A distributed heterogeneous task scheduling method and system based on an evolutionary algorithm

Abstract

The application discloses a distributed heterogeneous task scheduling method and system based on an evolutionary algorithm, and the method comprises the following steps: analyzing and quantifying influencing factors of distributed heterogeneous task scheduling; taking the heterogeneous chip, the server and the quantified influencing factors of the distributed heterogeneous task scheduling as an optimization target function, and establishing a whole power consumption planning model; using an improved evolutionary optimization algorithm, taking the whole power consumption planning model as an evaluation basis, scheduling and optimizing the distributed heterogeneous task to obtain an optimal scheduling scheme; through the improved evolutionary algorithm, the whole power consumption planning model and dynamic floating power consumption calculation, the optimal solution can be quickly converged in a large-scale data scene, and the scheduling efficiency and stability are improved.

Description

Technical Field

[0001] This invention relates to the field of distributed heterogeneous task scheduling technology, specifically to a distributed heterogeneous task scheduling method and system based on an evolutionary algorithm. Background Technology

[0002] The Internet of Things (IoT) primarily utilizes sensing and network communication technologies to achieve ubiquitous connectivity between people, machines, and things, providing the infrastructure for information sensing, transmission, and processing. Leveraging the widespread, stable, and efficient data communication capabilities of the IoT, various applications have permeated numerous fields such as industrial manufacturing, urban management, smart homes, autonomous driving, and environmental monitoring. Currently, the IoT is transitioning from its initial stage of isolated applications to a rapid growth phase encompassing multi-dimensional sensing, digital twins, and intelligent interconnection. However, with the explosive growth of IoT devices, effectively collecting, processing, analyzing, and utilizing this distributed, diverse, and heterogeneous massive amount of data has become a key bottleneck restricting the further development of IoT technology.

[0003] In distributed heterogeneous environments, the allocation of computing resources needs to consider factors such as task priority, device performance, and energy consumption. However, there is currently a lack of effective energy consumption estimation models and scheduling algorithms, making it difficult to achieve a balance between performance and energy consumption. Traditional scheduling algorithms are designed based on static power consumption models, assuming that device power consumption characteristics remain constant throughout the task execution cycle. This assumption does not match reality, especially in mobile edge computing scenarios, where devices experience real-time power consumption fluctuations due to changes in task load, network status, and ambient temperature. Furthermore, the integration of heterogeneous devices introduces dynamic parameter coupling effects. When a new device type is added, its unique power consumption curves will be non-linearly superimposed, leading to deviations in system-level energy efficiency predictions.

[0004] In the field of computer science, evolutionary computation, as a branch of artificial intelligence, possesses global search capabilities and the ability to adapt to complex optimization problems. It requires minimal auxiliary information and is not limited by the search space, providing an effective solution for distributed heterogeneous task scheduling problems. However, currently, there are no performance prediction models designed specifically for heterogeneous computing platforms, and most current heterogeneous scheduling algorithms generally use static energy consumption coefficients for their fitness functions. This design ignores the relationship between the real-time status of devices and environmental factors. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a distributed heterogeneous task scheduling method and system based on evolutionary algorithms. The aim is to improve scheduling efficiency by optimizing the algorithm, balancing performance and energy consumption, reducing the risk of privacy leaks, and enhancing system security.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a distributed heterogeneous task scheduling method based on an evolutionary algorithm, comprising the following steps:

[0007] Step S1: Analyze and quantify the influencing factors of distributed heterogeneous task scheduling;

[0008] Step S2: Using the influencing factors of heterogeneous chips, servers, and quantified distributed heterogeneous task scheduling as the objective function for optimization, establish an overall power consumption planning model;

[0009] Step S3: Using optimization algorithms and the overall power consumption planning model as the evaluation basis, optimize the scheduling of distributed heterogeneous tasks to obtain the optimal scheduling scheme;

[0010] The overall power consumption planning model is expressed as:

[0011] ;

[0012] In the formula, This indicates the overall power consumption of the server; Indicates floating power consumption. Indicates the base power consumption of heterogeneous chips. This indicates the standby power consumption of the memory device. This indicates the standby power consumption of the storage device. Indicates fixed costs; This indicates the system's energy efficiency.

