Multi-modal based computing power scheduling method and device, server and storage medium

Abstract

The application discloses a kind of based on multi-modal computing power scheduling method, device, server and storage medium, it is related to computing power distribution technical field, comprising: receiving multi-modal computing task and parsing task description file while obtaining the available computing resource snapshot of each computing node currently;According to the preset static performance library, construct decision variable set, establish multi-objective optimization task, utilize the improved NSGA-III algorithm to solve multi-objective optimization task, form optimal decision variable set, and according to the preset business strategy carries out computing power scheduling decision.The improved algorithm generates optimal decision variable set to improve the accuracy of computing power scheduling decision, while introducing static performance library and the current resource utilization of computing node for performance correction;Through the collaborative optimization of multi-objective, not only can ensure service quality, but also can reduce energy consumption, improve the utilization efficiency of computing resource, can effectively balance service quality QoS and energy consumption and other conflicting indicators.

Description

Technical Field

[0001] This invention relates to the field of computing power allocation technology, and in particular to a multimodal computing power scheduling method, device, server and storage medium. Background Technology

[0002] With the rapid development of artificial intelligence technology, the demand for computing resources for multimodal inference and training tasks, including those involving images, text, and speech, is becoming increasingly diverse and complex. Computing platforms, such as multi-cloud and hybrid cloud architectures, exhibit heterogeneous (e.g., NPUs, GPUs) and distributed computing resources. Furthermore, different users, based on their own needs and environments, often have highly specific and differentiated requirements for task execution latency, accuracy, frame rate, and other QoS (Quality of Service) aspects, as well as energy consumption costs.

[0003] Because heterogeneous and distributed computing resources differ in terms of accuracy, power consumption, and latency, traditional scheduling schemes can often only optimize for a single objective. Furthermore, there are often performance conflicts between objectives such as accuracy, power consumption, and latency. For example, ensuring QoS often ignores energy consumption, resulting in high energy consumption. If energy consumption is reduced, QoS may not be met, making it difficult to effectively balance the trade-offs and provide users with suitable and differentiated computing resources. Summary of the Invention

[0004] This invention provides a multimodal computing power scheduling method, device, server, and storage medium to solve the technical problem of difficulty in comprehensively optimizing QoS quality and energy consumption and accurately scheduling computing power resources.

[0005] In a first aspect, embodiments of the present invention provide a multimodal computing power scheduling method, comprising:

[0006] S101 receives multimodal computing tasks, obtains service quality requirement information by parsing the task description file, and obtains the current available computing resource information of each computing node to form a snapshot of available computing resources;

[0007] S102, based on the preset static performance library, construct a set of decision variables using Path to represent the computation path, Batch to represent the batch size, Kernel to represent the approximate computation kernel state, and Freq to represent the frequency level;

[0008] S103, Based on the preset static performance library, service quality requirement information and available computing resource snapshots, establish a multi-objective optimization task, which includes a delay objective function, an energy consumption objective function and an accuracy loss objective function;

[0009] S104 uses the improved NSGA-III algorithm to solve the multi-objective optimization task. An initial variable population is generated based on the set of decision variables. Each individual in the population corresponds to a combination of decision variables. Then, according to the preset number of generations, iterative evolution is performed on all combinations of decision variables. The non-dominated solution is selected as the Pareto optimal solution to form the optimal set of decision variables.

[0010] S105. Select the appropriate combination of computing power scheduling decision variables based on the optimal set of decision variables and the preset business strategy.

[0011] Furthermore, S104 includes:

[0012] An initial variable population is generated based on the set of decision variables, and each individual in the population corresponds to a combination of decision variables;

[0013] Obtain historical service quality requirement information from the historical computing power scheduling task database, determine the QoS critical region based on the historical service quality requirement information, and increase the number of individuals in the population within the QoS critical region.

[0014] Iterative evolution is performed on all individuals in the population according to the preset number of generations, and the Pareto front solution set is selected from the evolution results;

[0015] Simulated annealing algorithm is used to perform local search enhancement on Pareto front solution set. Simulated annealing iteration is performed on the adjacent frequency ranges of each individual in Pareto front solution set to optimize the frequency range of each individual in Pareto front solution set.

[0016] In the optimized Pareto front solution set, non-dominated solutions are selected as Pareto optimal solutions to form the optimal decision variable set.

[0017] Furthermore, S104 also includes:

[0018] Based on historical service quality demand information, the QoS critical region is determined by using the range of ±10% of the historical service quality demand boundary value, and the number of individuals in the QoS critical region is increased to 1.5-2 times the standard density.

[0019] In the Pareto front solution set, ±1 frequency ranges of each individual frequency range are used as its adjacent frequency ranges to perform simulated annealing local search enhancement with an initial temperature of 5℃, a temperature attenuation coefficient of 0.85, and 20 iterations.

[0020] Furthermore, S103 includes:

[0021] The latency load correction factor is calculated based on the baseline latency data and available computing resource snapshots in the preset static performance library. The corrected latency data is calculated using the latency load correction factor and the baseline latency data. The latency objective function is established using the corrected latency data and the latency quality requirements in the service quality requirement information.

[0022] The energy consumption correction factor is calculated based on the baseline energy consumption data and available computing resource snapshots in the preset static performance library. The corrected energy consumption data is calculated using the energy consumption correction factor and the baseline energy consumption data. The energy consumption objective function is established using the corrected energy consumption data and the energy consumption quality requirements in the service quality requirements.

[0023] The accuracy correction factor is calculated based on the baseline accuracy data in the preset static performance library and the snapshot of available computing resources. The corrected accuracy data is then calculated using the accuracy correction factor and the baseline accuracy data. Finally, the accuracy loss objective function is established using the corrected accuracy data.

