows: The definitions and value rules of each parameter in the formula are as follows: The average channel quality indicator (CQI) of all online users is calculated from the CQI of each user and reflects the overall channel quality of the system. a: Basic weight coefficient, reflecting the fundamental impact of the number of users and modulation order on computational complexity. The initial value is obtained through offline training using historical operating data; that is, using historical multi-scenario measured operating data as samples, with the goal of minimizing prediction error, it is solved through data fitting. A typical value is, for example, 0.08.

[0026] b: The nonlinear power coefficient of the number of users is a core innovative parameter that adapts to the superlinear growth characteristics of baseband processing computing power consumption in multi-user concurrent scenarios. The value range is 1.1~1.3, and the initial value is obtained through offline training using historical running data. d: Channel compensation coefficient, reflecting the compensatory impact of channel quality degradation on computational complexity. The worse the channel quality, the higher the computational cost of channel estimation and equalization operations in baseband processing. Preferably, based on historical measured operating data of the system under different channel qualities (good / medium / poor), with the goal of minimizing the error between the model's predicted computational complexity and the system's actual computational complexity, the coefficient d is solved using conventional data fitting / least squares / gradient descent methods in this field to obtain an initial value suitable for this system. The value range is 2.0 ~ 3.0, with a typical value of 2.5.

[0027] K: Exponential correction term, a core innovative parameter, adapted to the sharp increase in computing power under high-order modulation scenarios. Its value ranges from 0.9 to 1.3, with the following rules: when the modulation order M < 6 (QPSK, 16QAM), K = 1.0; when the modulation order M ≥ 6 (64QAM, 256QAM), K increases linearly with increasing M, and the linear growth formula is as follows: It accurately matches the changing computing power requirements of baseband processing under high-order modulation.

[0028] The initial values ​​of a, b, d, and K of the model were obtained through offline training using six months of historical full-scenario operation data from the communication system. The training dataset covers all typical scenarios, including peak / off-peak load, good / poor channel quality, and low / high-order modulation, ensuring the generalization ability of the initial model.

[0029] S3: Calculation of Demand Index and Determination of Load Level Based on the predicted computational complexity C obtained in step S2, the system computational demand index I is calculated. This index is a dimensionless parameter that reflects the proportion of the system's future computational demand to its full-load computational capacity within a future time window. The calculation formula is as follows: In the formula The computational complexity at full system load is determined by the BBU hardware configuration and is a fixed value.

[0030] Based on the range of the calculated demand index I, the load status of the system within the next 500ms time window is divided into three levels. The division rule takes into account both fine-grained scheduling and engineering feasibility. The specific judgment rule is as follows: High load state: I > 0.8. At this time, the system's computing power demand is close to full load. It is necessary to fully activate computing resources and perform business diversion to ensure the operation of core businesses. Medium load state: 0.3≤I≤0.8. At this time, the system computing power demand is moderate, and computing power resources can be activated proportionally, while the remaining resources enter shallow sleep, so as to realize the on-demand allocation of computing power. Low load state: I < 0.3. At this time, the system computing power requirement is extremely low. Only the core computing power resources need to be activated, and the remaining resources enter deep power-off hibernation to minimize power consumption.

[0031] S4: Tiered computing power dynamic scheduling strategy based on load level Based on the load level determined in step S3, a differentiated hierarchical computing power scheduling strategy is implemented to dynamically adjust the operating status of each computing unit (DSP core, general computing core) within the BBU. Simultaneously, combined with a core / non-core service offloading mechanism, this strategy achieves dual optimization of computing resource utilization and system power consumption while ensuring communication service quality. This strategy is another core innovation of the invention. The specific scheduling scheme is as follows: S41 High Load State (I > 0.8) Scheduling Strategy Activate all computing units within the BBU (all DSP cores and general-purpose computing cores) to maximize the release of computing resources and meet the computing power demands of the surge in business. Implement core / non-core business offloading: allocate core businesses such as URLLC and eMBB to high-performance DSP cores for processing to ensure their low latency and high reliability requirements; offload non-core businesses such as mMTC to backup general-purpose computing cores for processing to avoid non-core businesses occupying core computing resources and causing core business latency to exceed the standard. All computing units are running at full load, with all hibernation mechanisms disabled until the load level drops to medium or low.

