Energy-saving server operation method and system based on dynamic load regulation

Abstract

The invention discloses an energy-saving server operation method and system based on dynamic load adjustment, and relates to the technical field of servers. The energy-saving server operation method comprises the steps that a plurality of heterogeneous resources in a single physical machine are made to generate a plurality of virtual heterogeneous resource units through a pooling method and a virtualization method, and the heterogeneous resources in each virtual heterogeneous resource unit are different in proportion. According to the method, through establishment of a first framework, the first framework supports collaborative management and scheduling among various heterogeneous resources, the first framework dynamically adjusts the scale of a virtual heterogeneous resource unit for a single physical machine through load calculation, comparison and triggering conditions, and the virtual heterogeneous resource unit is optimized on the basis of the reduced scale. And meanwhile, different voltage reduction and frequency reduction are carried out on each heterogeneous resource in a single physical machine, so that the energy consumption of the physical machine is reduced, the voltage reduction and frequency reduction are respectively carried out on each heterogeneous resource, and the effective cooperative scheduling of the heterogeneous resources is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of servers, and specifically to an energy-saving server operation method and system based on dynamic load regulation. Background Art

[0002] On servers, the energy consumption cost accounts for the vast majority of the total expenditure, and energy consumption is regarded as a key issue on servers.

[0003] For example, the patent publication number "CN103970256A", with the name "An Energy-Saving Method and System Based on Memory Compression and CPU Dynamic Frequency Modulation", the above method includes compressing pages with the same content in the memory space, placing the saved space in a low-energy state, and reducing the frequency of the CPU to the lowest frequency required to run the memory space. The above invention can well control the energy consumption of the server, effectively reduce the energy consumption brought by the CPU and memory, and improve the memory usage efficiency while reducing the server energy consumption.

[0004] When the above method realizes energy saving, the CPU dynamic frequency modulation technology is combined with the memory compression technology to reduce the energy consumption in the memory compression process. However, on servers, there are not only CPU and memory resources, but also heterogeneous resources such as graphics processing unit GPU, field programmable gate array FPGA, and disk network. The above method fails to effectively coordinate and schedule heterogeneous resources such as CPU and GPU, resulting in some redundant heterogeneous resources being overloaded or idle, thus still having energy consumption on other heterogeneous resources. Therefore, an energy-saving server operation method and system based on dynamic load regulation are invented. Summary of the Invention

[0005] The purpose of the present invention is to provide an energy-saving server operation method and system based on dynamic load regulation to solve the problems raised in the above background art.

[0006] To achieve the above purpose, the present invention provides the following technical solution: An energy-saving server operation method based on dynamic load regulation, the energy-saving server operation method includes: Generating a number of virtual heterogeneous resource units from a number of heterogeneous resources in a single physical machine through a pooling method and a virtualization method, and the ratio of each heterogeneous resource in each virtual heterogeneous resource unit is different; Establishing a second load threshold for the physical machine and a first load threshold for the virtual heterogeneous resource unit according to the device parameters of the physical machine; Constructing a first framework for optimizing the energy consumption of a single physical machine and a second framework for optimizing the energy consumption of a server cluster; The first framework includes: Obtaining the load rates of a number of virtual heterogeneous resource units and the real-time global load rate through load calculation , establish the first trigger condition for physical machine energy consumption optimization; Through the comparison method, compare the load rates of several virtual heterogeneous resource units with the load thresholds of virtual heterogeneous resource units, and compare the real-time global load rate with the second load threshold to obtain the first comparison result; Input the first comparison result into the first trigger condition to determine whether to trigger. If it is determined to trigger, based on the elastic resource scaling technology, adjust the scale of several virtual heterogeneous resource units in the physical machine to obtain the total heterogeneous scale of the combination of several adjusted virtual heterogeneous resource units. Compare the total heterogeneous scale with the total scale of the physical machine to obtain a set of comparison values, and dynamically adjust the load of several heterogeneous resources in a single physical machine through the set of comparison values and the adjustment method; The set of comparison values includes the comparison values between the actual scales of several heterogeneous resources in the physical machine and the virtual scales of several heterogeneous resources in the total heterogeneous scale , where i represents the i-th heterogeneous resource; The adjustment method includes: dynamically adjusting the load of the devices from which several heterogeneous resources are sourced, and the energy consumption comparison value between the unit energy consumption of several modified heterogeneous resources and the unit energy consumption of the heterogeneous resources before modification , the energy consumption comparison value and the comparison value are positively correlated.

[0007] Furthermore, the second framework includes: Obtain the real-time global load rate of the physical machine through load calculation ; Through the comparison method, compare the real-time global load rate with the second load threshold to obtain the second comparison result; Establish the second trigger condition for the energy consumption of the server cluster, and determine whether to trigger the second trigger condition through the second comparison result. If it is determined to trigger; Obtain the difference between the real-time global load rate and the second load threshold, establish a safety value, and divide the tasks in the physical machine proportionally through the real-time global load rate to obtain the task scale ratio. Based on the difference, safety value, and task scale ratio, obtain the scale of the task to be processed, and obtain the task to be processed from the virtual heterogeneous resource unit with the highest virtual heterogeneous resource unit load rate; The tasks to be processed in the physical machine are transferred between physical machines through the feature selection method; The feature selection method includes selecting from the server cluster, based on the scale of the task to be processed, physical machines that can receive the task to be processed and will not cause the real-time global load rate after reception For a physical machine exceeding the second load threshold, the to-be-processed tasks are delivered to the selected physical machine; The to-be-processed tasks are processed, and the task processing includes: Feature processing is performed on the to-be-processed tasks to obtain the collaborative degree requirements of the to-be-processed tasks. The collaborative degree requirements represent the requirements of the tasks for different heterogeneous computing resources. Based on the collaborative degree requirements, a feature vector of the tasks is established, and a ratio feature vector of the virtual heterogeneous resource unit is established based on the ratio of each heterogeneous resource in the virtual heterogeneous resource unit; The to-be-processed tasks are input into the corresponding virtual heterogeneous resource units through a similarity selection method.