[0024] Furthermore, S103 also includes:

[0025] The preset static performance library includes benchmark latency data, benchmark power consumption data, and benchmark accuracy data for different computing paths under different batch sizes and frequency levels.

[0026] The delay objective function is f1(L) = (LL) min ) / (L max -L min ), where L is the corrected delay data, L min For the minimum latency quality requirement in the service quality requirements information, L max The highest latency quality requirement in the service quality demand information;

[0027] The energy consumption objective function is f2(P) = P / P max Where P is the corrected energy consumption data, P max The maximum energy consumption requirement in the service quality demand information;

[0028] The objective function for the accuracy loss is f3(A)=1-A, where A is the corrected accuracy data.

[0029] Furthermore, S105 includes:

[0030] The preset business strategies include high-performance mode, energy-saving mode, and balanced mode;

[0031] When the preset business strategy is high-performance mode, select the combination of computing power scheduling decision variables that minimizes the value of the delay objective function;

[0032] When the preset business strategy is energy-saving mode, select the combination of computing power scheduling decision variables that minimizes the value of the energy consumption objective function;

[0033] When the preset business strategy is balanced mode, the combination of computing power scheduling decision variables that minimizes the sum of the delay objective function, energy consumption objective function, and accuracy loss objective function is selected.

[0034] Furthermore, the method also includes:

[0035] After performing computing power scheduling operations using the selected combination of computing power scheduling decision variables, the actual service quality indicators and energy consumption information are collected in real time and stored in the historical computing power scheduling task database.

[0036] Secondly, embodiments of the present invention provide a multimodal computing power scheduling device, comprising:

[0037] The computing task response module is used to receive multimodal computing tasks, parse and obtain service quality requirement information, and obtain the current available computing resource information of each computing node to form a snapshot of available computing resources.

[0038] The decision variable construction module is used to construct a set of decision variables, including computation path, batch size, approximate computation kernel state, and frequency level, based on a preset static performance library.

[0039] The optimization task building module is used to create multi-objective optimization tasks based on a preset static performance library, service quality requirement information, and snapshots of available computing resources.

[0040] The task calculation module is optimized to solve multi-objective optimization tasks using the improved NSGA-III algorithm, select non-dominated solutions as Pareto optimal solutions, and form the optimal set of decision variables.

[0041] The computing power scheduling decision module is used to select the appropriate combination of computing power scheduling decision variables based on the optimal set of decision variables and the preset business strategy.

[0042] Thirdly, embodiments of the present invention provide a server, including:

[0043] One or more processors;

[0044] Storage device for storing one or more programs.

[0045] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-described multimodal computing power scheduling method.

[0046] Fourthly, embodiments of the present invention provide a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to execute the above-described multimodal computing power scheduling method.

[0047] This invention provides a multimodal computing power scheduling method, apparatus, server, and storage medium. The method establishes a multi-objective optimization task including three key indicators: latency, energy consumption, and accuracy. Simultaneously, it corrects the baseline performance of computing nodes based on their current resource utilization. With a configuration of decision variables across four dimensions, an improved NSGA-III algorithm is used for multi-objective collaborative optimization of the three key indicators. The optimal Pareto non-dominated solution is used to form an optimal set of decision variables for precise computing power allocation based on business strategies. The improved algorithm automatically generates the optimal set of decision variables, improving the accuracy of computing power scheduling decisions. Furthermore, the introduction of a static performance library and the current resource utilization of computing nodes during the computation process allows for performance correction. This multi-objective collaborative optimization, compared to a single optimization objective, not only ensures service quality but also reduces energy consumption and improves the utilization of computing resources. It can balance conflicting multi-dimensional indicators such as accuracy, latency, and frame rate, especially between service quality (QoS) and energy consumption, providing precise computing power scheduling flexibly based on the task description of the multimodal computing task. Attached Figure Description

[0048] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0049] Figure 1 This is a flowchart of a multimodal computing power scheduling method according to Embodiment 1 of the present invention;

[0050] Figure 2 This is a flowchart of a multimodal computing power scheduling method according to Embodiment 2 of the present invention;

[0051] Figure 3 This is a flowchart of a multimodal computing power scheduling method according to Embodiment 3 of the present invention;

[0052] Figure 4 This is a schematic diagram of the structure of a multimodal computing power scheduling device according to Embodiment 4 of the present invention;

[0053] Figure 5 This is a structural diagram of the server described in Embodiment 5 of the present invention. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0055] In edge computing, cloud computing, and other collaborative architectures, there are situations where various heterogeneous computing resources, such as ARM-NPU, GPU, CPU-AVX, and FPGA, are utilized. These heterogeneous computing resources exhibit significant differences in computing accuracy, power consumption characteristics, and processing latency. Existing computing power scheduling methods, when optimizing a single objective, inevitably face optimization conflicts between Quality of Service (QoS) and energy consumption. Furthermore, they cannot comprehensively consider multiple QoS metrics such as accuracy, latency, and frame rate, making it difficult to balance multiple dimensions, especially QoS and energy consumption, to achieve precise computing power scheduling and ensure efficient utilization of computing resources while maintaining QoS. The following solution is provided:

[0056] Example 1

[0057] Figure 1 This is a flowchart of a multimodal computing power scheduling method according to Embodiment 1 of the present invention. This embodiment obtains the service quality requirement information of multimodal computing tasks, constructs a multi-objective optimization task using snapshots of the currently available computing resources of each node, and constructs a set of decision variables using a pre-set static performance library. The improved NSGA-III algorithm is then used to solve the multi-objective optimization task, forming an optimal set of decision variables for matching computing power with business strategies. Specifically, the method includes the following steps:

[0058] S101 receives multimodal computing tasks, obtains service quality requirement information by parsing the task description file, and obtains the current available computing resource information of each computing node to form a snapshot of available computing resources.