[0032] S42 Medium Load Condition (0.3≤I≤0.8) Scheduling Strategy Activate computing units proportionally: Based on the value of the computing demand index I, activate a corresponding number of computing units proportionally. The activation formula is as follows: ,in This represents the total number of computing units in the BBU, and the result is rounded to the nearest integer, enabling fine-grained allocation of computing resources on demand. Unused computing units enter a shallow sleep state: the clock signal of the computing unit is turned off, only the power supply is retained. In this sleep state, the computing unit can be quickly woken up within 10ms to meet the rapid response requirements of sudden increases in computing power demand. All services are processed by the enabled computing units, eliminating the need for service splitting and ensuring the simplicity of scheduling.

[0033] S43 Low-load state (I < 0.3) scheduling strategy Only 30% of the core computing units within the BBU are activated to maintain low-rate basic communication services such as mMTC and basic voice / data, in order to meet the minimum computing power requirements. The remaining 70% of computing units perform deep power-off hibernation: completely cut off the power supply and clock signal of the computing units, and completely eliminate their idle power consumption. This is the core measure to reduce the overall power consumption of the system. The computing unit in deep sleep retains the wake-up trigger mechanism, which can quickly power on and start running within a preset delay when the load level rises to medium or high load.

[0034] S5: Online Iterative Optimization and Update of Predictive Model To address the issues of existing models lacking adaptive capabilities and exhibiting declining prediction accuracy over time, this invention employs an online iterative optimization mechanism, a core innovation of the invention. With a fixed update cycle of one hour, the coefficients of the nonlinear model are adaptively fine-tuned using actual operational data, ensuring the long-term stability of the model's prediction accuracy without requiring manual intervention. Figure 3 As shown, the specific steps are as follows: S51 extracts all actual operational data within a fixed update cycle at each interval: Actual data extraction: Every hour, extract all actual operational data for the update cycle, including: standardized input parameters (U, ...). (M, etc.) actual computational complexity of the system ; S52 uses the model's predicted value C and The mean squared error is the loss function used to quantify the model's prediction error; Loss function calculation: The model's predicted value C is compared with the actual value. The mean squared error (MSE) is the loss function, which quantifies the model's prediction accuracy. The formula for the loss function is: In the formula, N is the number of sampling points within the update period. The predicted value for the i-th sampling point. This represents the actual value of the i-th sampling point; S53 employs the gradient descent algorithm, aiming to minimize the loss function, and fine-tunes the coefficients of a, b, d, and K of the model with small step sizes, ranging from 0.001 to 0.01. Model coefficient fine-tuning: Using the gradient descent algorithm, with the goal of minimizing the loss function, the coefficients of a, b, d, and K of the model are fine-tuned with small steps. The fine-tuning step size is set to 0.001~0.01 to avoid sudden changes in coefficients that could cause a sharp drop in model prediction accuracy. S54 updates the fine-tuned coefficients into the nonlinear prediction model, completing the online iterative optimization of the model; Update model parameters: Update the fine-tuned coefficients into the nonlinear prediction model to complete the online iterative optimization of the model. The new model is then immediately put into the next cycle of computational complexity prediction.

[0035] This iterative mechanism enables the model to adaptively match changes in the channel environment (such as CQI changes caused by rain, snow, or obstruction), service model iterations (such as changes in the proportion of service types at different times), and equipment performance aging (such as the attenuation of baseband chip computing power), ensuring that the error between the model prediction and the actual value is always controlled within 3% throughout the entire life cycle of the communication system.

[0036] Example 2 Through the above-mentioned technical solutions, this invention achieves closed-loop scheduling of computing resources in communication systems from four dimensions: nonlinear modeling, pre-scheduling, hierarchical allocation, and adaptive optimization. Compared with existing static configuration and passive response scheduling schemes, it has outstanding substantial technical advantages and significant beneficial effects.

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed embodiments of this invention are described using the 5G macro base station baseband processing unit (BBU) as a specific application scenario. These embodiments are only used to explain this invention and are not intended to limit the scope of protection of this invention. All equivalent substitutions based on the concept of this invention fall within the scope of protection of this invention.