[0008] Furthermore, the similarity selection method includes a direct pairing method and a combined pairing method; The direct pairing method includes: The similarity degree values are calculated by calculating the similarity between the feature vector of the to-be-processed tasks and the ratio feature vectors of several virtual heterogeneous resource units in the physical machine. The virtual heterogeneous resource unit feature with the highest similarity degree value is selected for the input of the to-be-processed tasks, and a degree threshold is established. Tasks with similarity degree values lower than the degree threshold are input through the combined pairing method; The combined pairing method includes: Intelligent pairing combinations are performed among the to-be-processed tasks with similarity degree values lower than the degree threshold to obtain combined tasks and the feature vectors of the combined tasks. The similarity degree values are calculated by calculating the similarity between the feature vector of the combined tasks and the feature vectors of several virtual heterogeneous resource units in the physical machine to obtain a set of similarity degree values, and the virtual heterogeneous resource unit feature with the highest similarity degree value is selected for the input of the combined tasks: A time threshold is established. For tasks in the combined pairing method that still cannot be combined when exceeding the time threshold, the degree threshold is ignored, and the virtual heterogeneous resource unit feature with the highest similarity degree value is directly selected for the input of the combined tasks.

[0009] Furthermore, load calculation includes normalization processing and load operation; The normalization processing includes establishing a normalization formula: , where the minimum load rate and the maximum load rate are the theoretical minimum load rate and the theoretical maximum load rate of each heterogeneous resource respectively; The loads of each heterogeneous resource are calculated through the normalization formula to obtain the normalized load rates of each heterogeneous resource.

[0010] Furthermore, the elastic scaling technology includes dynamically scaling the scale of each heterogeneous resource of the virtual heterogeneous resource unit, and keeping the ratio of heterogeneous resources unchanged during the scaling process; Dynamic load regulation includes adjusting the voltage and frequency of heterogeneous resources and controlling the source device to be in a low-power state; The comparison value set includes the comparison values between the actual scale of each heterogeneous resource in the physical machine and the virtual scale of each heterogeneous resource in the total heterogeneous scale; Energy consumption comparison value and the comparison value are positively correlated, and a comparison coefficient is established , .

[0011] Furthermore, the load operation includes: Real-time global load rate ; , where N represents the number of groups in the heterogeneous resource group set, represents the overall occupancy ratio of the i-th heterogeneous resource in the physical machine, represents the normalized load rate of the i-th heterogeneous computing resource, represents the real-time global load rate calculated based on the heterogeneous resource load; The overall occupancy ratio is replaced by the energy consumption ratio, and the energy consumption ratio represents the energy consumption weight of several heterogeneous resources in the physical machine, and the sum of the energy consumption ratios is 1.

[0012] Furthermore, the second trigger condition includes: Real-time global load rate continuously greater than or the maximum value of the second load threshold, and at least maintained for three minutes.

[0013] Furthermore, the load operation includes: Establish a virtual heterogeneous resource unit load rate formula ; , represents the load rate of the j-th virtual heterogeneous resource unit, the normalized processor load , the normalized memory load , the normalized disk load , the normalized network bandwidth utilization rate ; , , and represent the energy consumption ratios of the virtual heterogeneous resource units, and the sum of the energy consumption ratios is 1; , where Y represents the number of virtual heterogeneous resource units in the physical machine, represents the ratio of the scale of the Y-th virtual heterogeneous resource unit to the total scale of the physical machine, Indicates the real-time global load rate based on the load calculation of virtual heterogeneous resource units.

[0014] Furthermore, the first trigger condition includes: Condition 1: The virtual heterogeneous resource unit continuously remains below the virtual heterogeneous resource unit threshold and lasts for at least three minutes; Condition 2: The real-time global load rate continuously remains less than the minimum value of the second load threshold and lasts for at least three minutes; Both Condition 1 and Condition 2 need to be satisfied simultaneously.

[0015] An energy-saving server operation system based on dynamic load regulation adopts the above-mentioned energy-saving server operation method based on dynamic load regulation. The energy-saving server operation system includes: Resource allocation module: Generates a number of virtual heterogeneous resource units from several heterogeneous resources in a single physical machine through pooling and virtualization methods. The ratio of each heterogeneous resource in each virtual heterogeneous resource unit is different; Threshold establishment module: Establishes the second load threshold of the physical machine and the first load threshold of the virtual heterogeneous resource unit according to the device parameters of the physical machine; Energy consumption optimization module: The energy consumption optimization module includes a first optimization module and a second optimization module; First optimization module: Obtains the load rates of a number of virtual heterogeneous resource units and the real-time global load rate through load calculation , and establishes the first trigger condition for physical machine energy consumption optimization; Through a comparison method, compares the load rates of a number of virtual heterogeneous resource units with the virtual heterogeneous resource unit load threshold, and compares the real-time global load rate with the second load threshold to obtain the first comparison result; Inputs the first comparison result into the first trigger condition to determine whether to trigger. If it is determined to trigger, based on the elastic resource scaling technology, reduces the scale of a number of virtual heterogeneous resource units in the physical machine to obtain the combined heterogeneous total scale among the reduced number of virtual heterogeneous resource units. Compares the heterogeneous total scale with the physical machine total scale to obtain a set of comparison values, and dynamically adjusts the load of a number of heterogeneous resource source devices in a single physical machine through the set of comparison values and adjustment methods; Second optimization module: Obtains the real-time global load rate of the physical machine through load calculation ; Through a comparison method, compares the real-time global load rate with the second load threshold to obtain the second comparison result; Establishes the second trigger condition for the energy consumption of the server cluster, and determines whether to trigger the second trigger condition through the second comparison result. If it is determined to trigger: Obtain the real-time global load rate The difference from the second load threshold is used to establish a safety value, and based on the real-time global load rate The tasks in the physical machine are proportionally divided to obtain the task scale ratio. According to the difference, safety value and task scale ratio, the scale of the tasks to be processed is obtained, and the tasks to be processed are obtained from the virtual heterogeneous resource unit with the highest load rate among the virtual heterogeneous resource units; The tasks to be processed in the physical machine are transmitted between physical machines through a feature selection method.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: In the energy-saving server operation method and system based on dynamic load regulation, through the establishment of the first framework, the first framework supports collaborative management and scheduling among multiple heterogeneous resources. The first framework targets a single physical machine and dynamically adjusts the scale of the virtual heterogeneous resource unit through load calculation, comparison, and trigger conditions. Based on the reduced scale, different voltage and frequency reductions are simultaneously performed on each heterogeneous resource in the single physical machine, reducing the energy consumption of the physical machine while performing voltage and frequency reduction on each heterogeneous resource separately, ensuring the effective collaborative scheduling of heterogeneous resources. At the same time, the second framework targets the server cluster. According to the comparison between the real-time global load rate and the load threshold, when a specific second trigger condition is met, through task migration and resource reallocation, the physical machine is kept in a normal state, avoiding being in a high-energy consumption state, and optimizing the energy consumption of the entire cluster.

