Low-delay task scheduling system for edge computing nodes based on improved swarm intelligence algorithm

Abstract

The application relates to the technical field of edge computing, and discloses an edge computing node low-delay task scheduling system based on an improved swarm intelligence algorithm, which comprises the following steps: obtaining micro-operation parameters to extract transient heat flux; obtaining a shared memory controller pre-play temperature through pre-play thermal wave interference; mapping the pre-play temperature into dynamic available memory bandwidth by using a thermal scattering attenuation law; calculating the real execution time of a task and constructing a scheduling fitness evaluation value on the basis of the bandwidth; initializing particle swarm position probability by using a benchmark delay reciprocal and setting initial evolution speed to zero; driving particle optimization by using the ratio of the fitness evaluation value and the ratio of the initial temperature and the pre-play temperature; converting the optimal scheduling solution into a deterministic assignment; and updating the starting core temperature of the next batch by using a thermal index attenuation law. The application effectively reduces the long tail delay of the computing node in a closed environment.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and more specifically, to a low-latency task scheduling system for edge computing nodes based on an improved swarm intelligence algorithm. Background Technology

[0002] Edge computing systems are often deployed in special application scenarios such as deep, high-gas mines to perform tasks such as high-frequency visual servo obstacle avoidance for mining equipment. To meet stringent explosion-proof requirements, these computing nodes typically employ a fully sealed encapsulation system, with high-density packaging of internal components and an explicit prohibition on the use of active cooling mechanisms. When a neural network processor intensively processes multiple high-resolution visual information streams, it generates a significant amount of heat. Due to the low thermal conductivity of the external sealing material, heat is rapidly transferred laterally inward, causing a sudden temperature rise in the adjacent shared memory controller. To protect the underlying hardware, the system's temperature protection mechanism is triggered, thereby reducing the operating frequency and available bandwidth of the memory bus.

[0003] Existing task scheduling systems and conventional swarm intelligence scheduling algorithms typically allocate tasks based solely on processor utilization or task queue status. In situations where internal heat buildup leads to a hidden decrease in global memory bandwidth, assigning time-sensitive computational tasks to a seemingly idle central processing unit (CPU) can cause severe data congestion when the CPU tries to access memory. This traditional scheduling approach, which fails to consider the dynamic changes in memory bandwidth in conjunction with heat dissipation, results in significant system latency, failing to meet the demands of microsecond-level high-frequency response and ultimately impacting the overall reliability of the device. Summary of the Invention

[0004] This invention provides a low-latency task scheduling system for edge computing nodes based on an improved swarm intelligence algorithm, which solves the technical problems mentioned in the background art.

[0005] This invention provides a low-latency task scheduling system for edge computing nodes based on an improved swarm intelligence algorithm. It is applied to edge computing nodes comprising a computing core consisting of a central processing unit and a neural network processor, a shared memory controller, and an explosion-proof sealing system, and is configured to execute: The absolute initial temperature of the environment is obtained, and the microscopic operating parameters of the calculation core are extracted to calculate the transient heat flux. The pre-simulated memory temperature of the shared memory controller is calculated based on the transient heat flux pre-simulated spatial thermal wave interferometry. The pre-simulated memory temperature is mapped to dynamic available memory bandwidth using the thermal scattering attenuation law; The actual execution time of the task is calculated based on the dynamically available memory bandwidth, and a scheduling fitness evaluation value is constructed in combination with the system energy consumption. The initial position probability of the particle swarm is initialized using the inverse ratio of the single-task baseline delay, and the initial evolution velocity is set to zero. The dynamic inertia weight is calculated based on the ratio of the scheduling fitness evaluation value, and the individual cognitive factor is calculated based on the ratio of the absolute initial temperature of the environment to the pre-simulation memory temperature, driving the particles to iterate towards the global optimal scheduling solution. The global optimal scheduling solution is transformed into deterministic assignment actions, and the starting core temperature of the next batch is updated using the thermal exponential decay law.

[0006] Furthermore, the step of extracting the microscopic operating parameters of the computational core to calculate the transient heat flux includes: Obtain the single-cycle equivalent dynamic capacitance, transient supply voltage, and transient clock frequency of each computing core; The transient power supply voltage of each of the aforementioned calculation cores is squared to obtain the voltage square value; The transient heat flux of the corresponding computing core is calculated by continuously multiplying the single-cycle equivalent dynamic capacitance, the squared voltage value, and the transient clock frequency of each computing core.