[0059] When users input multimodal computing tasks, they often input a task description file along with the task. This file describes the task's purpose, performance requirements, resource consumption, and other details. The task description file can be in JSON format. By parsing the JSON file, information related to quality of service (QoS) requirements can be extracted, including: minimum frame rate requirement F. min Minimum accuracy requirement A min Maximum latency tolerance L max and maximum energy consumption requirement P maxAt the same time, by querying the resource manager of each computing node of the computing service provider, the current available resources of each computing node can be obtained, including available computing resource information such as CPU performance, memory capacity, GPU memory, as well as the real-time load rate of each computing node, device temperature information, and network topology, forming a snapshot of available computing resources.

[0060] S102, based on the preset static performance library, constructs a set of decision variables using Path to represent the computation path, Batch to represent the batch size, Kernel to represent the approximate computation kernel state, and Freq to represent the frequency level.

[0061] The pre-built static performance library can be a pre-established static performance database that stores performance benchmark data for each computing path under different configurations. Using this pre-built static performance library, the benchmark latency, benchmark power consumption, and benchmark accuracy provided by each computing path at different batch sizes and frequency levels can be determined. For example, due to the heterogeneity of computing resources, the computing path can include ARM-NPU paths, GPU paths, CPU-AVX paths, and FPGA paths; the batch size can be an integer between 1 and 32; the approximate computing core status (Kernel) can be a Boolean value of 1 (enabled) or 0 (disabled), indicating whether the INT4 approximate computing core is enabled; the frequency level (Freq) can be divided into five discrete levels, each with an integer value between 0 and 4, corresponding to frequencies of 1350MHz, 1470MHz, 1590MHz, 1710MHz, and 1830MHz. Each individual is constructed with a set of decision variables containing four dimensions: Path (representing the computing path), Batch (representing the batch size), Kernel (representing the approximate computing core status), and Freq (representing the frequency level). Based on the values ​​of four variables among individuals that meet the service quality requirements of multimodal computing tasks—baseline latency, baseline power consumption, and baseline accuracy—corresponding computing resources can be scheduled to provide computing power services for multimodal computing tasks.

[0062] S103. Based on the preset static performance library, service quality requirement information and available computing resource snapshots, establish a multi-objective optimization task, which includes a delay objective function, an energy consumption objective function and an accuracy loss objective function.

[0063] Based on benchmark performance data in a static performance database, and combined with real-time load information from available computing resource snapshots, available computing resources are adjusted. Three objective functions—latency, energy consumption, and accuracy loss—are established to calculate expected performance indicators under each decision variable configuration. Service quality requirements are used as constraints to form a multi-objective optimization task. By solving this multi-objective optimization task and selecting the optimal decision variables, precise computing power scheduling can be achieved by configuring the resources according to these optimal variables.

[0064] S104 uses the improved NSGA-III algorithm to solve the multi-objective optimization task. An initial variable population is generated based on the set of decision variables. Each individual in the population corresponds to a combination of decision variables. Then, according to the preset number of generations, iterative evolution is performed on all combinations of decision variables. The non-dominated solution is selected as the Pareto optimal solution to form the optimal set of decision variables.

[0065] The NSGA-III algorithm maintains population diversity and effectively balances and converges this diversity by introducing widely distributed reference points. It is suitable for finding high-quality decision variables in multi-objective optimization tasks, thus providing more efficient and suitable computational power allocation for multimodal computing tasks. The algorithm first initializes the population, then performs evolutionary iterations, executing simulated binary crossover and polynomial mutation operations in each generation. Next, it selects the next generation of the population through fast non-dominated sorting and crowding calculation, ultimately outputting multiple non-dominated solutions as a Pareto optimal solution set, forming the optimal set of decision variables. This optimal set of decision variables represents the best combination of optimization objectives achievable under different combinations of decision variables. It serves as a reference for computational power scheduling in business strategies, allowing for the selection of appropriate combinations of decision variables based on the service quality requirements of multimodal computing tasks. To further enhance population diversity, we can identify critical regions and increase the density of reference points within those regions. We can also combine simulated annealing with simulated annealing to enhance local search for Pareto front solutions, thereby improving search and convergence efficiency. Simultaneously, we can optimize frequency settings to improve energy efficiency, resulting in an improved NSGA-III algorithm that is better suited to the characteristics of balancing multiple conditions during computing power scheduling.

[0066] S105. Select the appropriate combination of computing power scheduling decision variables based on the optimal set of decision variables and the preset business strategy.

[0067] The final solution is selected from the optimal set of decision variables based on different business strategies. The high-performance mode prioritizes latency performance, the energy-saving mode focuses on energy consumption optimization, and the balanced mode comprehensively considers the balance of multiple objective functions.

[0068] In one optional implementation of this embodiment, the preset business strategy includes a high-performance mode, an energy-saving mode, and a balanced mode.

[0069] Business strategies can be pre-specified through configuration parameters and can be dynamically adjusted according to the priority of multimodal computing tasks and user needs. The provided API interface allows users to specify strategy modes in the GUI interface or command line, and can also support automatic strategy selection based on historical task characteristics.

[0070] When the preset business strategy is high-performance mode, select the combination of computing power scheduling decision variables that minimizes the value of the delay objective function.

[0071] In order to meet the needs of high-efficiency and low-latency computing, the high-performance mode pays less attention to factors such as energy consumption. It can traverse the optimal set of decision variables, calculate the latency objective function under the configuration of each individual variable in the set, and select the combination of decision variables with the smallest latency objective function value as the final computing power scheduling scheme.

