Power management method and device based on heterogeneous computing power scheduling

Abstract

The application discloses a power management method and device based on heterogeneous computing power scheduling, relates to the technical field of heterogeneous computing power, and comprises the following steps: calculating the computing power evaluation coefficient of heterogeneous computing power and the computing weight of each kind of heterogeneous computing power for different power management tasks, and calculating the regular weight of each kind of heterogeneous computing power for different power management tasks; distributing the to-be-allocated power management task to the idle heterogeneous computing power with the maximum regular weight, if there is no idle heterogeneous computing power, then calculating the real-time weight of each kind of heterogeneous computing power for the to-be-allocated power management task, calculating the comprehensive weight of each kind of heterogeneous computing power for the to-be-allocated power management task, and distributing the to-be-allocated power management task to the heterogeneous computing power with the maximum comprehensive weight; calculating the busy coefficient of each kind of power management task, calculating the maximum difference value of the busy coefficients between each kind of power management task, denoted as the busy coefficient difference, and issuing an heterogeneous computing power optimization instruction outward. The utilization rate of overall computing power resources is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of heterogeneous computing, specifically to a power management method and device based on heterogeneous computing scheduling. BACKGROUND

[0002] With the rapid development of cloud computing, big data and artificial intelligence technology, heterogeneous computing scheduling has become an increasingly important topic. Heterogeneous computing refers to the computing power provided by different types of computing resources (such as CPU, GPU, FPGA, etc.) when performing parallel computing. These heterogeneous computing resources have different architectures and performance characteristics, and can handle different types of computing tasks. In the field of power management, traditional management methods often ignore the diversity and complexity of computing resources, resulting in unreasonable allocation of power resources, low energy efficiency, and other problems. Therefore, combining heterogeneous computing scheduling technology to achieve intelligent and efficient power management has become an important direction of current research and application.

[0003] In Chinese patent application CN119149230A, a distributed heterogeneous computing inference task dynamic scheduling method and system are disclosed, which includes a computing power architecture adaptation module, a task scheduling module, a model updating module, a capability service module, a dynamic scheduling module, a resource evaluation module, and a state perception module. The computing power architecture adaptation module is used to identify and match local and remote heterogeneous computing power architectures. The task scheduling module is used to select appropriate computing power terminals from a distributed heterogeneous computing environment. The model updating module is used to update AI models in real time. The capability service module is used to provide a unified AI model service calling interface. The dynamic scheduling module is used to dynamically allocate computing power resources. The resource evaluation module is used to monitor and evaluate the state of computing power resources. The state perception module is used to monitor the health status of computing power nodes in real time and respond to failures in a timely manner. This improves the scheduling efficiency of computing power resource scheduling and ensures the stable operation of inference tasks.

[0004] In the above patent application, device failures or computing power fluctuations are considered to cause interruption or delay of inference tasks, but the allocation is still based on existing heterogeneous computing power, and optimization of computing power according to task conditions is not considered. When the task demand does not match the allocated computing power, some computing power resources may be idle or underutilized. For example, if high computing power resources are allocated to low-load tasks, these resources will not be fully utilized, resulting in waste. Conversely, if low computing power resources are allocated to high-load tasks, the task processing speed may be slow, and even real-time requirements may not be met, thereby affecting the overall system performance.

[0005] To this end, the present application provides a power management method and device based on heterogeneous computing scheduling. SUMMARY

[0006] The technical problems solved by the present application

[0007] In view of the deficiencies of the prior art, the present application provides a power management method and device based on heterogeneous computing power scheduling, which obtains the average waiting time De j , processing time Ci j , energy consumption Xh j and total number per unit time Sg j of each power management task from the running records of heterogeneous computing power, calculates the busy factor Fm j of each power management task, and calculates the maximum difference of the busy factors Fm j between each power management task, denoted as busy factor difference Cz. When the busy factor difference exceeds 2ΔCz, an heterogeneous computing power optimization instruction is issued, which can comprehensively reflect the load condition of the task and the tension of the computing power resource, reveal the degree of load imbalance between tasks, and help to redistribute the computing power resource, so that the high-load task gets more computing power support and the low-load task releases the computing power resource, thereby improving the utilization rate of the overall computing power resource, and solving the technical problems described in the background art.