[0007] Furthermore, the step of calculating the simulated memory temperature of the shared memory controller includes: The total number of instructions corresponding to the task, the instruction throughput per cycle of the corresponding computing core, the memory thermal capacity of the shared memory controller, the spatial thermal decay constant, and the current temperature of the shared memory are obtained. Divide the total number of instructions by the product of the transient clock frequency of the corresponding computing core and the instruction throughput per cycle, and then multiply by the transient heat flux to obtain the heat dissipation of a single task. Multiply the particle's current position probability by the corresponding heat dissipation of a single task, and sum the results for all tasks assigned to the same computing core to obtain the total core heat dissipation; divide the total core heat dissipation by the memory heat capacity to obtain the base temperature rise. By combining the aforementioned spatial thermal decay constant, an exponential decay calculation is performed on the base temperature rise to obtain the decayed temperature rise. The pre-simulated memory temperature is obtained by summing the temperature rises after decay of all the computing cores and then adding the current shared memory temperature.

[0008] Furthermore, the step of mapping the simulated memory temperature to dynamically available memory bandwidth includes: Obtain the system's factory-set maximum bandwidth; Divide the absolute initial temperature of the environment by the pre-simulated memory temperature to obtain the absolute temperature quotient. The nonlinear decay rate is obtained by performing a cube root operation on the absolute temperature quotient and then taking the square root. The dynamic available memory bandwidth is obtained by multiplying the factory-set maximum bandwidth by the nonlinear attenuation rate.

[0009] Furthermore, the steps for constructing the scheduling fitness evaluation value include: Get the memory access data volume of the tasks to be assigned; The data blocking time is obtained by dividing the amount of memory accessed by the dynamic available memory bandwidth; the computation time is obtained by dividing the total number of instructions by the product of the transient clock frequency and the instruction throughput per cycle; the actual execution time of the task is obtained by adding the data blocking time and the computation time. Multiply the current position probability by the actual execution time of the task and sum the results across all tasks. Take the maximum value among the calculation cores to obtain the global execution time extreme value. The system energy consumption is obtained by multiplying the current location probability by the heat emitted by a single task, then summing the results by iterating through all tasks and each computing core. The scheduling fitness evaluation value is obtained by multiplying the extreme value of the global execution time by the system energy consumption.

[0010] Furthermore, the initialization steps include: The single-task baseline latency of the computing task under the condition that no thermal latency occurs on the corresponding computing core; The single-task response rate is obtained by taking the reciprocal of the single-task baseline delay corresponding to the task to be assigned. The initial position probability assigned to the particle is obtained by dividing the single-task response rate of the task to be assigned on the current computing core by the sum of the single-task response rates of the task on all computing cores. The initial evolution velocity of the particles is set to zero.

[0011] Furthermore, the steps to drive the particles to iterate towards the globally optimal scheduling solution include: Obtain the current evolution velocity of the particle and the individual historical optimal solution, and read the first voltage noise floor value and the second voltage noise floor value of the underlying system; The dynamic inertia weight is obtained by dividing the minimum value of the scheduling fitness evaluation value of all particles in the current iteration by the scheduling fitness evaluation value of the current particle. The individual cognitive factor is obtained by dividing the absolute initial temperature of the environment by the pre-simulation memory temperature of the current iteration round, and the social cognitive factor is obtained by subtracting the individual cognitive factor from one. The inertial component is obtained by multiplying the dynamic inertial weight by the current evolution velocity; the self-cognition component is obtained by continuously multiplying the individual cognitive factor, the first voltage noise floor value, and the difference between the individual's historical optimal solution and the current position probability; the social cognitive component is obtained by continuously multiplying the social cognitive factor, the second voltage noise floor value, and the difference between the current group's global optimal solution and the current position probability. The inertial component, the self-cognition component, and the social cognition component are added together to obtain the updated evolution velocity. The updated position probability is obtained by combining the evolution position normalization rule. The converged global optimal scheduling solution is output through iterative loops.

[0012] Further steps to derive the initial core temperature for the next batch include: Obtain the core thermal resistance, core thermal capacity, and current absolute temperature of the corresponding computing core; The computational core that maximizes the probability in the global optimal scheduling solution is extracted as the target, and the deterministic assignment action is constituted. The actual execution times of all tasks assigned to the same computing core are summed, and the maximum value among all computing cores is taken as the actual execution window delay. The sum of the heat dissipated by the single task assigned to the current computing core is divided by the core's heat capacity to obtain the new temperature rise. The current absolute core temperature is added to the new temperature rise and the ambient absolute initial temperature is subtracted to obtain the transient temperature difference. Calculate the cooling attenuation term, which is the ratio of the actual execution window delay to the product of the core thermal resistance and the core thermal capacity, with the natural constant as the base and the exponent being negative; Multiply the transient temperature difference by the cooling attenuation term, and finally add the ambient absolute initial temperature to obtain the starting core temperature of the corresponding computational core in the next batch.