[0072] When the preset business strategy is energy-saving mode, select the combination of computing power scheduling decision variables that minimizes the value of the energy consumption objective function.

[0073] In order to meet the computing needs of low energy consumption and low power consumption, the energy-saving mode takes into account factors such as latency less. It can traverse the optimal decision variable set, calculate the energy consumption objective function under the configuration of each individual variable in the set, and select the decision variable combination with the smallest energy consumption objective function value as the final computing power scheduling scheme.

[0074] When the preset business strategy is balanced mode, the combination of computing power scheduling decision variables that minimizes the sum of the delay objective function, energy consumption objective function, and accuracy loss objective function is selected.

[0075] The balanced approach is a solution that balances computational performance, energy consumption, and model accuracy loss. It seeks the optimal equilibrium among these three factors, ensuring sufficient computational performance without excessive energy consumption, while keeping the model's accuracy loss within acceptable limits. This can be achieved by iterating through the optimal set of decision variables, calculating the sum of the delay objective function, energy consumption objective function, and accuracy loss objective function for each individual variable configuration in the set, and selecting the combination of decision variables with the smallest total sum as the final computational power scheduling scheme. The delay objective function, energy consumption objective function, and accuracy loss objective function can also be weighted and summed. The sum of the weight coefficients of the three objective functions is 1. The weight coefficients can be configured according to actual needs or usage scenarios. For example, if computational efficiency needs to be prioritized, but energy consumption and model accuracy loss also need to be considered, the weight coefficient of the delay objective function can be increased accordingly. For example, the weight coefficient of the delay objective function can be set to w1=0.4, the weight coefficient of the energy consumption objective function to w2=0.3, and the weight coefficient of the accuracy loss objective function to w3=0.3. The decision variable with the smallest value after the weighted sum of the three objective functions is selected as the final computing power scheduling scheme.

[0076] This embodiment establishes a multi-objective optimization task encompassing three key indicators: latency, energy consumption, and accuracy. Simultaneously, it adjusts the baseline performance of computing nodes based on their current resource utilization. With a configuration of decision variables across four dimensions, an improved NSGA-III algorithm is employed for multi-objective collaborative optimization of the three key indicators. The optimal Pareto non-dominated solution is used to form an optimal set of decision variables for precise computing power allocation based on business strategies. The improved algorithm automatically generates the optimal set of decision variables, enhancing the accuracy of computing power scheduling decisions. Furthermore, the introduction of a static performance database and the current resource utilization of computing nodes during computation allows for performance correction. This multi-objective collaborative optimization, compared to a single optimization objective, not only ensures service quality but also reduces energy consumption and improves the utilization of computing resources. It can balance conflicting multi-dimensional indicators such as accuracy, latency, and frame rate, especially between service quality (QoS) and energy consumption, providing precise computing power scheduling flexibly based on the task description of multimodal computing tasks.

[0077] Optionally, the method further includes:

[0078] After performing computing power scheduling operations using the selected combination of computing power scheduling decision variables, the actual service quality indicators and energy consumption information are collected in real time and stored in the historical computing power scheduling task database.

[0079] To continuously optimize the computing power scheduling system and improve algorithm accuracy, model applicability, and stability, monitoring agents deployed on each computing node collect actual service quality indicators (such as actual frame rate, accuracy, latency, and power consumption) provided to multimodal computing tasks after computing power scheduling operations are performed based on computing power scheduling decision variables. The actual achievement of these indicators and actual energy consumption information are stored in a historical computing power scheduling task database for subsequent model optimization. During optimization, parameters in the computing power scheduling calculation process (such as parameter values ​​in individual decision variables) can be adjusted based on the degree of deviation between the computing power scheduling decision variables and the actual service quality indicators. Alternatively, corresponding weighting coefficients or correction coefficients can be added during the multi-objective optimization task solution process. The update frequency can be set to one optimization update every 100 new data samples received to ensure the accuracy of computing power scheduling.

[0080] Example 2

[0081] Figure 2 This is a flowchart of a multimodal computing power scheduling method according to Embodiment 2 of the present invention. This embodiment is based on the above embodiment and optimized. In this embodiment, S104 is specifically optimized as follows:

[0082] An initial variable population is generated based on the set of decision variables, and each individual in the population corresponds to a combination of decision variables;

[0083] Obtain historical service quality requirement information from the historical computing power scheduling task database, determine the QoS critical region based on the historical service quality requirement information, and increase the number of individuals in the population within the QoS critical region.

[0084] Iterative evolution is performed on all individuals in the population according to the preset number of generations, and the Pareto front solution set is selected from the evolution results;

[0085] Simulated annealing algorithm is used to perform local search enhancement on Pareto front solution set. Simulated annealing iteration is performed on the adjacent frequency ranges of each individual in Pareto front solution set to optimize the frequency range of each individual in Pareto front solution set.

[0086] In the optimized Pareto front solution set, non-dominated solutions are selected as Pareto optimal solutions to form the optimal decision variable set.

[0087] Accordingly, the multimodal computing power scheduling method provided in this embodiment specifically includes:

[0088] S201: Receive multimodal computing tasks, obtain service quality requirement information by parsing the task description file, and obtain the current available computing resource information of each computing node to form a snapshot of available computing resources.

[0089] S202, based on the preset static performance library, uses Path to represent the computation path, Batch to represent the batch size, Kernel to represent the approximate computation kernel state, and Freq to represent the frequency level to construct a set of decision variables.