[0017] Through the feature selection and similarity selection methods, tasks are reasonably allocated to the most suitable virtual heterogeneous resource unit, reducing resource fragmentation and avoiding the situation of task-resource mismatch, improving task processing efficiency. When optimizing the energy consumption of a single physical machine subsequently, voltage and frequency reduction are performed on the physical machine. At the same time, the reduction of resource fragmentation results in the reduction of idle resources, ensuring that the physical machine after voltage and frequency reduction can still handle the tasks in the physical machine even at a lower energy consumption. At the same time, during the process of scaling up and down, the heterogeneous resource ratio remains unchanged, ensuring the stability and performance of the system. The scale of the virtual heterogeneous resource unit is dynamically adjusted in combination with the load situation, which can not only meet the task requirements under high load but also avoid resource waste under low load. Through the feature selection and similarity selection methods, the optimization of the first framework can be ensured, and based on the task allocation method of the feature selection and similarity selection methods, there are fewer resource fragments in the heterogeneous resources after voltage and frequency reduction.

[0018] By establishing a load rate formula for virtual heterogeneous resource units and a calculation method for real-time global load rate, unified management and scheduling of different heterogeneous resources are achieved. The system can intelligently allocate tasks to the most suitable resource units according to the feature vectors of tasks and the ratio feature vectors of virtual heterogeneous resource units through similarity calculation. This intelligent scheduling method improves the accuracy of task allocation and optimizes the task processing flow of the entire server system. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the operation method of the present invention; Figure 2 It is a schematic diagram of the feature selection method of the present invention; Figure 3 It is a schematic diagram of the first framework of the present invention; Figure 4 It is a schematic diagram of the operation system of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0021] As Figures 1-4 shown, the present invention provides a technical solution: an energy-saving server operation method based on dynamic load regulation, and the energy-saving server operation method includes: Generating a number of virtual heterogeneous resource units from several heterogeneous resources in a single physical machine through pooling and virtualization methods, and the ratios of various heterogeneous resources in each virtual heterogeneous resource unit are different; Establishing a second load threshold for the physical machine and a first load threshold for the virtual heterogeneous resource units according to the device parameters of the physical machine; Constructing a first framework for optimizing the energy consumption of a single physical machine and a second framework for optimizing the energy consumption of a server cluster; The first framework includes: Obtaining the load rates of a number of virtual heterogeneous resource units and the real-time global load rate through load calculation , and establishing a first trigger condition for optimizing the energy consumption of the physical machine; Implementing the comparison between the load rates of a number of virtual heterogeneous resource units and the load thresholds of the virtual heterogeneous resource units through a comparison method, and the real-time global load rate and the second load threshold, and obtaining a first comparison result; The first comparison result is input to the first trigger condition to determine whether to trigger. If it is determined to trigger, based on the elastic resource scaling technology, the scale of several virtual heterogeneous resource units in the physical machine is adjusted, and the total heterogeneous scale of the combined virtual heterogeneous resource units after adjustment is obtained. The total heterogeneous scale is compared with the total scale of the physical machine to obtain a comparison value set, and the dynamic load adjustment of several heterogeneous resources in a single physical machine is performed through the comparison value set and the adjustment method; The comparison value set includes the comparison values between the actual scales of several heterogeneous resources in the physical machine and the virtual scales of several heterogeneous resources in the total heterogeneous scale , where i represents the i-th heterogeneous resource; The adjustment method includes: dynamically adjusting the load of several heterogeneous resource source devices, and obtaining the energy consumption comparison value between the unit energy consumption of several heterogeneous resources after modification and the unit energy consumption of the heterogeneous resources before modification , the energy consumption comparison value and the comparison value are positively correlated.

[0022] The second framework includes: Obtaining the real-time global load rate of the physical machine through load calculation ; Implementing the comparison between the real-time global load rate and the second load threshold through a comparison method to obtain the second comparison result; Establishing a second trigger condition for the energy consumption of the server cluster, and determining whether to trigger the second trigger condition through the second comparison result. If it is determined to trigger: Obtaining the difference between the real-time global load rate and the second load threshold, establishing a safety value, and proportionally dividing the tasks in the physical machine through the real-time global load rate to obtain the task scale ratio, and obtaining the scale of the task to be processed based on the difference, the safety value, and the task scale ratio, and obtaining the task to be processed from the virtual heterogeneous resource unit with the highest virtual heterogeneous resource unit load rate; The tasks to be processed in the physical machine are transmitted between physical machines through a feature selection method; The feature selection method includes selecting, from the server cluster based on the scale of the task to be processed, the physical machines that can receive the task to be processed and will not cause the real-time global load rate to exceed the second load threshold after reception, and transporting the task to be processed to the selected physical machines; Performing task processing on the task to be processed, and the task processing includes: Perform feature processing on the task to be processed to obtain the collaboration degree requirement of the task to be processed. The collaboration degree requirement represents the demand of the task for different heterogeneous computing resources. Based on the collaboration degree requirement, establish the feature vector of the task, and based on the ratio of each heterogeneous resource in the virtual heterogeneous resource unit, establish the ratio feature vector of the virtual heterogeneous resource unit; The task to be processed is input into the corresponding virtual heterogeneous resource unit through the similarity selection method.

[0023] The similarity selection method includes the direct pairing method and the combined pairing method; The direct pairing method includes: Calculate the similarity between the feature vector of the task to be processed and the ratio feature vectors of several virtual heterogeneous resource units in the physical machine to obtain the similarity degree value. Select the virtual heterogeneous resource unit feature with the highest similarity degree value for the input of the task to be processed, establish a degree threshold, and tasks with similarity degree values lower than the degree threshold are input through the combined pairing method; The combined pairing method includes: Intelligently pair and combine the tasks to be processed with similarity degree values lower than the degree threshold to obtain the combined task and the feature vector of the combined task. Calculate the similarity between the feature vector of the combined task and the feature vectors of several virtual heterogeneous resource units in the physical machine to obtain a set of similarity degree values, and select the virtual heterogeneous resource unit feature with the highest similarity degree value for the input of the combined task; Establish a time threshold. For tasks in the combined pairing method that have not been combined when exceeding the time threshold, ignore the degree threshold and directly select the virtual heterogeneous resource unit feature with the highest similarity degree value for the input of the combined task.