[0013] The beneficial effects of this invention are as follows: By extracting the microscopic operating parameters of the computing core to obtain transient heat flux, and by pre-simulating spatial thermal wave interferometry to evaluate the temperature state of the shared memory controller, it is mapped to dynamically available memory bandwidth, and a fitness evaluation value including latency and energy consumption is constructed. At the same time, the algorithm initializes the position probability with the inverse ratio of the baseline latency, drives the optimization iteration with the ratio of the initial ambient temperature to the pre-simulated temperature, and finally transforms the global optimal solution into a deterministic assignment, and updates the initial core temperature according to the thermal exponential decay law. This method effectively overcomes the problem of implicit memory bandwidth reduction caused by heat accumulation in computing nodes in a closed environment without heat dissipation, reduces the long-tail latency caused by processor waiting for data, and improves the real-time performance of task processing and system operation stability in high-frequency and high-load application scenarios. Attached Figure Description

[0014] Figure 1 This is a flowchart of the low-latency task scheduling system for edge computing nodes based on the improved swarm intelligence algorithm of the present invention. Detailed Implementation

[0015] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0016] like Figure 1 As shown, a low-latency task scheduling system for edge computing nodes based on an improved swarm intelligence algorithm is applied to an edge computing node that includes a computing core consisting of a central processing unit and a neural network processor, a shared memory controller, and an explosion-proof sealing system. The system is configured to execute: The absolute initial temperature of the environment is obtained, and the microscopic operating parameters of the calculation core are extracted to calculate the transient heat flux. The predicted memory temperature of the shared memory controller is calculated based on the predicted spatial thermal wave interferometry of the transient heat flux. The pre-simulated memory temperature is mapped to dynamic available memory bandwidth using the thermal scattering attenuation law; The actual execution time of the task is calculated based on the dynamically available memory bandwidth, and a scheduling fitness evaluation value is constructed in combination with the system energy consumption. The initial position probability of the particle swarm is initialized using the inverse ratio of the single-task baseline delay, and the initial evolution velocity is set to zero. The dynamic inertia weight is calculated based on the ratio of the scheduling fitness evaluation value, and the individual cognitive factor is calculated based on the ratio of the absolute initial temperature of the environment to the pre-simulation memory temperature, driving the particles to iterate towards the global optimal scheduling solution. The global optimal scheduling solution is transformed into deterministic assignment actions, and the starting core temperature of the next batch is updated using the thermal exponential decay law.

[0017] The absolute initial ambient temperature is obtained. This absolute initial ambient temperature is the thermodynamic temperature of the environment in which the edge computing node is located. It can be acquired in real time using a digital temperature sensor built into the node with a sampling period of 10 milliseconds to 1 second. The sampling period can be adjusted according to system response requirements. The unit is Kelvin. During system initialization, the ambient temperature sensor needs to be calibrated at three points: 273.15 Kelvin, 303.15 Kelvin, and 333.15 Kelvin. The calibration error should not exceed 0.5 Kelvin.

[0018] The microscopic operating parameters of the computational core are extracted to calculate the transient heat flux. The specific steps are as follows: Obtain the single-cycle equivalent dynamic capacitance, transient supply voltage, and transient clock frequency for each computing core. The single-cycle equivalent dynamic capacitance is the equivalent capacitance of the computing core during logic switching within one clock cycle. It can be obtained directly from the chip manufacturer's technical datasheet or measured under standard operating conditions using a chip power consumption test platform; the unit is farads. The transient supply voltage is the actual operating voltage of the computing core at the current moment. It can be read in real-time through the internal voltage monitoring register of the system power management unit at a sampling period of 1 microsecond; the unit is volts. The transient clock frequency is the operating clock frequency of the computing core at the current moment. It can be read in real-time through the frequency status register of the system clock controller; the unit is hertz.

[0019] The transient supply voltage of each computing core is squared to obtain the voltage square value. , For the first Each computing core at time... The transient supply voltage.

[0020] The transient heat flux of the corresponding computing core is calculated by continuously multiplying the single-cycle equivalent dynamic capacitance, the square of the voltage, and the transient clock frequency of each computing core. The calculation formula is as follows: In the formula, For the first Each computing core at time... Transient heat flux, measured in watts; For the first The single-cycle equivalent dynamic capacitance of a computing core; For the first Each computing core at time... The transient clock frequency.