[0090] S203. Based on the preset static performance library, service quality requirement information and available computing resource snapshots, establish a multi-objective optimization task, which includes a delay objective function, an energy consumption objective function and an accuracy loss objective function.

[0091] S204. Generate an initial variable population based on the set of decision variables. Each individual in the population corresponds to a combination of decision variables.

[0092] Based on the dimensions of the decision variable set, an initial variable population containing multiple individuals is generated. Each individual in the population is a quadruple (Path, Batch, Kernel, Freq) with four dimensions, and the individuals in the population are uniformly distributed in the decision space. For example, the Latin hypercube sampling method can be used to generate the initial population, with a population size of 50.

[0093] S205: Obtain historical service quality requirement information from the historical computing power scheduling task database, determine the QoS critical region based on the historical service quality requirement information, and increase the number of individuals in the population within the QoS critical region.

[0094] By querying the historical service quality information stored in the historical computing power scheduling task database, and based on the historical service quality requirements of each task, the distribution of historical QoS quality (frame rate, accuracy, latency) is analyzed. Based on the distribution, the QoS critical region is determined, and the density of reference points (population individuals) in the critical region is increased to improve the search capability in the critical region.

[0095] For example, based on historical QoS demand information, a QoS critical region is defined within ±10% of the historical QoS demand boundary values, and the number of individuals in the QoS critical region is increased to 1.5-2 times the standard density. For instance, if the historical QoS demand boundary values ​​are a frame rate of 30fps, an accuracy of 85%, and a latency of 100ms, then the frame rate range [27-33]fps, the accuracy range [76.5%-93.5%], and the latency range [90-110]ms are defined as the QoS critical region. Within this region, an adaptive reference point generation strategy is adopted, increasing the reference point density from the standard density of 10 points / unit to 15-20 points / unit by 1.5-2 times to ensure sufficient search granularity in the critical region.

[0096] S206, perform iterative evolution on all individuals in the population according to the preset number of generations, and select the Pareto front solution set from the evolution results.

[0097] In the process of population evolution and iteration, each generation performs selection, crossover, and mutation operations. The selection scale is 2, the simulated binary crossover method is used, the distribution index is set to 20, the polynomial mutation method is used, the mutation probability is 25%, and the preset number of generations can be 100. Finally, the top 15% of non-dominated solutions are selected as the Pareto front solution set through non-dominated sorting.

[0098] S207 utilizes the simulated annealing algorithm to perform local search enhancement on the Pareto front solution set. Simulated annealing iterations are performed on the adjacent frequency ranges of each individual in the Pareto front solution set to optimize the frequency range of each individual in the Pareto front solution set.

[0099] In the Pareto front solution set selected after evolutionary iteration, for each front solution, a search is performed again in its adjacent frequency ranges centered on its own frequency range. The impact of the adjacent frequency ranges on the objective function is evaluated through local search enhancement.

[0100] For example, in the Pareto front solution set, ±1 of each individual frequency level is used as its adjacent frequency levels to perform simulated annealing local search enhancement with an initial temperature of 5°C, a temperature attenuation coefficient of 0.85, and 20 iterations.

[0101] The simulated annealing algorithm is adopted. In the Pareto front solution set, the frequency range ±1 of each individual's frequency range Freq is taken as its neighborhood. The initial temperature of simulated annealing is set to 5℃, the temperature decay coefficient is 0.85, and the number of iterations is 20. With a probability of 0.15, inferior solutions are accepted to escape the local optimum. Through 20 iterations, better frequencies are found to further optimize the frequency configuration.

[0102] S208. Select the non-dominated solution from the optimized Pareto front solution set as the Pareto optimal solution to form the optimal decision variable set.

[0103] The Pareto front solution set enhanced by local search is then subjected to another non-dominated sorting. All non-dominated solutions in the first non-dominated layer are selected as the final optimal decision variable set, typically containing 15-20 non-dominated solutions. By increasing the density in the QoS critical region and employing the neighborhood local search enhancement optimization strategy of the simulated annealing algorithm, the algorithm can improve convergence speed while maintaining the diversity of the generated solution set distribution.

[0104] S209. Select the appropriate combination of computing power scheduling decision variables based on the optimal set of decision variables and the preset business strategy.

[0105] This embodiment enhances the algorithm's search efficiency and convergence speed in critical regions by using adaptive reference points in QoS critical areas, thereby reducing computational resource consumption. It also enhances local search by introducing simulated annealing, and further optimizes energy consumption by performing fine-grained searches at different frequency levels while keeping other objectives constant. Furthermore, improvements to the NSGA-III algorithm increase the uniformity of the Pareto front solution set, resulting in a richer set of optimal decision variables with more granular indicators, enabling more precise allocation of computing power based on multimodal computing tasks and the current resource occupancy status of computing nodes.

[0106] Example 3

[0107] Figure 3 This is a flowchart of a multimodal computing power scheduling method according to Embodiment 3 of the present invention. This embodiment is an optimization based on the above embodiment. In this embodiment, S103 includes:

[0108] The latency load correction factor is calculated based on the baseline latency data and available computing resource snapshots in the preset static performance library. The corrected latency data is calculated using the latency load correction factor and the baseline latency data. The latency objective function is established using the corrected latency data and the latency quality requirements in the service quality requirement information.

[0109] The energy consumption correction factor is calculated based on the baseline energy consumption data and available computing resource snapshots in the preset static performance library. The corrected energy consumption data is calculated using the energy consumption correction factor and the baseline energy consumption data. The energy consumption objective function is established using the corrected energy consumption data and the energy consumption quality requirements in the service quality requirements.

[0110] The accuracy correction factor is calculated based on the baseline accuracy data in the preset static performance library and the snapshot of available computing resources. The corrected accuracy data is then calculated using the accuracy correction factor and the baseline accuracy data. Finally, the accuracy loss objective function is established using the corrected accuracy data.