[0024] The load calculation through includes normalization processing and load operation; The normalization processing includes establishing a normalization formula: , where the minimum load rate and the maximum load rate are the theoretical minimum load rate and the theoretical maximum load rate of each heterogeneous resource respectively; Calculate the load of each heterogeneous resource through the normalization formula to obtain the normalized load rate of each heterogeneous resource.

[0025] The elastic scaling technology includes dynamically scaling the scale of each heterogeneous resource in the virtual heterogeneous resource unit, and keeping the ratio of heterogeneous resources unchanged during the scaling process; The dynamic load regulation includes adjusting the voltage and frequency of the heterogeneous resource and controlling the source device to be in a low-power state; The comparison value set includes the comparison value between the actual scale of each heterogeneous resource in the physical machine and the virtual scale of each heterogeneous resource in the total heterogeneous scale; The energy consumption comparison value and the comparison value Show a positive correlation and establish a comparison coefficient , .

[0026] The load operation includes: Real-time global load rate ; , where N represents the number of groups in the heterogeneous resource grouping set, represents the overall occupancy ratio of the i-th heterogeneous resource in the physical machine, represents the normalized load rate of the i-th heterogeneous computing resource, represents the real-time global load rate calculated based on the heterogeneous resource load; The overall occupancy ratio is replaced by the energy consumption ratio. The energy consumption ratio represents the energy consumption weight of several heterogeneous resources in the physical machine, and the sum of the energy consumption ratios is 1.

[0027] The second trigger condition includes: Real-time global load rate continuously greater than or the maximum value of the second load threshold, and maintained for at least three minutes.

[0028] The load operation includes: Establish a formula for the load rate of virtual heterogeneous resource units ; , represents the load rate of the j-th virtual heterogeneous resource unit, the normalized processor load , the normalized memory load , the normalized disk load , the normalized network bandwidth utilization rate ; , , and represent the energy consumption ratio of the virtual heterogeneous resource unit, and the sum of the energy consumption ratios is 1; , where Y represents the number of virtual heterogeneous resource units in the physical machine, represents the ratio of the scale of the Y-th virtual heterogeneous resource unit to the total scale of the physical machine, represents the real-time global load rate calculated based on the virtual heterogeneous resource unit load.

[0029] The first trigger condition includes: Condition 1: The virtual heterogeneous resource unit is continuously lower than the virtual heterogeneous resource unit threshold and maintained for at least three minutes; Condition 2: The real-time global load rate continuously less than the minimum value of the second load threshold and maintain this state for at least three minutes; Both Condition 1 and Condition 2 need to be satisfied simultaneously.

[0030] An energy-saving server operation system based on dynamic load regulation adopts the above-mentioned energy-saving server operation method based on dynamic load regulation. The energy-saving server operation system includes: Resource allocation module: generating a number of virtual heterogeneous resource units from several heterogeneous resources in a single physical machine through pooling and virtualization methods, with different ratios of heterogeneous resources in each virtual heterogeneous resource unit; Threshold establishment module: establishing the second load threshold of the physical machine and the first load threshold of the virtual heterogeneous resource unit according to the device parameters of the physical machine; Energy consumption optimization module: The energy consumption optimization module includes a first optimization module and a second optimization module; First optimization module: obtaining the load rates of a number of virtual heterogeneous resource units and the real-time global load rate through load calculation , and establishing a first trigger condition for physical machine energy consumption optimization; Through a comparison method, compare the load rates of a number of virtual heterogeneous resource units with the load thresholds of the virtual heterogeneous resource units, and compare the real-time global load rate with the second load threshold to obtain a first comparison result; Input the first comparison result into the first trigger condition to determine whether to trigger. If it is determined to trigger, based on the elastic resource scaling technology, reduce the scale of a number of virtual heterogeneous resource units in the physical machine to obtain the combined heterogeneous total scale among the reduced virtual heterogeneous resource units. Compare the heterogeneous total scale with the total scale of the physical machine to obtain a set of comparison values, and perform dynamic load regulation on a number of heterogeneous resource source devices in a single physical machine through the set of comparison values and adjustment methods; Second optimization module: obtaining the real-time global load rate of the physical machine through load calculation ; Through a comparison method, compare the real-time global load rate with the second load threshold to obtain a second comparison result; Establish a second trigger condition for the energy consumption of the server cluster, and determine whether to trigger the second trigger condition through the second comparison result. If it is determined to trigger: Obtain the difference between the real-time global load rate and the second load threshold, establish a safety value, and divide the tasks in the physical machine proportionally through the real-time global load rate to obtain a task scale ratio. Obtain the scale of the task to be processed based on the difference, the safety value, and the task scale ratio, and obtain the task to be processed from the virtual heterogeneous resource unit with the highest load rate among the virtual heterogeneous resource units; The tasks to be processed in the physical machine are transmitted between physical machines through the feature selection method.

[0031] The server runs on the physical machine, and the physical machine provides the necessary hardware resources for the server, such as the computing power of each processor, the storage space of the memory, the data storage of the hard disk, etc., enabling the server to execute various tasks and services. The energy-saving server operation method actually enables the physical machine carrying the server to save energy during operation. By dynamically adjusting the load of the physical machine and the load of each heterogeneous resource on the physical machine, the physical machine can save energy. At the same time, the server cluster constructed by multiple physical machines will also adopt the same method to achieve energy saving. On a single physical machine, there are multiple processors. Therefore, during the operation of the physical machine, if a single physical machine is in a low-load state, the load of the processors on the physical machine will be reduced, causing some processors to be in a low-power consumption mode, so that the remaining processors can be in the high-efficiency area and can adapt to the physical machine. When a single physical machine is in an overloaded state, there will also be processors in an overloaded state. At this time, task migration will be triggered and tasks will stop being input to this physical machine. Some tasks in this physical machine will be migrated to other physical machines in a low-load state or normal state, while ensuring that other physical machines receiving tasks will not be in a high-load state after receiving the tasks. Dynamic voltage and frequency scaling (DVFS), according to the load conditions of the virtual machines, Dynamically adjust the processor frequency and voltage of the physical machine. Reduce the processor frequency and voltage at low load to reduce dynamic power consumption, and increase the frequency and voltage at high load to meet performance requirements and achieve energy saving. By dynamically adjusting the voltage of the processor through dynamic voltage and frequency scaling, energy saving is achieved. The heterogeneous resources on a single physical machine are obtained by combining multiple identical devices. Therefore, when reducing the voltage and frequency of a single physical machine and it can meet the need to process the remaining tasks, it is manifested in multiple identical devices. Some of the identical devices are in a low-consumption or even dormant state, and the remaining identical devices are in the normal mode. The identical devices in the normal mode will process the remaining tasks.