[0038] Basic configuration This embodiment uses the BBU of a 5G macro base station as the application object. This BBU is the mainstream configuration of operators, and its hardware and system parameters are as follows: Computing unit configuration: 12 high-performance baseband processing DSP cores that can be started and stopped independently, and 2 spare general-purpose computing cores; Single DSP core computing power: 100 GOPS; Single general-purpose computing core computing power: 50 GOPS; System full load computational complexity The total computing power of 12 DSP cores running at full load, i.e. GOPS; System rated maximum channel quality (Compliant with 3GPP protocol standards); Data acquisition period: 100ms, prediction time window: 500ms, model update period: 1 hour; Business types: Covering multiple business integration scenarios including eMBB (high-definition video, high-speed download), URLLC (industrial control, telemedicine), and mMTC (smart meters, IoT sensors).

[0039] Using six months of historical full-scenario operation data from this 5G macro base station (covering peak / off-peak load, good / poor channel quality, and low / high-order modulation), the initial coefficients of the nonlinear computational complexity prediction model were obtained through offline training. , (User count is exponentially linear, adapting to multi-user concurrency characteristics). , For the exponential correction term (value rules: hour , hour ).

[0040] S1: Real-time acquisition and preprocessing of system operating parameters The raw values ​​of the following core operating parameters are collected through the BBU main control unit at a sampling period of 100ms: The current number of online users in the community is U=600; CQI values ​​collected by each user: mainly distributed between 10 and 14, the average value was calculated. =12; The current modulation order of the system is M=6 (64QAM, which is the mainstream modulation method for eMBB services). System rated maximum channel quality The computational complexity of the system at full load =1200 GOPS. Three-stage preprocessing is performed on the above original parameters: 3-point moving average filtering: This parameter is the collected values ​​of 598, 600 and 602 for 3 consecutive periods, and the filtered value is U=600; Three consecutive periods are 11.8, 12, and 12.2. After filtering... =12; Outlier removal: All the above parameters are within 3 standard deviations of the historical mean, so no outliers are removed. Parameter normalization: Normalize U, M is normalized to the 0~1 interval to obtain standardized input parameters, which are used for model prediction.

[0041] S2: Prediction of computational complexity in nonlinear communication The preprocessed, standardized input parameters are input into the pre-trained nonlinear model, and the exponential correction term is first determined. : Due to modulation order According to the value selection rules, .

[0042] Substituting all known parameters into the model's core mapping function, we can accurately calculate the predicted system computational complexity for the next 500ms. : ; Calculate the value of each item step by step: calculate : ; Calculate the number of users - modulation order term: ; Calculate the channel quality compensation term : Calculate the predicted value : GOPS (Due to the compatibility between hardware computing power and model coefficients, the actual calculation results are approximate values, which meet the requirements of engineering implementation).

[0043] S3: Calculation of Demand Index and Determination of Load Level According to the formula for calculating the demand index, calculate... : ; According to the load level determination rules: It was determined that the 5G macro base station would be under high load within the next 500ms.

[0044] S4: Dynamic scheduling and execution based on high-load computing power Based on the scheduling policy for high load conditions, perform the following operations: DSP cores are activated proportionally to the calculated demand index: Total number of DSP cores Number of activations One (the result is an integer); All 12 high-performance DSP cores are enabled at full capacity (due to insufficient core count, the computing power demand cannot be fully met proportionally, so all available cores are enabled); the current mixed eMBB, URLLC, and mMTC services are handled with full force, prioritizing the latency and bandwidth requirements of various services under high load scenarios to avoid service blockage; S5: Online Iterative Optimization and Update of Predictive Model With an update cycle of 1 hour, extract all actual operational data within that cycle, including: Standardized input parameters for each sampling point ( , , wait); Actual computational complexity of the system (Collected via BBU computing power monitoring module).

[0045] In this embodiment, the mean of the model's predicted values ​​within this update cycle is 1544 GOPS (the average of all predicted values ​​in 1 hour), and the mean of the actual values ​​is 1542 GOPS (the measured average of the actual computing power in 1 hour). The loss function MSE is calculated as follows: ; (Note: 1 hour = 36,000 100ms data acquisition cycles) With the goal of minimizing the loss function, the gradient descent algorithm is used to fine-tune the model coefficients with small steps (fine-tuning step size = 0.001). The coefficients after fine-tuning are: , , , The value selection rules remain unchanged. The fine-tuned coefficients are updated into the model to complete this iteration of optimization, and the new model is immediately put into the prediction of the next cycle.