[0013] Furthermore, the optimization algorithm employs an improved evolutionary algorithm. Using the overall power consumption planning model as the evaluation basis, this improved evolutionary optimization algorithm optimizes the scheduling of distributed heterogeneous tasks, yielding the optimal scheduling scheme. The specific process is as follows:

[0014] Step S3.1: Initialize the population for the improved evolutionary optimization algorithm;

[0015] Step S3.2: Calculate the fitness value of each individual based on the overall power consumption planning model;

[0016] Step S3.3: Select individuals in the current population using a selection strategy that combines roulette wheel selection with random competitive selection.

[0017] Step S3.4: Update the individual position through a sensed pulse evolution strategy, which includes an energy-efficient crossover strategy and a power-sensitive quantum mutation strategy;

[0018] Step S3.5: Optimize the current best individual using a local search optimization mechanism;

[0019] Step S3.6: Determine whether the current iteration meets the termination condition. If not, continue iterating; if it does, stop iterating and output the best individual in the current population as the optimal task scheduling scheme.

[0020] Furthermore, floating power consumption Represented as:

[0021] ;

[0022] In the formula, Indicates the first The overhead of completing a distributed heterogeneous task; This indicates the server's power consumption fluctuation value; Indicates the first The runtime required for a distributed heterogeneous task; Indicates the first The number of heterogeneous chips required for distributed heterogeneous tasks. Indicates the number of distributed heterogeneous tasks;

[0023] Heterogeneous chip base power consumption Represented as:

[0024] ;

[0025] In the formula, Indicates the base power consumption of heterogeneous chips; Indicates the operating voltage of the heterogeneous chip; This indicates the operating current in standby mode; This represents the longest runtime required to complete all distributed heterogeneous tasks.

[0026] Memory device standby power consumption Represented as:

[0027] ;

[0028] In the formula, This indicates the standby power consumption of the memory device. Indicates the operating voltage of the memory device. This represents the operating current of a single memory device in standby mode. Indicates the number of memory devices;

[0029] The standby power consumption of a storage device is expressed as:

[0030] ;

[0031] In the formula, This indicates the standby power consumption of the storage device. This indicates the operating voltage of the storage device. This represents the operating current of a single storage device in standby mode. This indicates the number of storage devices.

[0032] Furthermore, improve the population of the evolutionary optimization algorithm. Represented as:

[0033] ;

[0034] In the formula, Indicates the first in the population The position of each individual , Indicates the first The servers to which distributed heterogeneous tasks are assigned. , Indicates the number of distributed heterogeneous tasks; This represents the total number of individuals in the population, with each individual representing a possible distributed heterogeneous task scheduling scheme.

[0035] Based on the overall power consumption planning model, the fitness value of each individual is calculated and expressed as:

[0036] ;

[0037] In the formula, Indicates the first The fitness value of each individual; express The resulting floating power consumption; Indicates the first The runtime of a distributed heterogeneous task.

[0038] Furthermore, the specific process of step S3.3 is as follows: calculate the probability of each individual being selected in the current population and the cumulative probability of each individual being selected, expressed as:

[0039] ;

[0040] ;

[0041] In the formula, Indicates the first The probability of an individual being selected; Indicates the first The cumulative probability of each individual being selected is used to determine the individual's position in the roulette wheel selection; Indicates the first in the population The individual, that is, the first individual. A distributed heterogeneous task scheduling scheme; Indicates the first The probability of an individual being selected, that is, the ratio of the individual's fitness value to the total fitness value of the population; This represents the sum of fitness values ​​of all individuals in the population, used to normalize probabilities.

[0042] Generate probabilistic random numbers Choose to satisfy individual , Indicates the first The cumulative probability of each individual being selected;

[0043] From satisfaction individual Two individuals are randomly selected and their fitness values ​​are compared. The individual with the lower fitness value is selected.