[0032] The device parameters of a physical machine are obtained by combining the operating parameters of various heterogeneous resources. The heterogeneous resources include processing resources, memory resources, storage resources, and network resources, and also represent the physical devices providing such resources. The pooling method is a technology that integrates dispersed computing, storage, or network resources into a unified resource pool to improve resource utilization, enhance elasticity, and reduce energy consumption. The pooling method physically pools the CPUs, GPUs, FPGAs, and other heterogeneous resources in a physical server to build a unified heterogeneous resource pool containing logical operation units, parallel computing units, and hardware acceleration units, breaking the physical isolation between resources and realizing centralized management and scheduling of resources. This enables resources to be dynamically allocated to different tasks or applications according to demand, improving resource sharing and utilization, and enhancing the elasticity and flexibility of the system. Through virtualization technology, a single physical server is divided into multiple virtual machines that share resources such as processors and memory, avoiding resource idleness during low load. Virtualization technology logically partitions the hardware resources (such as CPUs, memory, storage, and network) of a single physical server through an abstraction layer into multiple independent and isolated virtual machines, each of which can run different operating systems and applications, achieving resource sharing and efficient utilization.

[0033] Intelligent pairing combination refers to the combination of tasks to be processed with similarity degree values lower than the degree threshold. Since the feature vectors and task volumes of the tasks to be processed are already clear, and the ratio feature vectors of virtual heterogeneous resource units are also known, during calculation, through intelligent pairing, it is analyzed which tasks to be processed can be combined, so that the similarity degree between the obtained combined tasks and the ratio feature vectors is maximized, and calculations are carried out based on this. In the combination pairing method, if a task still has not been input into the virtual heterogeneous resource unit within the set time, the limit of the degree threshold can be directly ignored, and the task can be input according to the characteristics of the virtual heterogeneous resource unit with the highest similarity degree value selected. This design is mainly to avoid the situation where tasks are shelved in the absence of adaptation, ensuring that tasks can proceed. This situation occurs less frequently, and the task volume is relatively small compared to the capacity of the virtual heterogeneous resource unit, so the impact on the fragmentation rate of the virtual heterogeneous resource unit is also relatively small.

[0034] Aligning tasks and resources through similarity matching can reduce the fragmentation rate in virtual heterogeneous resource units. By setting up several virtual heterogeneous resource units in a single server, multiple virtual heterogeneous resources can handle multiple tasks simultaneously. At the same time, it ensures that tasks on the same physical machine are restricted by virtual heterogeneous resources, thus avoiding interference between tasks. The system can more precisely allocate tasks to virtual heterogeneous resource units that match their resource requirements, which reduces resource fragmentation caused by mismatches between tasks and resources, thereby reducing the fragmentation rate. Through similarity calculation, the system can allocate resources more reasonably, ensuring that each task can run on the most suitable resources and reducing the situation of residual resources in virtual heterogeneous resource units. This optimized allocation reduces resource overload and idle situations, improving resource utilization. Therefore, when subsequent voltage and frequency scaling is performed on the processor in a low-load state, the system can release excess resources, enabling the processor to be in a low-consumption state, reducing processor energy consumption, and thus reducing the overall energy consumption. After voltage and frequency scaling of the heterogeneous resources in a single physical machine, the tasks originally in the virtual heterogeneous resource units in the single physical machine can still continue to operate normally. By means of feature selection and similarity selection methods, tasks are reasonably allocated to the most suitable virtual heterogeneous resource units, reducing resource fragmentation and avoiding mismatches between tasks and resources, improving task processing efficiency. When this method is used for subsequent energy consumption optimization of a single physical machine, voltage and frequency scaling are performed on the physical machine. At the same time, the reduction of resource fragmentation also reduces idle resources, ensuring that the physical machine after voltage and frequency scaling can still handle the tasks in the physical machine with lower energy consumption. At the same time, during the process of scaling up and down, the ratio of heterogeneous resources remains unchanged, ensuring the stability and performance of the system. Dynamically adjusting the scale of virtual heterogeneous resource units in combination with the load situation can not only meet the task requirements during high load but also avoid resource waste during low load. The feature selection and similarity selection methods can ensure the optimization of the first framework. Task allocation based on feature selection and similarity selection methods, resulting in less resource fragmentation in heterogeneous resources after voltage and frequency scaling.

[0035] Normalize the load rates of different heterogeneous resources to make them comparable and additive. Different heterogeneous resources (such as CPUs, GPUs, FPGAs, etc.) have different working characteristics and load ranges. The normalization process unifies data with different dimensions or ranges to the same scale (such as between 0 and 1), eliminating the impact of dimensional differences, enabling direct comparison and analysis of the load conditions of different resources. By establishing a load rate formula for virtual heterogeneous resource units and a calculation method for real-time global load rate, unified management and scheduling of different heterogeneous resources are achieved. The system can intelligently allocate tasks to the most suitable resource units based on the feature vectors of tasks and the matching feature vectors of virtual heterogeneous resource units using similarity calculations. This intelligent scheduling method improves the accuracy of task allocation and optimizes the task processing flow of the entire server system.