[0021] Based on the transient heat flux prediction of spatial thermal wave interferometry, the predicted memory temperature of the shared memory controller is calculated. The specific steps are as follows: This section describes the process for obtaining the total number of instructions for a given task, the instruction throughput per cycle of the corresponding computing core, the memory thermal capacity of the shared memory controller, the space thermal decay constant, and the current temperature of the shared memory. The total number of instructions is the total number of binary instructions generated after compiling the task. This can be extracted from the task's executable file using static analysis tools or obtained from the mapping file generated by the compiler during the compilation phase. The instruction throughput per cycle is the maximum number of instructions that the computing core can execute in one clock cycle. This is an inherent performance parameter of the chip and can be obtained from the chip's technical datasheet. For superscalar processors, this value represents the number of instructions that can be issued per clock cycle. The memory thermal capacity is the amount of heat absorbed or released by the shared memory controller per unit temperature change. It can be calculated based on the shared memory controller's material composition, volume, and specific heat capacity, and is measured in joules per Kelvin. For silicon, the specific heat capacity is taken as 700 joules per kilogram per Kelvin. The spatial thermal decay constant characterizes the degree of heat attenuation during its propagation from the computing core to the shared memory controller within the chip. It is determined by the physical distance between the computing core and the shared memory controller, as well as the thermal conductivity of silicon. The thermal conductivity of silicon ranges from 148 watts per meter per Kelvin to 156 watts per meter per Kelvin, with a standard value of 150 watts per meter per Kelvin. The equivalent heat dissipation coefficient of the edge computing node within the current packaging system is obtained to characterize the ability of heat to diffuse from the chip's interior to the surrounding medium. Its unit is watts per square meter per Kelvin, and it can be pre-calibrated using thermal simulation software. The current shared memory temperature is the actual temperature of the shared memory controller at the current moment, which can be read in real-time using an on-chip temperature sensor integrated within the shared memory controller with a sampling period of 1 millisecond, and its unit is Kelvin.

[0022] Multiplying the transient clock frequency of the corresponding computing core by the instruction throughput per cycle yields the instruction processing rate per cycle, calculated as follows: In the formula, For the first Instruction throughput per cycle of each computing core.

[0023] Dividing the total number of instructions by the single-cycle instruction processing rate yields the theoretical execution time of a single task, calculated using the following formula: In the formula, For the first The total number of instructions for each task.

[0024] Multiplying the theoretical execution time of a single task by the transient heat flux yields the heat dissipation of a single task, calculated using the following formula: In the formula, For the first The task was assigned to the first The heat generated per computing core is measured in joules.

[0025] The particle's current position probability is multiplied by the corresponding single-task heat dissipation, and the sum is applied to all tasks assigned to the same computational core to obtain the total core heat dissipation. The particle's current position probability is the [missing value] in the particle swarm optimization algorithm. The i-th particle represents the first The task was assigned to the first The probability of each computational core is denoted as... The range of values ​​is And for any task The sum of the location probabilities corresponding to all computational cores is The number of particles in a particle swarm is typically between 20 and 100, and can be adjusted based on the number of tasks and computing cores. When the number of tasks is less than or equal to 10, the number of particles is 20; when the number of tasks is greater than 10 but less than or equal to 50, the number of particles is 50; and when the number of tasks is greater than 50, the number of particles is 100. The formula for calculating the total heat generated by the core is: In the formula, This represents the total number of tasks to be assigned.

[0026] Divide the total core heat generation by the memory heat capacity to obtain the base temperature rise. The calculation formula is as follows: In the formula, This refers to the hot capacity of the shared memory controller.

[0027] By combining the spatial thermal decay constant with the exponential decay calculation of the base temperature rise, the decayed temperature rise is obtained. The calculation formula is as follows: In the formula, For the first The physical distance between each computing core and the shared memory controller, in meters; The thermal conductivity of silicon is expressed in watts per meter per Kelvin. is the equivalent heat dissipation coefficient.

[0028] The predicted memory temperature is calculated by summing the temperature rises of all computing cores after decay, and then adding the current shared memory temperature. The formula is as follows: In the formula, For the first The pre-simulation memory temperature corresponding to each particle is expressed in Kelvin. Shared memory temperature at the current moment; To calculate the total number of cores.

[0029] The pre-simulated memory temperature is mapped to dynamically available memory bandwidth using the thermal scattering attenuation law. The specific steps are as follows: Obtain the system's factory-base maximum bandwidth. The factory-base maximum bandwidth is the maximum data transfer bandwidth that the shared memory controller can achieve at a standard operating temperature of 303.15 Kelvin. It is a nominal performance parameter of the chip and can be obtained from the chip's technical datasheet. The unit is bytes per second.

[0030] Divide the initial absolute ambient temperature by the pre-simulated memory temperature to obtain the absolute temperature quotient. The calculation formula is as follows: In the formula, The absolute initial temperature of the environment.