[0044] Furthermore, the specific process of step S3.4 is as follows:

[0045] By calculating the standard deviation of the population To measure population diversity and then dynamically adjust the crossover probability. , represented as:

[0046] ;

[0047] ;

[0048] In the formula, in the formula, This represents the average fitness value of the population. A threshold representing population diversity, used to determine whether a population is too concentrated or too dispersed. This represents a temperature parameter used to control the rate of change of the crossover probability; The base of the natural logarithm;

[0049] After calculating the crossover probability, pairwise comparisons are performed on the individuals selected by the selection strategy in the order of the population individuals; for each pair of individuals, a first random number between 0 and 1 is generated. Second random number and the third random number ,when Less than Then, perform crossover operations on this pair of individuals, and simultaneously, when and satisfy , Given the individual encoding length, then this pair of individuals will be... arrive The codes between them are cross-interchangeable to complete the individual crossover;

[0050] For mutation operations, power-sensitive quantized mutation is used, and the mutation strength is calculated as follows:

[0051] ;

[0052] In the formula, Indicates the first The variable and asynchronous length of each individual; Indicates the first The total power consumption of each individual unit, including operating power consumption and standby power consumption; , These are the upper and lower boundaries of the population power consumption, i.e., the maximum and minimum power consumption of all individuals in the current population. This indicates the maximum variation length, i.e., the maximum adjustment range of the mutation operation;

[0053] Simultaneously, if the best individual in the current population remains unchanged for several consecutive generations, tunneling mutation is triggered, and the probability of tunneling mutation is... The calculation method is expressed as follows:

[0054] ;

[0055] ;

[0056] In the formula, It represents potential energy, which is the ratio of an individual's current power consumption to its energy efficiency; Represents kinetic energy, indicating the rate of dynamic change of a population; This represents the scaling factor, used to adjust the magnitude of the kinetic energy. Indicates the current iteration number; Indicates the first The standard deviation of population fitness at the next iteration;

[0057] Generate the fourth random number ,when Less than This triggers tunneling mutation, at which point the mutation length is determined according to the variable time. The mutation process involves altering gene values ​​to complete an individual mutation.

[0058] Furthermore, the specific process of using a local search optimization mechanism to optimize the current best individual is as follows:

[0059] In each iteration, select the individual with the lowest fitness in the current population. Perform a local search; the local search optimizes the individual's fitness value by adjusting task allocation;

[0060] For the individual with the lowest fitness in the current population Randomly select a distributed heterogeneous task and a server Distributed heterogeneous tasks Reassigned to server And calculate the redistribution fitness value , is represented as:

[0061] ;

[0062] ;

[0063] In the formula, Indicates the redistributed ; Indicates the first A distributed heterogeneous task in Servers allocated in China ;

[0064] when ,take over A redistribution scheme; otherwise, Restore to its original state; repeat the above steps. Second-rate; express The fitness value.

[0065] A distributed heterogeneous task scheduling system based on an evolutionary algorithm includes:

[0066] The analysis module is used to analyze and quantify the influencing factors of distributed heterogeneous task scheduling;

[0067] The model building module is used to take the influencing factors of heterogeneous chips, servers, and quantified distributed heterogeneous task scheduling as the objective function for optimization and establish an overall power consumption planning model.

[0068] The solution module is used to optimize the scheduling of distributed heterogeneous tasks by using an improved evolutionary optimization algorithm and an overall power consumption planning model as the evaluation basis, so as to obtain the optimal scheduling scheme.

[0069] Compared with existing technologies, the present invention has the following advantages:

[0070] (1) This invention, through an improved evolutionary algorithm, combined with an overall power consumption planning model and dynamic floating power consumption calculation, can quickly converge to the optimal solution in large-scale data scenarios, thereby improving the efficiency and stability of scheduling. This invention not only considers task completion time, but also takes into account the overall power consumption of the server, including standby power consumption and operating power consumption. By dynamically adjusting the floating power consumption range and optimizing the task allocation strategy, a balance between performance and energy consumption is achieved. At the same time, in view of the complexity of heterogeneous computing platforms, this invention constructs a flexible and reliable energy consumption prediction model that can adapt to the addition of different heterogeneous devices and provide accurate energy consumption assessment for task scheduling.