[0036] There is a load range for the use of processors. The load range includes a low-load area, an efficient area, and an overload area. At the same time, the corresponding processors are in a low-load state, a normal state, and an overload state. For example, in the low-load area (0% - 30%): The power consumption rises slowly with the load, and the energy efficiency ratio is low. In the efficient area (30% - 70%): The power consumption has an approximate linear relationship with the load, and the energy efficiency ratio is optimal. In the overload area (70%): The power consumption increases steeply, but the marginal increase in performance decreases (such as the CPU frequency being throttled due to heat limitations). Therefore, the physical machine and physical resources can be in the efficient area to ensure energy conservation and processing efficiency. For physical machines in the overload area, their loads will be reduced, and some of their tasks will be transferred to other physical machines. When triggering task migration, virtual heterogeneous resource units with the same heterogeneous resource ratio in other servers will be preferentially selected. However, if there are no virtual heterogeneous resource units with the same heterogeneous resource ratio in other servers, virtual heterogeneous resource units on other physical machines will be reselected through a similarity selection method. Note that other physical machines cannot be in a high-load state after receiving tasks. Similarly, there will be a similar situation for the load range in physical machines.

[0037] Connect virtual heterogeneous resource units with the same heterogeneous resource ratio between multiple servers through a cross-node method to form a global resource pool. When triggering task migration, virtual heterogeneous resource units with the same heterogeneous resource ratio in other servers will be selected through the global resource pool. However, if there are no virtual heterogeneous resource units with the same heterogeneous resource ratio in other servers, other virtual heterogeneous resource units will be reselected through a similarity selection method.

[0038] The real-time global load rate refers to the overall load rate of each heterogeneous resource in a single physical machine. Since a physical machine is composed of multiple heterogeneous resources, the real-time global load rate can be obtained through calculation. It represents the overall occupancy ratio of the i-th heterogeneous resource in the physical machine. The overall occupancy ratio can be calculated by substituting the energy consumption of the heterogeneous resource under normal circumstances. When different heterogeneous resources (such as CPUs, GPUs, FPGAs, etc.) process tasks, their energy consumption performances are different. The energy consumption of these resources is not only closely related to their load ratios (i.e., the ratio of the actual processed task volume to their maximum processing capacity). By measuring the energy consumption of the heterogeneous resource, its load ratio can be estimated, and then the overall occupancy ratio of this resource in the physical machine can be calculated. Moreover, through the energy consumption ratio, it can be determined which group has the highest energy consumption. Since its energy consumption is the highest, its impact on the real-time global load ratio is also the largest. Therefore, in this application, the energy consumption ratio is used to replace the overall occupancy ratio.

[0039] Establish task load adjustment for the physical machine. For example, when the real-time global load ratio is higher than 75% for 3 consecutive minutes, it indicates that the current task load of the server is heavy, and this physical machine can no longer meet the task requirements. At this time, according to the growth ratio of the task requirements, the scale of the virtual unit is expanded proportionally. For example, if the task requirements increase by 20%, the scale of the virtual unit also increases by 20% accordingly. This can allocate more computing resources for the task, improve the processing ability of the system, and ensure that the task can be processed in a timely and efficient manner. When the real-time global load ratio is lower than 30% for 5 consecutive minutes, it means that the current task load of the server is light, and the resource utilization rate of some virtual units is low, resulting in waste of resources. These virtual units with low utilization rates are shut down, the tasks in them are migrated to other virtual units with higher utilization rates, and the power supply of the redundant physical nodes after merging is turned off. When transferring tasks, it is necessary to ensure that the transferred tasks are also selectively transferred according to the similar selection method and cooperation degree requirements.

[0040] Select virtual heterogeneous resource units by identifying the cooperation degree requirements, so that the cooperation computing degree of the heterogeneous resources in the virtual heterogeneous resource units is close to the identified cooperation degree requirements. Thus, when the virtual heterogeneous resource units process the corresponding tasks, they can make full use of each heterogeneous resource in the virtual heterogeneous resource units, avoiding the situation where the utilization rate of a single heterogeneous resource is too high while the utilization rate of other heterogeneous resources is too low.

[0041] At the same time, since the processor processes multiple tasks, virtual heterogeneous resource units with corresponding relationships can be combined and input among multiple tasks. After the combination of two tasks, the original feature vector changes, so as to adapt to the ratio feature vector of the virtual heterogeneous resource unit. The similarity between the two feature vectors is calculated through the cosine similarity to obtain a similarity degree value. If the similarity degree value of a single task is lower than the degree threshold, it means that the similarity between the task and the virtual heterogeneous resource unit is low. Therefore, it is not recommended to directly input it into the virtual heterogeneous resource unit. At this time, the combination pairing method will be used to combine it with other single tasks or multiple tasks with low similarity degree in order to obtain a combined task with a high similarity degree. The combined task will be input according to the direct pairing method. At the same time, within a set time for a single task, such as within two seconds, if a task with a low similarity degree fails to be combined, it will be input regardless of the degree threshold to avoid the task being idle for too long.

[0042] The load rate of the virtual heterogeneous resource unit is compared with the load threshold of the virtual heterogeneous resource unit, and the total energy supply scale of the virtual heterogeneous resource unit is dynamically scaled based on the comparison result. During the scaling process, the heterogeneous resource ratio is maintained. When the load rate of all virtual heterogeneous resource units is higher than 75% for 3 consecutive minutes, and the global load rate is less than the maximum value in the second load threshold, the scale of the virtual unit will be expanded proportionally. When the load rate of the virtual heterogeneous resource unit is lower than 25% for 5 consecutive minutes, the scale of the virtual unit will be reduced proportionally. When several virtual heterogeneous resource units in a single physical machine synchronously reduce their scales, at this time, the voltage reduction and frequency reduction of the single physical machine can be triggered, so that the single physical machine can achieve energy saving while being able to process tasks.

[0043] The actual scale of different heterogeneous resources can be regarded as the maximum operating load of the heterogeneous resources in the normal state. The total heterogeneous scale refers to the lowest load at which several virtual heterogeneous resource units can process tasks and maintain efficiency after reduction. By comparing the two, the dynamic load adjustment of the source devices of the heterogeneous resources is carried out, so that after adjustment, the physical machine can meet the lowest load at which tasks can be processed and efficiency can be maintained.

[0044] Comparison coefficient is between 0.8 and 1.2. Through the comparison value the dynamic load adjustment of the heterogeneous resources is directly carried out. The comparison coefficient will change. It is necessary to ensure that the unit energy consumption of several heterogeneous resources after modification can meet the processing of tasks in the physical machine. The total heterogeneous scale is compared with the total physical machine scale to obtain a set of comparison values. The set of comparison values is the separate comparison of the total scale of each heterogeneous resource in the total heterogeneous scale and the total scale of each heterogeneous resource in the physical machine to obtain a set of comparison values.