[0046] The scheduling method of this invention was subjected to a 7-day live network comparison test on a 5G macro base station, and compared with the traditional static configuration method and the traditional passive dynamic scheduling method. The test indicators included computing resource utilization, overall system power consumption, service interruption probability, and model prediction error. The test results are shown in the table below: Table 1. Comparison of the effects of Embodiment 2 of the present invention with the prior art. The test results show that the scheduling method of the present invention has achieved significant improvements in all core indicators, fully achieving the purpose of the invention. The solution is stable and fault-free, with strong engineering feasibility. It can be directly deployed to commercial 5G base stations and can also seamlessly adapt to the computing power scheduling requirements of 6G base stations.

[0047] Example 3 This embodiment targets ultra-high throughput service scenarios such as eMBB high-definition video and high-speed download of large files. In this scenario, the 256QAM ultra-high-order modulation method is used, and the computing power requirements for baseband demodulation, decoding, and soft bit processing increase exponentially. Traditional models do not consider the characteristics of the sharp increase in computing power of high-order modulation, which can easily lead to problems such as insufficient computing power supply, failure to meet transmission rate standards, and degraded user experience.

[0048] The hardware and system basic parameters in this embodiment are consistent with the original basic embodiment. The scenario-based input parameters are as follows: Current number of online users in the cell U=300; average channel quality indication of all online users (Excellent channel conditions, compatible with 256QAM modulation); The system currently uses a modulation order of M=8, corresponding to a 256QAM ultra-high-order modulation scheme; The system's rated maximum channel quality The computational complexity of the system at full load =1200 GOPS; the initial coefficients of the nonlinear prediction model are a=0.08, b=1.2, and d=2.5.

[0049] First, determine the value of the exponential correction term K: Since the modulation order M = 8 ≥ 6, according to the value rules set by this invention, .

[0050] Substituting all parameters into the core mapping function of the nonlinear computational complexity prediction model of this invention, the predicted value C of the system computational complexity within the next 500ms time window is calculated: Calculate each item step by step: User count superlinear term: ; Basic computing power items: ; Channel quality compensation terms: ; Final predicted value: GOPS; Calculate the demand index based on the predicted value calculation system: According to the load level determination rule, 0.3≤I≤0.8, the system is determined to be in a medium load state within the next 500ms.

[0051] The corresponding medium load scheduling strategy is implemented: DSP cores are activated exponentially according to computing demand, and the number of activated cores is [number missing]. One; the 6 DSP cores that are enabled process all services, while the other 6 DSP cores enter a shallow sleep state, turn off the clock signal and only keep the power supply, and set a 10ms fast wake-up mechanism.

[0052] The live network test results of this embodiment are as follows: The error between the predicted value of the nonlinear model of this invention and the actual computing power consumption of the system is 1.8%, which is much lower than the 24.5% prediction error of the traditional linear model. Through the precise adaptation of the modulation order exponent correction term K, the service transmission rate in the 256QAM ultra-high order modulation scenario reaches 99.98%, which is much higher than the 92.3% of the traditional solution. The system computing power resource utilization rate reaches 76%, which is 22 percentage points higher than the traditional passive scheduling solution. This verifies the rationality of the K value selection rules and the medium load threshold setting of 0.3-0.8 in this invention. It is an exclusive innovative design for the computing power characteristics of high-order modulation baseband, and is not a conventional technical means in this field.

[0053] Table 2 Comparison of the effects of Embodiment 3 of the present invention with the prior art Example 4 This embodiment targets massive connectivity scenarios such as mMTC smart meters and IoT sensors, where the number of online users exceeds 1,000. The computing power consumption of baseband processing for multi-user scheduling, interference cancellation, and random access processing increases superlinearly with the number of users. Traditional linear models (number of users to the power of one) are completely unable to adapt to this characteristic, and are prone to problems such as severely insufficient computing power estimation, user access failure, and data reporting timeout.

[0054] The hardware and system basic parameters in this embodiment are consistent with the original basic embodiment. The scenario-based input parameters are as follows: The current number of online users in the cell is U=1200, which meets the criteria for a massive connection scenario where U>1000; the average channel quality indicator of all online users... The system currently uses a modulation order of M=2, corresponding to QPSK modulation (the mainstream modulation mode suitable for massive connection scenarios); the system's rated maximum channel quality... The computational complexity of the system at full load The initial coefficients of the nonlinear prediction model are a=0.08, b=1.2, and d=2.5.