[0071] (2) This invention incorporates the overall power consumption of the server, including standby power consumption and operating power consumption, into the optimization objective, enabling a comprehensive assessment of the impact of task scheduling on energy consumption. It provides a comprehensive energy consumption assessment framework for task scheduling, enabling dynamic adjustment of power allocation to adapt to the characteristics of different tasks and devices. By introducing a dynamic floating power consumption range to address power consumption changes during heterogeneous chip operation, it can more accurately reflect energy consumption fluctuations during task execution, avoiding energy consumption estimation errors caused by data read / write conflicts or queuing. This improves the flexibility and accuracy of energy consumption estimation, adapting to the characteristics of different heterogeneous devices and providing a more reliable basis for optimized scheduling.

[0072] (3) This invention significantly improves the performance and efficiency of the distributed heterogeneous task scheduling algorithm by introducing adaptive crossover and mutation probabilities and local search optimization. The adaptive crossover and mutation probability mechanism dynamically adjusts the operation intensity according to the population diversity, avoiding premature convergence or getting trapped in local optima due to fixed probabilities, thereby enhancing the global search capability and convergence speed. Local search optimization further optimizes the task allocation of the current best individual, and further reduces energy consumption and improves the quality of solutions by fine-tuning the exploration of the neighborhood solution space. This enables the algorithm to efficiently find the global optimal solution in large-scale distributed heterogeneous task scheduling, while taking into account energy consumption optimization and task completion time, thus improving the overall performance and adaptability of the system. Attached Figure Description

[0073] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0074] like Figure 1 As shown, the present invention provides a technical solution: a distributed heterogeneous task scheduling method based on an evolutionary algorithm, comprising the following steps:

[0075] Step S1: Analyze and quantify the influencing factors of distributed heterogeneous task scheduling.

[0076] Among them, the influencing factors of distributed heterogeneous task scheduling include the overall standby power consumption of the server and the power consumption increase brought about by the operation; the overall standby power consumption of the server is divided into the power consumption brought about by hardware devices and the standby power consumption of heterogeneous chips, and the power consumption increase brought about by the operation of the server includes the power consumption increase brought about by heterogeneous chip computing and the power consumption increase brought about by adding containers.

[0077] The quantification process includes:

[0078] Data collection involves gathering relevant data by consulting official statistics, research reports, energy consumption records, and other sources.

[0079] Data standardization involves standardizing the collected data to facilitate comparisons of data from different times and locations.

[0080] Data transformation converts standardized data into numerical values ​​that can be used for subsequent model calculations, such as scaling data to a specific range through normalization.

[0081] Data validation, through cross-validation or other methods, ensures the accuracy and reliability of the transformed data.

[0082] Step S2: Using the influencing factors of heterogeneous chips, servers, and quantified distributed heterogeneous task scheduling as the objective function for optimization, establish an overall power consumption planning model.

[0083] For each distributed heterogeneous task, floating power consumption is calculated. Floating power consumption includes power consumption caused by heterogeneous chip operation, data conflicts, queuing, and other reasons.

[0084] Once heterogeneous chips begin operation, the increased frequency due to computation leads to increased power consumption. This increased power consumption varies significantly depending on the specific heterogeneous chip, and the performance loss is not linear. Due to factors such as data read / write conflicts and computer system queuing mechanisms, this power consumption needs a reasonable fluctuation range. This fluctuation range can be dynamically adjusted based on the number of heterogeneous chips used to ensure that the fluctuation value remains within a reasonable range. Represented as:

[0085] ;

[0086] In the formula, Indicates the first The overhead of completing a distributed heterogeneous task; This indicates the server's power consumption fluctuation value; Indicates the first The runtime required for a distributed heterogeneous task; Indicates the first The number of heterogeneous chips required for distributed heterogeneous tasks. This indicates the number of distributed heterogeneous tasks.