[0045] Embodiment 1: In the scenario of three resources, namely CPU, GPU, and memory, with resource types and parameter settings. The normalized CPU load rate L1 = 0.7, the energy consumption ratio A1 = 0.5, the normalized GPU load rate L2 = 0.6, the energy consumption ratio A2 = 0.3, the normalized memory load rate L3 = 0.5, and the energy consumption ratio A3 = 0.2. Calculate the global load rate as follows: =(0.5×0.7)+(0.3×0.6)+(0.2×0.5)=0.35 + 0.18 + 0.10 = 0.63. The global load rate is 0.63. Compare the global load rate with the second load threshold to determine the load situation.

[0046] Embodiment 2: Parameter settings in the virtual heterogeneous resource unit. Resource allocation weights: α = 0.4 (CPU), β = 0.3 (memory), γ = 0.2 (disk), θ = 0.1 (network). Normalized load rates: Processor load = 0.8 (high CPU load rate), memory load = 0.6, disk load = 0.4, network bandwidth utilization = 0.5. The calculation process =(0.4×0.8)+(0.3×0.6)+(0.2×0.4)+(0.1×0.5)=0.32 + 0.18 + 0.08 + 0.05 = 0.63. Calculate the load data in this virtual heterogeneous resource unit. At the same time, combine the load data in this virtual heterogeneous resource unit with the ratio of the scale of this virtual heterogeneous resource unit to the total scale of the physical machine, so as to obtain the impact of the load data in this virtual heterogeneous resource unit on the total load of the physical machine.

[0047] Through the establishment of the first framework, the first framework supports collaborative management and scheduling among multiple heterogeneous resources. The first framework is for a single physical machine. Through load calculation, comparison, and trigger conditions, it dynamically adjusts the scale of the virtual heterogeneous resource unit. Based on the reduced scale, it simultaneously performs different voltage reduction and frequency reduction on each heterogeneous resource in the single physical machine, reducing the energy consumption of the physical machine while performing voltage reduction and frequency reduction on each heterogeneous resource separately, ensuring the effective collaborative scheduling of heterogeneous resources. At the same time, the second framework is for the server cluster. According to the comparison of the real-time global load rate with the load threshold, when meeting the specific second trigger condition, through task migration and resource reallocation, the physical machine is kept in a normal state, avoiding being in a high-energy consumption state, and optimizing the energy consumption of the entire cluster.

[0048] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An energy-saving server operation method based on dynamic load regulation, characterized in that The method for operating the energy-saving server includes: Generating a number of virtual heterogeneous resource units from a number of heterogeneous resources in a single physical machine through pooling and virtualization methods, where the ratios of the heterogeneous resources in each virtual heterogeneous resource unit are different; Establishing a second load threshold for the physical machine and a first load threshold for the virtual heterogeneous resource units based on the device parameters of the physical machine; Constructing a first framework for optimizing the energy consumption of a single physical machine and a second framework for optimizing the energy consumption of a server cluster; The first framework includes: Obtain the load rates of several virtual heterogeneous resource units and the real-time global load rate through load calculation , and establish the first trigger condition for physical machine energy consumption optimization; Compare the load ratios of several virtual heterogeneous resource units with the load thresholds of the virtual heterogeneous resource units through a comparison method to obtain the real-time global load ratio Compare with the second load threshold to obtain the first comparison result; Inputting the first comparison result into a first trigger condition to determine whether to trigger. If it is determined to trigger, adjusting the scale of a number of virtual heterogeneous resource units in the physical machine based on the elastic resource scaling technology to obtain the total heterogeneous scale of the combined virtual heterogeneous resource units after adjustment. Comparing the total heterogeneous scale with the total scale of the physical machine to obtain a set of comparison values, and dynamically adjusting the load of a number of heterogeneous resources in the single physical machine through the set of comparison values and adjustment methods; The comparison value set includes the comparison values between the actual scale of several heterogeneous resources in the physical machine and the virtual scale of several heterogeneous resources in the total heterogeneous scale. , where i represents the i-th heterogeneous resource; The adjustment method includes: by dynamically adjusting the load of a number of heterogeneous resource source devices, the energy consumption comparison value between the unit energy consumption of a number of heterogeneous resources after modification and the unit energy consumption of the heterogeneous resources before modification , the energy consumption comparison value and the comparison value are positively correlated.

2. The energy-saving server operation method based on dynamic load adjustment according to claim 1, wherein: The second framework includes: Obtain the real-time global load rate of a physical machine through load calculation ; Real-time global load rate is achieved through a comparison method A second comparison result is obtained by comparing with a second load threshold Establishing a second trigger condition for the energy consumption of the server cluster, and determining whether to trigger the second trigger condition through the second comparison result. If it is determined to trigger; Obtain the real-time global load rate The difference from the second load threshold is used to establish a safety value, and through the real-time global load rate The tasks in the physical machine are proportionally divided to obtain the task scale ratio. Based on the difference, the safety value, and the task scale ratio, the scale of the task to be processed is obtained, and the task to be processed is obtained from the virtual heterogeneous resource unit with the highest load rate of the virtual heterogeneous resource units Transmitting the tasks to be processed in the physical machine between physical machines through a feature selection method; The feature selection method includes selecting, based on the scale of the task to be processed, a physical machine from the server cluster that can receive the task to be processed and will not cause the real-time global load rate after reception to exceed the second load threshold, and transporting the task to be processed to the selected physical machine; Processing the tasks to be processed, and the task processing includes: Performing feature processing on the tasks to be processed to obtain the collaborative degree requirement of the tasks to be processed. The collaborative degree requirement represents the demand of the tasks for different heterogeneous computing resources. Establishing a feature vector of the tasks based on the collaborative degree requirement, and establishing a ratio feature vector of the virtual heterogeneous resource units based on the ratios of the heterogeneous resources in the virtual heterogeneous resource units; Inputting the tasks to be processed into the virtual heterogeneous resource units with corresponding relationships through a similarity selection method.