[0055] First, determine the value of the exponential correction term K: Since the modulation order M=2<6, according to the value rule set by this invention, K=1.0.

[0056] Substituting all parameters into the core mapping function of the nonlinear computational complexity prediction model of this invention, the predicted value C of the system computational complexity within the next 500ms time window is calculated: Calculate each item step by step: User count superlinear term: ; Basic computing power items: ; Channel quality compensation terms: ; Final predicted value: GOPS; Calculate the demand index based on the predicted value calculation system: According to the load level determination rule, 0.3≤I≤0.8, the system is determined to be in a medium load state within the next 500ms.

[0057] The corresponding medium load scheduling strategy is implemented: DSP cores are activated exponentially according to computing demand, and the number of activated cores is [number missing]. Eight DSP cores are activated to process all IoT access services, while the remaining four DSP cores enter a shallow sleep state.

[0058] The live network test results of this embodiment are as follows: The error between the predicted value and the actual computing power consumption of the nonlinear model of this invention is 2.3%, which is much lower than the 38.2% prediction error of the traditional linear model. Through the precise adaptation of the superlinear power term of the number of users, the user access success rate in the massive connection scenario reaches 99.99%, which is much higher than the 96.1% of the traditional solution. The power coefficient value of 1.1-1.3 in this invention perfectly matches the superlinear growth law of baseband computing power in multi-user concurrent scenarios. If the traditional linear model with b=1 is used, the predicted value is only 558.71 GOPS, which deviates from the actual value by more than 30%, and cannot meet the business needs at all. This fully proves the creativity and rationality of the coefficient value range, which is not a conventional choice for those skilled in the art.

[0059] Table 3 Comparison of the effects of Embodiment 4 of the present invention with the prior art Example 5 This embodiment targets a low-traffic scenario during the early morning hours. In this scenario, the cell only carries basic voice calls and a small amount of low-rate services such as mMTC smart meter / sensor IoT reporting. The number of online users is extremely small, and the computing power requirements for baseband demodulation, decoding, and soft bit processing are very low. Traditional static configuration methods still fully activate computing units, resulting in a large number of computing units running idle, with system idle power consumption exceeding 60%, leading to serious resource waste. Traditional passive dynamic scheduling methods do not design a dedicated sleep mechanism for low-load scenarios, resulting in scheduling lag and inability to achieve precise control of deep sleep, with limited power optimization effects, failing to meet the energy-saving requirements of low-load scenarios.

[0060] The hardware and system basic parameters in this embodiment are consistent with the original basic embodiment. The scenario-based input parameters are as follows: The current number of online users in the cell is U=200 (the number of users decreases significantly during off-peak hours, consistent with low load characteristics); the average channel quality indicator for all online users. (Wireless interference is minimal in the early morning, and channel conditions are excellent); the system currently uses modulation order M=4, corresponding to 16QAM modulation (the mainstream adaptation modulation mode for low-rate basic services); the system's rated maximum channel quality... The computational complexity of the system at full load The initial coefficients of the nonlinear prediction model are a=0.08, b=1.2, and d=2.5.

[0061] First, determine the value of the exponential correction term K: Since the modulation order M=4<6, according to the value rule set by this invention, K=1.0.

[0062] Substituting all parameters into the core mapping function of the nonlinear computational complexity prediction model of this invention, the predicted value C of the system computational complexity within the next 500ms time window is calculated: Calculate each item step by step: User count superlinear term: ; Basic computing power items: ; Channel quality compensation terms: ; Final predicted value: ; Calculate the demand index based on the predicted value calculation system: According to the load level determination rule, if I < 0.3, the system is determined to be in a low load state within the next 500ms.

[0063] The corresponding low-load state scheduling strategy is implemented as follows: DSP cores are activated according to the exponential ratio of computing demand, with the number of activated cores being N = 12 × 0.162 ≈ 2. The two activated DSP cores only handle basic voice and mMTC low-rate reporting services to meet the minimum communication service requirements. The remaining 10 DSP cores enter a deep power-off sleep state, completely cutting off the clock signal and power supply, and completely eliminating idle power consumption. The two backup general-purpose computing cores remain in a power-off state and are not activated, further reducing the system's operating energy consumption.