[0111] Accordingly, the multimodal computing power scheduling classification method provided in this embodiment specifically includes:

[0112] S301 receives multimodal computing tasks, obtains service quality requirement information by parsing the task description file, and obtains the current available computing resource information of each computing node to form a snapshot of available computing resources.

[0113] S302, based on the preset static performance library, uses Path to represent the computation path, Batch to represent the batch size, Kernel to represent the approximate computation kernel state, and Freq to represent the frequency level to construct a set of decision variables.

[0114] S303: Calculate the latency load correction factor based on the baseline latency data and available computing resource snapshots in the preset static performance library, calculate the corrected latency data using the latency load correction factor and the baseline latency data, and establish the latency objective function using the corrected latency data and the latency quality requirements in the service quality requirement information.

[0115] Specifically, the preset static performance library includes benchmark latency data, benchmark power consumption data, and benchmark accuracy data for different computing paths under different batch sizes and frequency levels. The static performance library can be established through offline benchmark testing and contains complete performance data for computing paths such as ARM-NPU, GPU, CPU-AVX, and FPGA under 32 batch sizes and 5 frequency level combinations, totaling 4×32×5=640 performance records.

[0116] The baseline latency data L under the current decision variable configuration is obtained by querying the preset static performance database. baseThen, based on the actual load rate of the current node in the available computing resource snapshot, calculate the load correction factor α = 1 + 0.5 × (current load rate / 100), and correct the latency data L = L base ×α.

[0117] Establish the delay objective function using the corrected delay data: f1(L)=(LL) min ) / (L max -L min ), where L is the corrected delay data, L min For the minimum latency quality requirement in the service quality requirements information, L max This represents the highest latency quality requirement in the service quality requirements information. For example, the corrected latency data is calculated using baseline latency data retrieved from a static performance database, and then combined with the lowest latency quality requirement L from the service quality requirements information retrieved from the task description file. min and the highest latency quality requirement L max Establish a delay objective function. For example, L... min It is usually set to 0ms, L max Typically set to 100ms, the delay objective function also needs to normalize the absolute delay to the [0,1] interval.

[0118] It should be noted that when calculating the correction factor, the impact of GPU bandwidth utilization on latency also needs to be considered for the GPU path. Therefore, an additional bandwidth correction factor needs to be calculated based on GPU bandwidth utilization. When GPU bandwidth utilization exceeds 80%, an additional bandwidth correction factor β = 1 + 0.3 × (bandwidth utilization - 80) / 20 is calculated, and the recalculated corrected latency data L = L base ×α×β.

[0119] S304: Calculate the energy consumption correction factor based on the baseline energy consumption data and available computing resource snapshots in the preset static performance library, calculate the corrected energy consumption data using the energy consumption correction factor and the baseline energy consumption data, and establish the energy consumption objective function using the corrected energy consumption data and the energy consumption quality requirements in the service quality requirements.

[0120] For energy consumption calculations, the impact of device temperature on energy consumption at computing nodes needs to be considered, and an energy consumption correction factor needs to be calculated. Baseline energy consumption data P under the current decision variable configuration is obtained by querying a preset static performance database. base Then, based on the actual temperature of the current node in the available computing resource snapshot, the energy consumption correction factor γ = 1 + 0.01 × (current temperature - 70) is calculated. This method is used when the temperature is >70℃, and if the temperature is <70℃, the energy consumption correction factor is 1. The corrected energy consumption data P = P base ×γ.

[0121] Establish an energy consumption objective function using corrected energy consumption data: f2(P) = P / P max Where P is the corrected energy consumption data, P max This refers to the maximum energy consumption requirement in the service quality requirements information. For example, the corrected energy consumption data is calculated using baseline energy consumption data retrieved from the static performance database, and then combined with the maximum energy consumption quality requirement P from the service quality requirements information retrieved from the task description file. max Establish an energy consumption objective function. For example, P... max Typically, it's 180W. The energy consumption objective function also needs to normalize the absolute energy consumption to the [0,1] interval. If the calculated P > P max This indicates that the energy consumption exceeds the standard, and the solution should be discarded directly.

[0122] S305: Calculate the accuracy correction factor based on the baseline accuracy data and available computing resource snapshots in the preset static performance library, calculate the corrected accuracy data using the accuracy correction factor and the baseline accuracy data, and establish the accuracy loss objective function using the corrected accuracy data.

[0123] For precision calculations, it is necessary to consider whether the approximate calculation kernel of the calculation node is enabled, and obtain the baseline precision data A under the current decision variable configuration by querying the preset static performance library. base Then, the precision correction factor is calculated based on the Boolean value of the approximate computing kernel of the current node in the snapshot of available computing resources. When Kernel is 1, indicating that the INT4 approximate computing kernel is enabled, the precision correction factor is determined to be 0.95, that is, a loss of 5% precision, and the corrected precision data A = Abase × 0.95 is obtained.

[0124] An accuracy loss objective function is established using the corrected accuracy data: f3(A) = 1 - A, where A is the corrected accuracy data. For example, the corrected accuracy data is calculated using baseline accuracy data retrieved from a static performance database, and the accuracy loss objective function is established accordingly. This accuracy loss objective function represents the loss of model accuracy due to computational resource constraints.

[0125] S306 utilizes an improved NSGA-III algorithm to solve a multi-objective optimization task. An initial variable population is generated based on the set of decision variables, and each individual in the population corresponds to a combination of decision variables. Then, iterative evolution is performed on all combinations of decision variables according to a preset number of generations. The non-dominated solution is selected as the Pareto optimal solution, forming the optimal set of decision variables.