[0031] The nonlinear decay rate is obtained by taking the cube root of the absolute temperature quotient and then the square root. The calculation formula is as follows: Multiplying the factory-set maximum bandwidth by the non-linear attenuation rate yields the dynamic available memory bandwidth. The calculation formula is as follows: In the formula, For the first The dynamic available memory bandwidth corresponding to each particle; This is the system's factory-set maximum bandwidth. When the calculated dynamic available memory bandwidth is less than the system's minimum available bandwidth, the system's minimum available bandwidth is taken, which is 10% of the factory-set maximum bandwidth.

[0032] The actual execution time of the task is calculated based on the dynamically available memory bandwidth, and the scheduling fitness evaluation value is constructed by combining the system energy consumption. The specific steps are as follows: Obtain the memory access data volume of the task to be assigned. The task memory access data volume is the total number of bytes of data that need to be read or written from shared memory during task execution. It can be obtained by disassembling the task's executable file using static code analysis tools, or by calculating it during task execution using performance profiling tools. The unit is bytes. For tasks containing loop structures, it is necessary to calculate the product of the number of loop executions and the memory access data volume per loop.

[0033] The data blocking time is calculated by dividing the amount of data accessed by the dynamically available memory bandwidth. In the formula, For the first The amount of data accessed by each task.

[0034] The computation time is calculated by dividing the total number of instructions by the product of the transient clock frequency and the instruction throughput per cycle. The formula is as follows: The actual execution time of the task is obtained by adding the data blocking time to the computation time. The calculation formula is as follows: In the formula, For the first The task was assigned to the first The actual execution time per computing core, in seconds.

[0035] Multiply the current position probability by the actual task execution time, sum the results across all tasks, and take the maximum value among all computational cores to obtain the global execution time extremum. The calculation formula is as follows: Multiply the current location probability by the heat dissipation of a single task, then sum the results across all tasks and each computing core to obtain the system energy consumption. The calculation formula is as follows: Multiplying the extreme value of global execution time by the system energy consumption yields the scheduling fitness evaluation value, calculated using the following formula: In the formula, For the first The scheduling fitness evaluation value corresponding to each particle; the smaller the value, the better the corresponding task scheduling scheme.

[0036] The initial position probabilities of the particle swarm are initialized using the inverse ratio of the single-task baseline delay, and the initial evolution velocity is set to zero. The specific steps are as follows: The single-task baseline latency is the time taken for a task to execute on its corresponding computing core without thermal latency, assuming the shared memory controller is operating at its standard temperature of 303.15 Kelvin. At this temperature, the dynamically available memory bandwidth equals the factory-set maximum bandwidth. The calculation formula is as follows: In the formula, For the first The task in the first Single-task benchmark latency on each computing core.

[0037] The single-task response rate is obtained by taking the reciprocal of the baseline latency for the task to be assigned. The initial position probability assigned to the particle is obtained by dividing the single-task response rate of the task to be assigned on the current computing core by the sum of the single-task response rates of the task on all computing cores. The calculation formula is as follows: In the formula, For the first The first particle The task was assigned to the first The initial location probability of each computational core; For the first The task in the first Single-task baseline latency on each computational core. All particles have the same initial position probability to ensure the uniformity of the initial population.

[0038] The initial evolution velocity of the particles is set to zero, that is... ; In the formula, For the first The first particle The task was assigned to the first The initial evolution rate of each computing core.

[0039] The dynamic inertia weight is calculated by the ratio of the scheduling fitness evaluation value, and the individual cognitive factor is calculated by the ratio of the initial absolute temperature of the environment to the pre-simulation memory temperature. This drives the particles to iterate towards the globally optimal scheduling solution. The specific steps are as follows: The system obtains the current evolution velocity and the individual's historical optimal solution for each particle, and reads the first and second voltage-to-noise ratios of the underlying system. The individual's historical optimal solution is the position probability corresponding to the minimum scheduling fitness evaluation value obtained by a single particle during the iteration process, denoted as... Initially, the individual's historical optimal solution equals the particle's initial position probability. The first voltage noise floor ratio and the second voltage noise floor ratio are the ratios of the noise of the system power module's output voltage to the reference voltage; these are inherent system parameters, with values ​​ranging from [value range missing]. The values ​​can be predetermined through the system calibration process. The calibration method involves collecting 1000 samples of the power module output voltage under no-load conditions, calculating the ratio of the voltage standard deviation to the reference voltage, taking 1.2 times the first voltage noise floor ratio and 0.8 times the second voltage noise floor ratio.