[0087] Meanwhile, the power consumption generated by the server in standby mode while it is powered on is also considered. Under certain conditions, as long as there are no large-scale changes to the server, this part of the power consumption is basically within a stable range with minimal fluctuations. Therefore, in the calculation process, this part of the fluctuation is integrated into the power consumption model part of the heterogeneous chip operation. The power consumption generated by the server in standby mode while it is powered on includes the basic power consumption of the heterogeneous chip and the power consumption of peripheral components.

[0088] Heterogeneous chip base power consumption refers to the power consumption of each heterogeneous chip in a server in standby mode, expressed as:

[0089] ;

[0090] In the formula, Indicates the base power consumption of heterogeneous chips; Indicates the operating voltage of the heterogeneous chip; This indicates the operating current in standby mode; This represents the longest runtime required to complete all distributed heterogeneous tasks.

[0091] The power consumption of peripheral components includes standby power consumption of memory devices, standby power consumption of storage devices, and fixed overhead; the standby power consumption of memory devices is expressed as:

[0092] ;

[0093] In the formula, This indicates the standby power consumption of the memory device. Indicates the operating voltage of the memory device. This represents the operating current of a single memory device in standby mode. Indicates the number of memory devices.

[0094] The standby power consumption of a storage device is expressed as:

[0095] ;

[0096] In the formula, This indicates the standby power consumption of the storage device. This indicates the operating voltage of the storage device. This represents the operating current of a single storage device in standby mode. This indicates the number of storage devices.

[0097] Considering the overall energy efficiency of the server, the overall power consumption of the server is obtained by integrating the floating power consumption, the basic power consumption of heterogeneous chips, and the power consumption of peripheral components. :

[0098] ;

[0099] In the formula, Indicates floating power consumption. Indicates the base power consumption of heterogeneous chips. This indicates the standby power consumption of the memory device. This indicates the standby power consumption of the storage device. This indicates fixed overhead, specifically the fixed energy consumption of components such as the motherboard and low-speed fans. This indicates the system's energy efficiency.

[0100] Step S3: Using an improved evolutionary optimization algorithm and an overall power consumption planning model as the evaluation basis, the scheduling of distributed heterogeneous tasks is optimized to obtain the optimal scheduling scheme.

[0101] Step S3.1: Initialize the population for the improved evolutionary optimization algorithm.

[0102] Randomly generated Each individual represents a possible distributed heterogeneous task scheduling scheme.

[0103] Population of improved evolutionary optimization algorithms Represented as:

[0104] ;

[0105] In the formula, Indicates the first in the population The position of each individual , Indicates the first The servers to which distributed heterogeneous tasks are assigned. , This indicates the number of distributed heterogeneous tasks.

[0106] Step S3.2: The fitness value is used to evaluate the quality of each individual. A lower fitness value indicates lower energy consumption, while a higher fitness value indicates higher energy consumption. Based on the overall power consumption planning model, the fitness value of each individual is calculated and expressed as:

[0107] ;

[0108] In the formula, Indicates the first The fitness value of each individual; express The resulting floating power consumption; Indicates the first The runtime of a distributed heterogeneous task.

[0109] Step S3.3: Select individuals in the current population using a selection strategy that combines roulette wheel selection with random competitive selection.

[0110] Fitness percentage is directly proportional to the probability of an individual being selected; the probability of each individual being selected in the current population and the cumulative probability of each individual being selected are expressed as:

[0111] ;

[0112] ;

[0113] In the formula, Indicates the first The probability of an individual being selected; Indicates the first The cumulative probability of each individual being selected is used to determine the individual's position in the roulette wheel selection; Indicates the first in the population The individual, that is, the first individual. A distributed heterogeneous task scheduling scheme; Indicates the first The probability of an individual being selected, that is, the ratio of the individual's fitness value to the total fitness value of the population; This represents the sum of fitness values ​​of all individuals in the population, used to normalize probabilities.

[0114] Generate probabilistic random numbers Choose to satisfy individual , Indicates the first The cumulative probability of each individual being selected;

[0115] From satisfaction individual Two individuals are randomly selected and their fitness values ​​are compared. The individual with the lower fitness value is selected to proceed to the next iteration.