[0126] S307: Select the appropriate combination of computing power scheduling decision variables based on the optimal set of decision variables and the preset business strategy.

[0127] This embodiment introduces a multi-metric correction mechanism, utilizing the current computing resource occupancy of each computing node and a static performance database. By considering factors such as load rate, device temperature, and the activation status of approximate computing cores in the actual operating environment, it corrects the baseline performance, improving the accuracy of performance prediction and making the predicted performance closer to reality. Subsequently, modeling is performed based on the obtained performance data, enabling refined scheduling of computing resources and improving their utilization.

[0128] Example 4

[0129] Figure 4 This is a schematic diagram of a multimodal computing power scheduling device according to Embodiment 4 of the present invention. In this embodiment, the multimodal computing power scheduling device includes:

[0130] The computing task response module 810 is used to receive multimodal computing tasks and parse them to obtain service quality requirement information, and obtain the current available computing resource information of each computing node to form a snapshot of available computing resources.

[0131] The decision variable construction module 820 is used to construct a set of decision variables containing computation path, batch size, approximate computation kernel state and frequency level based on a preset static performance library;

[0132] The optimization task construction module 830 is used to build multi-objective optimization tasks based on a preset static performance library, service quality requirement information, and snapshots of available computing resources.

[0133] The optimization task calculation module 840 is used to solve multi-objective optimization tasks using the improved NSGA-III algorithm, select non-dominated solutions as Pareto optimal solutions, and form the optimal set of decision variables.

[0134] The computing power scheduling decision module 850 is used to select the appropriate combination of computing power scheduling decision variables based on the optimal set of decision variables and the preset business strategy.

[0135] This embodiment receives multimodal computing tasks and parses task description files through a computing task response module, while simultaneously acquiring available computing resource information. A decision variable construction module builds a set of decision variables encompassing four dimensions. An optimization task construction module constructs a multi-objective optimization task including latency, energy consumption, and accuracy. An optimization task computation module uses an improved NSGA-III algorithm to solve the multi-objective optimization task, generating an optimal set of decision variables. A computing power scheduling decision module selects appropriate combinations of computing power scheduling decision variables from the optimal set based on business strategies. The improved algorithm automatically generates the optimal set of decision variables, improving the accuracy of computing power scheduling decisions. Furthermore, the introduction of a static performance library and the current resource utilization of computing nodes during computation allows for performance correction. This multi-objective collaborative optimization, compared to a single optimization objective, not only ensures service quality but also reduces energy consumption and improves computing resource utilization. It can balance conflicting optimization metrics such as accuracy, latency, and frame rate, especially between service quality (QoS) and energy consumption, providing precise computing power scheduling based on the task description of the multimodal computing task.

[0136] The multimodal computing power scheduling device provided in this embodiment of the invention can execute the multimodal computing power scheduling method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0137] Example 5

[0138] Figure 5 This is a structural diagram of a server according to Embodiment 5 of the present invention. Figure 5 A block diagram of an exemplary server 12 suitable for implementing embodiments of the present invention is shown. Figure 5 The server 12 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0139] like Figure 5 As shown, server 12 is presented in the form of a general-purpose computing device. The components of server 12 may include, but are not limited to: one or more processors or processing units 16, system memory 28, and bus 18 connecting different system components (including system memory 28 and processing unit 16).

[0140] Bus 18 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0141] Server 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by server 12, including volatile and non-volatile media, removable and non-removable media.

[0142] System memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. Server 12 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 may be used to read and write non-removable, non-volatile magnetic media (… Figure 5 Not shown; usually referred to as a "hard drive"). Although Figure 5 Not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 18 via one or more data media interfaces. System memory 28 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0143] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in system memory 28. Such program modules 42 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 42 typically perform the functions and / or methods described in the embodiments of the present invention.

[0144] Server 12 can also communicate with one or more external devices 14 (e.g., keyboard, pointing device, display 24, etc.), and with one or more devices that enable users to interact with server 12 / electronic devices / computers, and / or with any device that enables server 12 to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed through input / output (I / O) interface 22. Furthermore, server 12 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 20. Figure 5 As shown, network adapter 20 communicates with other modules of server 12 via bus 18. It should be understood that, although... Figure 5As not shown, it can be combined with other hardware and / or software modules, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0145] The processing unit 16 executes various functional applications and data processing by running programs stored in the system memory 28, such as implementing the multimodal computing power scheduling method provided in the embodiments of the present invention.

[0146] Example 6

[0147] Embodiment 6 of the present invention also provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to execute the multimodal computing power scheduling method provided in the above embodiments.

[0148] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0149] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0150] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0151] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0152] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A multimodal computing power scheduling method, characterized in that, include: S101 receives multimodal computing tasks, obtains service quality requirement information by parsing the task description file, and obtains the current available computing resource information of each computing node to form a snapshot of available computing resources; S102, based on the preset static performance library, construct a set of decision variables using Path to represent the computation path, Batch to represent the batch size, Kernel to represent the approximate computation kernel state, and Freq to represent the frequency level; S103, Based on the preset static performance library, service quality requirement information and available computing resource snapshots, establish a multi-objective optimization task, which includes a delay objective function, an energy consumption objective function and an accuracy loss objective function; S103 includes: The latency load correction factor is calculated based on the baseline latency data in the preset static performance library and the actual node load rate in the available computing resource snapshot. The corrected latency data is obtained by correcting the baseline latency data, and then a latency objective function is established with the minimum and maximum latency quality requirements in the service quality requirement information. The energy consumption correction factor is calculated based on the baseline energy consumption data in the preset static performance library and the actual temperature of the current node in the snapshot of available computing resources. The corrected energy consumption data is obtained by correcting the baseline energy consumption data, and then the energy consumption objective function is established with the maximum energy consumption quality requirement in the service quality requirements. The accuracy correction factor is calculated based on the baseline accuracy data in the preset static performance library and the approximate calculation kernel Boolean value of the current node in the snapshot of available computing resources. The corrected accuracy data is obtained by correcting the baseline accuracy data, and an accuracy loss objective function is established. S104 uses the improved NSGA-III algorithm to solve the multi-objective optimization task. An initial variable population is generated based on the set of decision variables. Each individual in the population corresponds to a combination of decision variables. Then, according to the preset number of generations, iterative evolution is performed on all combinations of decision variables. The non-dominated solution is selected as the Pareto optimal solution to form the optimal set of decision variables. S105. Select the appropriate combination of computing power scheduling decision variables based on the optimal set of decision variables and the preset business strategy.