[0040] The dynamic inertia weight is obtained by dividing the minimum value of the scheduling fitness evaluation of all particles in the current iteration by the scheduling fitness evaluation value of the current particle. The calculation formula is as follows: In the formula, For the first In the first iteration The dynamic inertial weight of each particle; For the first In the first iteration The scheduling fitness evaluation value of each particle; For the first In the first iteration The scheduling fitness evaluation value of each particle. When the dynamic inertia weight is greater than 1, it is set to 1; when the dynamic inertia weight is less than 0.1, it is set to 0.1.

[0041] The individual cognitive factor is obtained by dividing the initial absolute temperature of the environment by the pre-simulation memory temperature of the current iteration. The calculation formula is as follows: In the formula, For the first Individual cognitive factors in round-reaction iteration; For the first In the first iteration The pre-simulation memory temperature corresponds to each particle. When the individual cognitive factor is greater than 1, it is set to 1; when the individual cognitive factor is less than 0.1, it is set to 0.1.

[0042] Subtracting the individual cognitive factor from 1 yields the social cognitive factor, calculated using the following formula: In the formula, For the first Social cognitive factors in round-trip iteration.

[0043] The inertial component is obtained by multiplying the dynamic inertial weight by the current evolution velocity. The calculation formula is as follows: In the formula, For the first In the first iteration The evolution rate of each particle.

[0044] The self-cognition component is obtained by continuously multiplying the individual cognitive factor, the first voltage-to-noise ratio, and the difference between the individual's historical best solution and the probability of the current position. The calculation formula is as follows: In the formula, This is the first voltage noise floor value; For the first In the first iteration The probability of the current position of each particle.

[0045] The social cognitive component is obtained by continuously multiplying the social cognitive factor, the second voltage noise floor ratio, and the difference between the global optimal solution of the current group and the probability of the current position. The calculation formula is as follows: In the formula, This is the second voltage noise floor value; The position probability corresponding to the minimum scheduling fitness evaluation value obtained by the entire particle swarm during the iteration process is the global optimal scheduling solution. Initially, the global optimal scheduling solution is equal to the one with the smallest scheduling fitness evaluation value among all the initial position probabilities of the particles.

[0046] The updated evolutionary velocity is obtained by adding the inertial component, the self-cognition component, and the social cognition component. The calculation formula is as follows: In the formula, For the first In the first iteration The evolution rate of each particle. When the updated evolution rate is greater than the maximum evolution rate, the maximum evolution rate is taken; when the updated evolution rate is less than the negative maximum evolution rate, the negative maximum evolution rate is taken, and the maximum evolution rate is taken as 0.1.

[0047] The updated position probability is obtained by combining the evolutionary position normalization rule. The evolutionary position normalization rule is: for any task... Normalize the sum of its updated position probabilities across all computational cores to The calculation formula is: In the formula, For the first In the first iteration The position probability of each particle. When the calculated position probability is less than 0, it is set to 0; when the calculated position probability is greater than 1, it is set to 1, and then the normalization operation is re-executed.

[0048] The algorithm iterates repeatedly to output the converged global optimal scheduling solution. Iterative convergence conditions include reaching a preset maximum number of iterations, or the decrease in the scheduling fitness evaluation value corresponding to the global optimal solution after a consecutive preset number of iterations being less than a preset threshold. The preset maximum number of iterations is typically between 100 and 1000. Specifically, it's 100 iterations when the number of tasks is less than or equal to 10; 500 iterations when the number of tasks is greater than 10 but less than or equal to 50; and 1000 iterations when the number of tasks is greater than 50. The preset number of iterations is 20, and the preset threshold is one-thousandth of the scheduling fitness evaluation value corresponding to the current global optimal scheduling solution. The iteration terminates when either condition is met, and the current global optimal scheduling solution is output.

[0049] The globally optimal scheduling solution is transformed into deterministic assignment actions, and the starting core temperature for the next batch is updated using the thermal exponential decay law. The specific steps are as follows: Obtain the core thermal resistance, core thermal capacity, and current absolute core temperature for the corresponding computing core. Core thermal resistance is the thermal resistance between the computing core and the environment, characterizing the ease with which heat is transferred from the computing core to the environment. It can be obtained from the chip's technical datasheet and is measured in Kelvin per watt. Core thermal capacity is the amount of heat absorbed or released by the computing core per unit temperature change. It can be calculated based on the computing core's material composition, volume, and specific heat capacity and is measured in joules per Kelvin. The current absolute core temperature is the actual temperature of the computing core at the current moment. It can be read in real-time using an on-chip temperature sensor integrated within each computing core with a sampling period of 1 millisecond and is measured in Kelvin.