3. A method for operating an energy-saving server based on dynamic load adjustment according to claim 2, characterized in that: The similarity selection method includes a direct pairing method and a combined pairing method; The direct pairing method includes: Calculating the similarity between the feature vector of the task to be processed and the ratio feature vectors of a number of virtual heterogeneous resource units in the physical machine to obtain a similarity degree value. Selecting the virtual heterogeneous resource unit feature with the highest similarity degree value for input of the task to be processed, and establishing a degree threshold. Tasks with similarity degree values all lower than the degree threshold are input through the combined pairing method; The combined pairing method includes: Performing intelligent pairing and combination between the tasks to be processed with similarity degree values all lower than the degree threshold to obtain a combined task and a feature vector of the combined task. Calculating the similarity between the feature vector of the combined task and the feature vectors of a number of virtual heterogeneous resource units in the physical machine to obtain a set of similarity degree values, and selecting the virtual heterogeneous resource unit feature with the highest similarity degree value for input of the combined task; Establishing a time threshold. For tasks in the combined pairing method, if they still cannot be combined when exceeding the time threshold, the degree threshold is ignored, and the virtual heterogeneous resource unit feature with the highest similarity degree value is directly selected for input of the combined task.

4. A method for operating an energy-saving server based on dynamic load adjustment according to claim 1, characterized in that: The load calculation is performed through normalization processing and load operation; The normalization process includes establishing a normalization formula: , where the minimum load rate and the maximum load rate are respectively the theoretical minimum load rate and the theoretical maximum load rate of each heterogeneous resource; Calculating the load of each heterogeneous resource through a normalization formula to obtain the normalized load rate of each heterogeneous resource.

5. A method for operating an energy-saving server based on dynamic load adjustment according to claim 1, characterized in that: The elastic scaling technology includes dynamically scaling the scales of various heterogeneous resources of the virtual heterogeneous resource unit, and keeping the heterogeneous resource ratio unchanged during the scaling process; The dynamic load adjustment includes adjusting the voltage and frequency of the heterogeneous resources and controlling the source device to be in a low-power state; The comparison value set includes the comparison values between the actual scales of various heterogeneous resources in the physical machine and the virtual scales of various heterogeneous resources in the total heterogeneous scale; Energy consumption comparison value and the comparison value are positively correlated, and a comparison coefficient is established , .

6. The energy-saving server operation method based on dynamic load adjustment according to claim 4, wherein: The load operation includes: Real-time global load rate ; , N represents the number of groups in the heterogeneous resource grouping set, represents the overall occupancy ratio of the i-th heterogeneous resource in the physical machine, represents the normalized load rate of the i-th heterogeneous computing resource, represents the real-time global load rate calculated based on the heterogeneous resource load; The overall occupancy ratio is replaced by the energy consumption ratio. The energy consumption ratio represents the energy consumption weights of several heterogeneous resources in the physical machine, and the sum of the energy consumption ratios is 1.

7. A method for operating an energy-saving server based on dynamic load regulation according to claim 2, characterized in that: The second trigger condition includes: Real-time global load rate Continuously greater than or equal to the maximum value of the second load threshold and maintained for at least three minutes.

8. A method for operating an energy-saving server based on dynamic load adjustment according to claim 6, wherein: The load operation includes: Establish the formula for the load rate of virtual heterogeneous resource units ; , represents the load rate of the j-th virtual heterogeneous resource unit, the normalized processor load , the normalized memory load , the normalized disk load , the normalized network bandwidth utilization ; , , and represent the energy consumption ratios of virtual heterogeneous resource units, and the sum of the energy consumption ratios satisfies 1; , Y represents the number of virtual heterogeneous resource units in a physical machine, represents the ratio of the scale of the Y-th virtual heterogeneous resource unit to the total scale of the physical machine, represents the real-time global load rate calculated based on the load of virtual heterogeneous resource units.

9. A method for operating an energy-saving server based on dynamic load regulation according to claim 1, characterized in that: The first trigger condition includes: Condition 1: The virtual heterogeneous resource unit is continuously lower than the virtual heterogeneous resource unit threshold and remains so for at least three minutes; Condition 2: Real-time global load rate Continuously less than the minimum value of the second load threshold and maintained for at least three minutes; Condition 1 and Condition 2 need to be satisfied simultaneously.

10. An energy-saving server operation system based on dynamic load regulation adopts an energy-saving server operation method according to any one of claims 1-9, characterized in that: The energy-saving server operation system includes: Resource allocation module: Generating a number of virtual heterogeneous resource units from several heterogeneous resources in a single physical machine through pooling and virtualization methods. The heterogeneous resource ratios in each virtual heterogeneous resource unit are different; Threshold establishment module: Establishing the second load threshold of the physical machine and the first load threshold of the virtual heterogeneous resource unit according to the device parameters of the physical machine; Energy consumption optimization module: The energy consumption optimization module includes a first optimization module and a second optimization module; Optimization Module 1: Obtain the load rates of several virtual heterogeneous resource units and the real-time global load rate through load calculation , and establish the first trigger condition for physical machine energy consumption optimization: By means of a comparison method, a comparison is made between the load ratios of a number of virtual heterogeneous resource units and the load thresholds of the virtual heterogeneous resource units, and the real-time global load ratio and the second load threshold, and a first comparison result is obtained; The first comparison result is input to the first trigger condition to determine whether to trigger. If it is determined to trigger, based on the elastic resource scaling technology, the scales of several virtual heterogeneous resource units in the physical machine are reduced to obtain the total heterogeneous scale formed by the combination of the reduced several virtual heterogeneous resource units. The total heterogeneous scale is compared with the total scale of the physical machine to obtain a comparison value set, and the dynamic load adjustment of several heterogeneous resource source devices in a single physical machine is performed through the comparison value set and the adjustment method; Second optimization module: Obtain the real-time global load rate of the physical machine through load calculation ; Implementing real-time global load rate through a comparison method Obtain the second comparison result by comparing with the second load threshold; Establishing a second trigger condition for the energy consumption of the server cluster, and determining whether to trigger the second trigger condition through the second comparison result. If it is determined to trigger: Obtain the real-time global load rate The difference from the second load threshold is used to establish a safety value, and based on the real-time global load rate The tasks in the physical machine are proportionally divided to obtain the task scale ratio. According to the difference value, the safety value and the task scale ratio, the scale of the tasks to be processed is obtained, and the tasks to be processed are obtained from the virtual heterogeneous resource unit with the highest load rate among the virtual heterogeneous resource units The tasks to be processed in the physical machine are transmitted between physical machines through the feature selection method.

Images