2. The method according to claim 1, characterized in that, S104 includes: An initial variable population is generated based on the set of decision variables, and each individual in the population corresponds to a combination of decision variables; Obtain historical service quality requirement information from the historical computing power scheduling task database, determine the QoS critical region based on the historical service quality requirement information, and increase the number of individuals in the population within the QoS critical region. Iterative evolution is performed on all individuals in the population according to the preset number of generations, and the Pareto front solution set is selected from the evolution results; Simulated annealing algorithm is used to perform local search enhancement on Pareto front solution set. Simulated annealing iteration is performed on the adjacent frequency ranges of each individual in Pareto front solution set to optimize the frequency range of each individual in Pareto front solution set. In the optimized Pareto front solution set, non-dominated solutions are selected as Pareto optimal solutions to form the optimal decision variable set.

3. The method according to claim 2, characterized in that, S104 further includes: Based on historical service quality demand information, the QoS critical region is determined by using the range of ±10% of the historical service quality demand boundary value, and the number of individuals in the QoS critical region is increased to 1.5-2 times the standard density. In the Pareto front solution set, ±1 frequency ranges of each individual frequency range are used as its adjacent frequency ranges to perform simulated annealing local search enhancement with an initial temperature of 5℃, a temperature attenuation coefficient of 0.85, and 20 iterations.

4. The method according to claim 1, characterized in that, S103 further includes: The preset static performance library includes benchmark latency data, benchmark power consumption data, and benchmark accuracy data for different computing paths under different batch sizes and frequency levels. The delay objective function is f1(L) = (LL) min ) / (L max -L min ), where L is the corrected delay data, L min For the minimum latency quality requirement in the service quality requirements information, L max The highest latency quality requirement in the service quality demand information; The energy consumption objective function is f2(P) = P / P max Where P is the corrected energy consumption data, P max The maximum energy consumption requirement in the service quality demand information; The objective function for the accuracy loss is f3(A)=1-A, where A is the corrected accuracy data.

5. The method according to claim 1, characterized in that, S105 includes: The preset business strategies include high-performance mode, energy-saving mode, and balanced mode; When the preset business strategy is high-performance mode, select the combination of computing power scheduling decision variables that minimizes the value of the delay objective function; When the preset business strategy is energy-saving mode, select the combination of computing power scheduling decision variables that minimizes the value of the energy consumption objective function; When the preset business strategy is balanced mode, the combination of computing power scheduling decision variables that minimizes the sum of the delay objective function, energy consumption objective function, and accuracy loss objective function is selected.

6. The method according to claim 1, characterized in that, The method further includes: After performing computing power scheduling operations using the selected combination of computing power scheduling decision variables, the actual service quality indicators and energy consumption information are collected in real time and stored in the historical computing power scheduling task database.

7. A multimodal computing power scheduling device, used to implement the multimodal computing power scheduling method as described in any one of claims 1-6, characterized in that, include: The computing task response module is used to receive multimodal computing tasks, parse and obtain service quality requirement information, and obtain the current available computing resource information of each computing node to form a snapshot of available computing resources. The decision variable construction module is used to construct a set of decision variables, including computation path, batch size, approximate computation kernel state, and frequency level, based on a preset static performance library. The optimization task building module is used to create multi-objective optimization tasks based on a preset static performance library, service quality requirement information, and snapshots of available computing resources. Specifically, a latency load correction factor is calculated based on the baseline latency data in the preset static performance library and the actual node load rate in the available computing resource snapshot. The baseline latency data is then corrected to obtain the corrected latency data, and a latency objective function is established by combining it with the minimum and maximum latency quality requirements in the service quality requirement information. The energy consumption correction factor is calculated based on the baseline energy consumption data in the preset static performance library and the actual temperature of the current node in the snapshot of available computing resources. The corrected energy consumption data is obtained by correcting the baseline energy consumption data, and then the energy consumption objective function is established with the maximum energy consumption quality requirement in the service quality requirements. The accuracy correction factor is calculated based on the baseline accuracy data in the preset static performance library and the approximate calculation kernel Boolean value of the current node in the snapshot of available computing resources. The corrected accuracy data is obtained by correcting the baseline accuracy data, and an accuracy loss objective function is established. The task calculation module is optimized to solve multi-objective optimization tasks using the improved NSGA-III algorithm, select non-dominated solutions as Pareto optimal solutions, and form the optimal set of decision variables. The computing power scheduling decision module is used to select the appropriate combination of computing power scheduling decision variables based on the optimal set of decision variables and the preset business strategy.

8. A server, characterized in that, The server includes: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the multimodal computing power scheduling method as described in any one of claims 1-6.

9. A storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the multimodal computing power scheduling method as described in any one of claims 1-6.

Images