[0064] The live network test results of this embodiment are as follows: The error between the predicted value and the actual computing power consumption of the nonlinear model in this invention is 1.2%, which is far lower than the 32.7% prediction error of the traditional linear model. Through the design of a low-load dedicated deep sleep mechanism, the system's idle power consumption is only 560W, which is 51.6% lower than the traditional static configuration method and 42.9% lower than the traditional passive scheduling method. The basic service disconnection rate is controlled below 0.01%, which is far lower than the 0.07% of the traditional solution. The sleep strategy and superlinear power term adaptation design in this invention for low-load scenarios accurately match the extremely low growth characteristics of baseband computing power under low business volume scenarios. If the traditional linear model with b=1 is used, the predicted value is only 106.67 GOPS, which deviates from the actual value by more than 45%, and cannot support the energy saving and service guarantee requirements of low-load scenarios. This fully demonstrates the creativity and rationality of the low-load scheduling strategy and coefficient values, which are not the conventional choices of those skilled in the art.

[0065] Table 4 Comparison of the effects of Embodiment 5 of the present invention with the prior art Example 6: A Dynamic Scheduling Device for Computational Resources in Communication Systems Based on Operational Complexity Prediction like Figure 4 As shown, in order to realize the above-mentioned dynamic scheduling method for communication system computing resources based on runtime complexity prediction, the present invention also provides a corresponding scheduling device. This device is deployed in the baseband processing unit of the communication system. Each functional module corresponds to each step of the scheduling method and works in coordination to ensure the stability and efficiency of the scheduling process.

[0066] The parameter acquisition and preprocessing module uses a high-precision data acquisition chip to capture core operating parameters such as the number of online users and channel quality indicators of the baseband processing unit in real time. Through built-in filtering, outlier removal and normalization algorithms, it completes the preprocessing of the original parameters and outputs standardized parameters that meet the model input requirements.

[0067] The nonlinear complexity prediction module integrates a dedicated computing chip and has a pre-trained nonlinear communication computational complexity prediction model built in. It can quickly receive standardized input parameters and complete the calculation, and output the computational complexity prediction value for a future preset time window in real time.

[0068] The load level determination module calculates the computing demand index based on the predicted value and automatically classifies the load level according to the preset determination rules, providing a basis for the execution of subsequent scheduling strategies. The computing power dynamic scheduling module is directly linked with the computing unit of the baseband processing unit, and outputs corresponding scheduling instructions according to the load level to realize the dynamic adjustment of the computing unit's operating status, reasonable diversion of services, and flexible switching of the hibernation mechanism.

[0069] The online model iteration and optimization module periodically extracts actual system operation data and fine-tunes and updates the coefficients of the prediction model through the built-in gradient descent optimization algorithm to ensure that the model prediction accuracy is always maintained at a high level. The modules work together to form a complete closed loop of computing power scheduling.

[0070] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0071] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for scheduling computing power in a communication system based on runtime complexity prediction, characterized in that, Includes the following steps: S1. Real-time acquisition of core operating parameters of the baseband processing unit of the communication system; preprocessing of the acquired raw parameters to obtain standardized input parameters; the core operating parameters include at least the current number of online users U in the cell and the average channel quality indication of all online users. The current modulation order M of the system, and the rated maximum channel quality of the system. System full-load computational complexity ; S2. Input the standardized input parameters into the pre-trained nonlinear communication computational complexity prediction model, and output the predicted system computational complexity value C within a future preset time window; the mapping function of the nonlinear communication computational complexity prediction model is: In the formula, a is the basic weight coefficient, b is the nonlinear power coefficient of the number of users, d is the channel compensation coefficient, and K is the modulation order exponential correction term. The initial values ​​of the above coefficients are obtained through offline training using historical system operation data. S3 Based on the predicted computational complexity C and the computational complexity at full load of the system. The computing system calculates the demand index I, and classifies the load level of the system within a future preset time window based on the value range of the computing demand index I. S4 executes a differentiated hierarchical computing power dynamic scheduling strategy based on the divided load levels, and dynamically adjusts the operating status of the computing units within the baseband processing unit. S5 uses a fixed update cycle to iteratively optimize the coefficients of the nonlinear communication computational complexity prediction model online using actual system operating data. The updated model is then used for the computational complexity prediction in the next cycle.