[0116] Step S3.4: Update the individual position through a sensed pulse evolution strategy, which includes an energy-efficient crossover strategy and a power-sensitive quantum mutation.

[0117] Crossover and mutation are crucial operators in evolutionary algorithms, primarily ensuring that the algorithm doesn't converge too quickly and fail to reach the global optimum. In traditional evolutionary algorithms, crossover and mutation probabilities are typically fixed, limiting the algorithm's optimization capabilities at different stages. To overcome this problem, this invention introduces a perceptual spurious evolution strategy, enabling the evolutionary algorithm to dynamically adjust the intensity of operations based on population diversity.

[0118] By calculating the standard deviation of the population To measure population diversity and then dynamically adjust the crossover probability. A larger standard deviation indicates higher population diversity, in which case increasing the crossover probability can facilitate global search; conversely, decreasing the crossover probability can avoid unnecessary computational overhead.

[0119] ;

[0120] ;

[0121] In the formula, This represents the average fitness value of the population. A threshold representing population diversity, used to determine whether a population is too concentrated or too dispersed. This represents a temperature parameter used to control the rate of change of the crossover probability; It represents the base of the natural logarithm.

[0122] After calculating the crossover probability, pairwise comparisons are performed on the individuals selected by the selection strategy in the order of the population individuals; for each pair of individuals, a first random number between 0 and 1 is generated. Second random number and the third random number ,like Less than Then, perform a crossover operation on this pair of individuals. Meanwhile, if... and satisfy ( (for the individual encoding length), then this pair of individuals will be in arrive The codes between them are cross-interchangeable, thereby completing the individual crossover.

[0123] For mutation operations, power-sensitive quantum mutation is used. Individuals with higher power consumption exhibit greater mutation strength. There is a greater tendency to perform larger mutation operations on high-power individuals to guide the population to evolve towards lower power consumption. The mutation strength is calculated as follows:

[0124] ;

[0125] In the formula, Indicates the first The variable and asynchronous length of each individual; Indicates the first The total power consumption of each individual unit, including operating power consumption and standby power consumption; , These are the upper and lower boundaries of the population power consumption, i.e., the maximum and minimum power consumption of all individuals in the current population. This indicates the maximum variable length, i.e., the maximum adjustment range of the mutation operation.

[0126] If the best individual in the current population remains unchanged for several consecutive generations, tunneling mutation is triggered. The probability of tunneling mutation is... The calculation method is expressed as follows:

[0127] ;

[0128] ;

[0129] In the formula, It represents potential energy, which is the ratio of an individual's current power consumption to its energy efficiency; Represents kinetic energy, indicating the rate of dynamic change of a population; This represents the scaling factor, used to adjust the magnitude of the kinetic energy. Indicates the current iteration number; Indicates the first The standard deviation of population fitness at each iteration.

[0130] Generate the fourth random number ,like Less than This triggers tunneling mutation, at which point the mutation length is determined according to the variable time. The mutation process involves altering gene values ​​to complete an individual mutation.

[0131] Step S3.5: To further improve the algorithm's optimization performance and avoid getting trapped in local optima, a local search optimization mechanism is used to optimize the current best individual. The purpose of local search optimization is to fine-tune the current best individual to explore its neighborhood solution space and find a better solution. This enhances the algorithm's search capability in the local solution space and can effectively improve the quality of the solution.

[0132] In each iteration, select the individual with the lowest fitness in the current population. A local search is performed. Local search optimizes the fitness value of an individual by adjusting the distribution of distributed heterogeneous tasks.

[0133] For the individual with the lowest fitness in the current population Randomly select a distributed heterogeneous task and a server Distributed heterogeneous tasks Reassigned to server And calculate the redistribution fitness value ;when ,take over A redistribution scheme; otherwise, Restore to its original state; repeat the above steps. Second-rate; express The fitness value.

[0134] ;

[0135] ;

[0136] In the formula, Indicates the redistributed ; Indicates the first A distributed heterogeneous task in Servers allocated in China .