[0050] Extract the computational core that maximizes the probability from the globally optimal scheduling solution as the target, and use it to constitute a deterministic assignment action. The calculation formula is as follows: In the formula, For the first The deterministic assignment action of the first task indicates that the first task will be assigned. The task was assigned to the first Each computing core executes the operation. When multiple computing cores have equal and maximum probabilities, the computing core with the smallest number is selected as the target.

[0051] The total execution time of all tasks assigned to the same computing core is obtained by summing the actual execution times of all tasks assigned to the same computing core, using the following formula: In the formula, The first corresponding to the globally optimal scheduling solution The task in the first The actual execution time on each computing core.

[0052] The maximum value among the total execution times of all computing cores is taken as the actual execution window latency. The calculation formula is as follows: In the formula, The actual execution window delay is in seconds.

[0053] The increase in temperature is calculated by dividing the sum of the heat dissipated by a single task assigned to the current computing core by the core's heat capacity. The formula is as follows: In the formula, For the first The core heat capacity of each computing core.

[0054] Adding the current core absolute temperature to the newly generated temperature rise and subtracting the ambient absolute initial temperature, the transient temperature difference is obtained using the following formula: In the formula, For the first Each computing core at time... The current core absolute temperature.

[0055] The cooling attenuation term, calculated as the ratio of the actual execution window delay to the product of core thermal resistance and core thermal capacity with a base of the natural constant and a negative exponent, is given by the following formula: In the formula, For the first The core thermal resistance of each computing core.

[0056] Multiply the transient temperature difference by the cooling decay term, and finally add the ambient absolute initial temperature to obtain the starting core temperature of the corresponding calculated core in the next batch. The calculation formula is as follows: In the formula, For the first The initial core temperature of each computing core during the next batch of task scheduling, in Kelvin. When the calculated initial core temperature is lower than the ambient absolute initial temperature, the ambient absolute initial temperature is used.

[0057] The task scheduling system is triggered at 10% of the actual execution window delay. When a new task is available in the system, a task scheduling process is triggered immediately.

[0058] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A low-latency task scheduling system for edge computing nodes based on an improved swarm intelligence algorithm, applied to an edge computing node comprising a computing core consisting of a central processing unit and a neural network processor, a shared memory controller, and an explosion-proof sealing system, characterized in that: Configured for execution: The absolute initial temperature of the environment is obtained, and the microscopic operating parameters of the calculation core are extracted to calculate the transient heat flux. The pre-simulated memory temperature of the shared memory controller is calculated based on the transient heat flux pre-simulated spatial thermal wave interferometry. The pre-simulated memory temperature is mapped to dynamic available memory bandwidth using the thermal scattering attenuation law; The actual execution time of the task is calculated based on the dynamically available memory bandwidth, and a scheduling fitness evaluation value is constructed in combination with the system energy consumption. The initial position probability of the particle swarm is initialized using the inverse ratio of the single-task baseline delay, and the initial evolution velocity is set to zero. The dynamic inertia weight is calculated based on the ratio of the scheduling fitness evaluation value, and the individual cognitive factor is calculated based on the ratio of the absolute initial temperature of the environment to the pre-simulation memory temperature, driving the particles to iterate towards the global optimal scheduling solution. The global optimal scheduling solution is transformed into deterministic assignment actions, and the starting core temperature of the next batch is updated using the thermal exponential decay law.

2. The low-latency task scheduling system for edge computing nodes based on an improved swarm intelligence algorithm according to claim 1, characterized in that, The steps for extracting the microscopic operating parameters of the computational core to calculate the transient heat flux include: Obtain the single-cycle equivalent dynamic capacitance, transient supply voltage, and transient clock frequency of each computing core; The transient power supply voltage of each of the aforementioned calculation cores is squared to obtain the voltage square value; The transient heat flux of the corresponding computing core is calculated by continuously multiplying the single-cycle equivalent dynamic capacitance, the squared voltage value, and the transient clock frequency of each computing core.

3. The low-latency task scheduling system for edge computing nodes based on an improved swarm intelligence algorithm according to claim 2, characterized in that, The steps for calculating the simulated memory temperature of the shared memory controller include: The total number of instructions corresponding to the task, the instruction throughput per cycle of the corresponding computing core, the memory thermal capacity of the shared memory controller, the spatial thermal decay constant, and the current temperature of the shared memory are obtained. Divide the total number of instructions by the product of the transient clock frequency of the corresponding computing core and the instruction throughput per cycle, and then multiply by the transient heat flux to obtain the heat dissipation of a single task. Multiply the particle's current position probability by the corresponding heat dissipation of a single task, and sum the results for all tasks assigned to the same computing core to obtain the total core heat dissipation; divide the total core heat dissipation by the memory heat capacity to obtain the base temperature rise. By combining the aforementioned spatial thermal decay constant, an exponential decay calculation is performed on the base temperature rise to obtain the decayed temperature rise. The pre-simulated memory temperature is obtained by summing the temperature rises after decay of all the computing cores and then adding the current shared memory temperature.