2. The method according to claim 1, characterized in that, The preprocessing of the original parameters in step S1 specifically includes: S11 performs 3-point moving average filtering and noise reduction on the same parameter in 3 consecutive acquisition cycles to eliminate random noise interference; S12 calculates the mean μ and standard deviation σ of the historical operating data of each parameter, and removes outlier sampling points that deviate from the historical mean by more than 3 times the standard deviation. S13 linearly normalizes the preprocessed parameters to the 0~1 range to obtain standardized input parameters.

3. The method according to claim 1, characterized in that, In step S2, the value range of the nonlinear power coefficient b for the number of users is 1.1 to 1.3; the value rule for the modulation order exponent correction term K is as follows: when the modulation order M < 6, K = 1.0; when the modulation order M ≥ 6, K increases linearly with the increase of M, and the linear growth formula is as follows: .

4. The method according to claim 1, characterized in that, In step S3, the formula for calculating the demand index I is as follows: The load level is divided into three levels, and the judgment rule is: I > 0.8 is a high load state, 0.3 ≤ I ≤ 0.8 is a medium load state, and I < 0.3 is a low load state.

5. The method according to claim 4, characterized in that, In step S4, when a high load state is determined, the scheduling strategy executed is as follows: enable all computing units within the baseband processing unit, allocate URLLC and eMBB core services to high-performance DSP cores for processing, divert mMTC non-core services to backup general-purpose computing cores for processing, and disable all hibernation mechanisms.

6. The method according to claim 4, characterized in that, In step S4, when the load is determined to be moderate, the scheduling strategy is as follows: A corresponding number of computing units are activated proportionally according to the computing demand index I, and the activation quantity formula is: ,in This represents the total number of baseband processing units; the result is rounded to the nearest integer. Unused computing units enter a shallow sleep state, turning off the clock signal while retaining only power supply, and are set to a 10ms fast wake-up mechanism.

7. The method according to claim 4, characterized in that, In step S4, when the system is determined to be in a low-load state, the scheduling strategy is as follows: only 30% of the core computing units within the baseband processing unit are enabled to maintain the operation of basic communication services; the remaining computing units are put into deep power-off hibernation, completely cutting off power supply and clock signals, while retaining the wake-up trigger mechanism.

8. The method according to claim 1, characterized in that, In step S5, the specific steps for online iterative optimization of the model are as follows: S51 extracts all actual operating data within a fixed update cycle, including standardized input parameters and the actual computational complexity of the system. ; S52 Using the model's predicted value C and The mean squared error is the loss function, and the formula for the loss function is: In the formula, N is the number of sampling points in the update period; This is the computational complexity of predicting the i-th sampling point; Let be the actual measured value of the system at the i-th sampling point; S53 employs the gradient descent algorithm, aiming to minimize the loss function, and fine-tunes the coefficients of a, b, d, and K of the model with small steps, the fine-tuning step size being 0.001~0.01; S54 updates the fine-tuned coefficients into the nonlinear prediction model, completing the online iterative optimization of the model.

9. The method according to claim 1, characterized in that, The core operating parameter acquisition cycle in step S1 is 100ms, the future preset time window in step S2 is 500ms, and the model fixed update cycle in step S5 is 1 hour.

10. A scheduling apparatus for the method according to any one of claims 1-9, characterized in that, The device is deployed in the baseband processing unit of the communication system and includes: The parameter acquisition and preprocessing module is used to acquire the core operating parameters of the baseband processing unit in real time and preprocess the raw parameters to obtain the standardized input parameters. The nonlinear complexity prediction module has a built-in pre-trained nonlinear communication computation complexity prediction model, which is used to receive standardized input parameters and output the predicted value C of the system computation complexity within a future preset time window. The load level determination module is used to calculate the system's demand index I based on the predicted computational complexity value C, and to classify the load level according to the range of values ​​of the computational demand index I. The computing power dynamic scheduling module is used to execute differentiated hierarchical computing power dynamic scheduling strategies according to the divided load levels, and dynamically adjust the operating status of computing units within the baseband processing unit. The online iterative optimization module is used to perform online iterative optimization of the coefficients of the nonlinear communication computational complexity prediction model using actual system operating data at a fixed update cycle.

Images