[0137] Step S3.6: Determine if the current iteration meets the termination condition (preset maximum number of iterations). If not, continue iterating; if yes, stop iterating and output the best individual in the current population (the individual with the lowest fitness). ( ) is the optimal distributed heterogeneous task scheduling scheme.

[0138] A distributed heterogeneous task scheduling system based on an evolutionary algorithm includes:

[0139] The analysis module is used to analyze and quantify the influencing factors of distributed heterogeneous task scheduling.

[0140] The model building module is used to establish an overall power consumption planning model by taking the influencing factors of heterogeneous chips, servers, and quantified distributed heterogeneous task scheduling as the objective function for optimization.

[0141] The solution module is used to optimize the scheduling of distributed heterogeneous tasks by using an improved evolutionary optimization algorithm and an overall power consumption planning model as the evaluation basis, so as to obtain the optimal scheduling scheme.

[0142] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for distributed heterogeneous task scheduling based on evolutionary algorithm, characterized in that, Comprising the following steps: Step S1: analyze the influencing factors of distributed heterogeneous task scheduling and quantify; Step S2: take the heterogeneous chips, servers and the quantified influencing factors of distributed heterogeneous task scheduling as the objective function of optimization, and establish a whole power consumption planning model; Step S3: use an optimization algorithm to optimize the scheduling of distributed heterogeneous tasks based on the whole power consumption planning model, and obtain an optimal scheduling scheme; The whole power consumption planning model is expressed as: ； wherein, represents the total power consumption of the server; represents the floating power consumption, represents the heterogeneous chip base power consumption, represents the memory device standby power consumption, represents the storage device standby power consumption, represents the fixed overhead; represents the system energy efficiency; The optimization algorithm uses an improved evolutionary algorithm. The specific process of using the improved evolutionary optimization algorithm to optimize the scheduling of distributed heterogeneous tasks based on the whole power consumption planning model to obtain an optimal scheduling scheme is as follows: Step S3.1: initialize the population of the improved evolutionary optimization algorithm; Step S3.2: calculate the fitness value of each individual according to the whole power consumption planning model; Step S3.3: select individuals in the current population through a selection strategy, which uses a combination of roulette selection and random competition selection; Step S3.4: update the individual position through a perception-based pulse evolution strategy, which includes an energy efficiency-driven crossover strategy and a power consumption-sensitive quantization mutation; Step S3.5: optimize the current optimal individual using a local search optimization mechanism; Step S3.6: determine whether the current iteration meets the termination condition. If not, continue iteration; if yes, stop iteration and output the optimal individual in the current population as the optimal task scheduling scheme; Floating power dissipation is represented as: ； In the formula, represents the consumption brought by the first distributed heterogeneous task; represents the server power consumption floating value; represents the running time required by the first distributed heterogeneous task; represents the running time required by the first distributed heterogeneous task; represents the number of distributed heterogeneous tasks; Heterogeneous chip base power consumption is represented as: ； wherein represents the heterogeneous chip base power consumption; represents the heterogeneous chip operating voltage; represents the operating current in standby state; represents the longest running time needed to complete all distributed heterogeneous tasks; Memory device standby power consumption is represented as: ； wherein represents the standby power consumption of the memory device, represents the operating voltage of the memory device, represents the operating current of the memory device in the standby state, represents the number of memory devices; The standby power consumption of the storage device is expressed as: ； wherein represents the standby power consumption of the storage device, represents the operating voltage of the storage device, represents the operating current of the storage device in the standby state of the single storage device, represents the number of the storage devices; Population of an improved evolutionary optimization algorithm is represented as: ； wherein, represents the position of the i-th individual in the population, , represents the i-th server to which a distributed heterogeneous task is assigned, , represents the number of distributed heterogeneous tasks; represents the total number of individuals in the population, each individual representing a possible distributed heterogeneous task scheduling scheme;​​ According to the whole power consumption planning model, the fitness value of each individual is calculated, which is expressed as: ； wherein represents the fitness value of the th individual; represents the generated floating power consumption; represents the running time of the th distributed heterogeneous task; The specific process of step S3.3 is to calculate the probability of each individual in the current population being selected and the cumulative probability of each individual being selected, which is expressed as: ； ； In the formula, Indicates the first The probability of an individual being selected; Indicates the first The cumulative probability of each individual being selected is used to determine the individual's position in the roulette wheel selection; Indicates the first in the population The individual, that is, the first individual A distributed heterogeneous task scheduling scheme; Indicates the first The probability of an individual being selected, that is, the ratio of the individual's fitness value to the total fitness value of the population; This represents the sum of fitness values ​​of all individuals in the population, used to normalize probabilities. Generating probabilistic random numbers , selecting individuals satisfying , denotes the cumulative probability of the th individual being selected;​ from satisfying the individual two individuals are randomly selected from the individuals satisfying the condition and the individual having a lower fitness value is selected by comparing the fitness values. The specific process of step S3.4 is: By calculating the standard deviation of the population to measure the diversity of the population, and then dynamically adjust the crossover probability is expressed as: ； ； In the formula, In the formula, represents the average fitness value of the population, represents the threshold value of population diversity, used to determine whether the population is too concentrated or dispersed, represents the temperature parameter, used to control the change speed of the crossover probability; represents the base number of natural logarithm; After calculating the crossover probability, pairwise comparisons are performed on the individuals selected by the selection strategy in the order of the population individuals; for each pair of individuals, a first random number between 0 and 1 is generated. Second random number and the third random number ,when Less than Then, perform crossover operations on this pair of individuals, and simultaneously, when and satisfy , Given the individual encoding length, then this pair of individuals will be... arrive The codes between them are cross-interchangeable to complete the individual crossover; For mutation operation, use power consumption-sensitive quantization mutation, and the mutation strength calculation is expressed as: ； In the formula, represents the maximum variation step length of the i-th individual; represents the total power consumption of the i-th individual, including the running power consumption and the standby power consumption; , are the upper and lower boundaries of the population power consumption, i.e. the maximum and minimum power consumption of all individuals in the current population; represents the maximum variation step length, i.e. the maximum adjustment range of the mutation operation;​​ Meanwhile, when the optimal individual in the current population does not change in continuous several generations of iterations, the tunneling mutation is triggered, and the tunneling mutation probability The calculation method is represented as: ； ； wherein, represents potential energy, i.e. the ratio of power consumption and energy efficiency of the current individual; represents kinetic energy, i.e. the dynamic change rate of the population; represents a scaling factor for adjusting the magnitude of the kinetic energy; represents the current iteration number; represents the standard deviation of the population fitness at the th iteration; Generating a fourth random number When Less than Then trigger tunneling mutation, at this time according to the mutation step Perform mutation, change gene value, complete individual mutation.