4. The low-latency task scheduling system for edge computing nodes based on the improved swarm intelligence algorithm according to claim 3, characterized in that, The step of mapping the simulated memory temperature to dynamically available memory bandwidth includes: Obtain the system's factory-set maximum bandwidth; Divide the absolute initial temperature of the environment by the pre-simulated memory temperature to obtain the absolute temperature quotient. The nonlinear decay rate is obtained by performing a cube root operation on the absolute temperature quotient and then taking the square root. The dynamic available memory bandwidth is obtained by multiplying the factory-set maximum bandwidth by the nonlinear attenuation rate.

5. The low-latency task scheduling system for edge computing nodes based on an improved swarm intelligence algorithm according to claim 4, characterized in that, The steps for constructing the scheduling fitness evaluation value include: Get the memory access data volume of the tasks to be assigned; The data blocking time is obtained by dividing the amount of memory accessed by the dynamic available memory bandwidth; the computation time is obtained by dividing the total number of instructions by the product of the transient clock frequency and the instruction throughput per cycle; the actual execution time of the task is obtained by adding the data blocking time and the computation time. Multiply the current position probability by the actual execution time of the task and sum the results across all tasks. Take the maximum value among the calculation cores to obtain the global execution time extreme value. The system energy consumption is obtained by multiplying the current location probability by the heat emitted by a single task, then summing the results by iterating through all tasks and each computing core. The scheduling fitness evaluation value is obtained by multiplying the extreme value of the global execution time by the system energy consumption.

6. The low-latency task scheduling system for edge computing nodes based on an improved swarm intelligence algorithm according to claim 5, characterized in that, The initialization steps include: The single-task baseline latency of the computing task under the condition that no thermal latency occurs on the corresponding computing core; The single-task response rate is obtained by taking the reciprocal of the single-task baseline delay corresponding to the task to be assigned. The initial position probability assigned to the particle is obtained by dividing the single-task response rate of the task to be assigned on the current computing core by the sum of the single-task response rates of the task on all computing cores. The initial evolution velocity of the particles is set to zero.

7. The low-latency task scheduling system for edge computing nodes based on an improved swarm intelligence algorithm according to claim 6, characterized in that, The steps to drive particles to iterate towards the globally optimal scheduling solution include: Obtain the current evolution velocity of the particle and the individual historical optimal solution, and read the first voltage noise floor value and the second voltage noise floor value of the underlying system; The dynamic inertia weight is obtained by dividing the minimum value of the scheduling fitness evaluation value of all particles in the current iteration by the scheduling fitness evaluation value of the current particle. The individual cognitive factor is obtained by dividing the absolute initial temperature of the environment by the pre-simulation memory temperature of the current iteration round, and the social cognitive factor is obtained by subtracting the individual cognitive factor from one. The inertial component is obtained by multiplying the dynamic inertial weight by the current evolution velocity; the self-cognition component is obtained by continuously multiplying the individual cognitive factor, the first voltage noise floor value, and the difference between the individual's historical optimal solution and the current position probability; the social cognitive component is obtained by continuously multiplying the social cognitive factor, the second voltage noise floor value, and the difference between the current group's global optimal solution and the current position probability. The inertial component, the self-cognition component, and the social cognition component are added together to obtain the updated evolution velocity. The updated position probability is obtained by combining the evolution position normalization rule. The converged global optimal scheduling solution is output through iterative loops.

8. The low-latency task scheduling system for edge computing nodes based on an improved swarm intelligence algorithm according to claim 7, characterized in that, The steps for updating and determining the starting core temperature for the next batch include: Obtain the core thermal resistance, core thermal capacity, and current absolute temperature of the corresponding computing core; The computational core that maximizes the probability in the global optimal scheduling solution is extracted as the target, and the deterministic assignment action is constituted. The actual execution times of all tasks assigned to the same computing core are summed, and the maximum value among all computing cores is taken as the actual execution window delay. The sum of the heat dissipated by the single task assigned to the current computing core is divided by the core's heat capacity to obtain the new temperature rise. The current absolute core temperature is added to the new temperature rise and the ambient absolute initial temperature is subtracted to obtain the transient temperature difference. Calculate the cooling attenuation term, which is the ratio of the actual execution window delay to the product of the core thermal resistance and the core thermal capacity, with the natural constant as the base and the exponent being negative; Multiply the transient temperature difference by the cooling attenuation term, and finally add the ambient absolute initial temperature to obtain the starting core temperature of the corresponding computing core in the next batch.

Images