2. The method of claim 1, wherein the method is characterized by: The specific process of optimizing the current optimal individual using a local search optimization mechanism is: In each iteration, the individual with the lowest fitness in the current population is selected A local search is performed; the local search optimizes the fitness value of the individual by adjusting the task assignments; for the individual with the lowest fitness in the current population , randomly selecting a distributed heterogeneous task and a server , reassigning the distributed heterogeneous task to the server and calculating a fitness value after the reassignment , denoted as: ； ； In the formula, represents the re-allocated ; represents the first distributed heterogeneous task allocated to the server in ; ; When , receive the re-allocated scheme; otherwise, restore the original; repeat the above operation times; indicates the fitness value.

3. A distributed heterogeneous task scheduling system based on evolutionary algorithm, applied to the distributed heterogeneous task scheduling method based on evolutionary algorithm in any one of claims 1-2, characterized in that, Comprise: An analysis module for analyzing the influencing factors of distributed heterogeneous task scheduling and quantifying; A model construction module for taking the heterogeneous chips, servers and the quantified influencing factors of distributed heterogeneous task scheduling as the objective function of optimization, and establishing a whole power consumption planning model; A solution module for using an improved evolutionary optimization algorithm to optimize the scheduling of distributed heterogeneous tasks based on the whole power consumption planning model, and obtaining an optimal scheduling scheme.

Images