[0008] (II) Technical solutions

[0009] To achieve the above object, the present application is implemented by the following technical solutions: a power management method based on heterogeneous computing power scheduling, comprising the following steps:

[0010] obtaining the computing power Sl i , bandwidth Dk i and computing density Sm i of each heterogeneous computing power, calculating the computing power evaluation coefficient Spg i of the heterogeneous computing power, and making each heterogeneous computing power execute the same kind of power management task, using a monitoring tool to record the processing time Cs ij and energy consumption Nu ij of the same power management task in different heterogeneous computing powers, calculating the computing weight α ij of each heterogeneous computing power to different power management tasks, calculating the regular weight μ ij of each heterogeneous computing power to different power management tasks according to the computing power evaluation coefficient Spg i of the heterogeneous computing power and the computing weight α ij of each heterogeneous computing power to different power management tasks;

[0011] allocating the power management task to be allocated to the idle heterogeneous computing power with the maximum regular weight μ ij , and if there is no idle heterogeneous computing power, obtaining the waiting time Da iand the real-time requirement level Ja of the power management task to be allocated, calculate the real-time weight p of each heterogeneous computing power on the power management task to be allocated i , according to the real-time weight p of each heterogeneous computing power on the power management task to be allocated i and the conventional weight m of each heterogeneous computing power on the power management task to be allocated i , calculate the comprehensive weight x of each heterogeneous computing power on the power management task to be allocated i , allocate the power management task to be allocated to the heterogeneous computing power with the maximum comprehensive weight x i .

[0012] From the running records of the heterogeneous computing powers, obtain the average waiting time De j , processing time Ci j , energy consumption Xh j and total number per unit time Sg j of each power management task, calculate the busy factor Fm j of each power management task, and calculate the maximum difference of the busy factors Fm j between each power management task, denoted as the busy factor difference Cz. When the busy factor difference exceeds 2ΔCz, an optimization instruction of the heterogeneous computing power is sent out.

[0013] Further, obtain the computing power Sl i , bandwidth Dk i and computing density Sm i of each heterogeneous computing power, and calculate the computing power evaluation coefficient Spg i of the heterogeneous computing power:

[0014]

[0015] wherein i represents the sequential number of each heterogeneous computing power, i=1, 2, …, n, and n represents the total number of all heterogeneous computing powers; max(Sl i ) represents the maximum value of the computing power of all heterogeneous computing powers, min(Sl i ) represents the minimum value of the computing power of all heterogeneous computing powers, max(Dk i ) represents the maximum value of the bandwidth of all heterogeneous computing powers, min(Dk i ) represents the minimum value of the bandwidth of all heterogeneous computing powers, max(Sm i ) represents the maximum value of the computing density of all heterogeneous computing powers, and min(Sm i ) represents the minimum value of the computing density of all heterogeneous computing powers.

[0016] Computing power refers to the number of operations that a computing unit can perform per second, usually expressed as FLOPs (Floating Point Operations Per Second). In heterogeneous computing, the computing power of different computing units can vary greatly. For example, GPUs usually have higher parallel computing capabilities than CPUs, so they perform better when processing large amounts of parallel data.

[0017] Bandwidth refers to the rate at which data is transmitted between a computing unit and memory or external devices, usually expressed as bytes per second (Byte / s). In heterogeneous computing, bandwidth is crucial for data transmission speed and overall system performance. High bandwidth ensures that computing units can efficiently access and process data, while low bandwidth can cause performance bottlenecks.

[0018] Computing density in heterogeneous computing refers to the computing power that can be provided within a given space or power consumption. It is usually related to the ratio of computing power and power consumption, but may also involve other factors such as heat dissipation performance, physical size, etc. A high computing density computing unit can provide higher computing power within limited resources.

[0019] Further, make each heterogeneous computing power perform the same kind of power management task, use monitoring tools to record the processing time of the same power management task on different heterogeneous computing powers Cs ij and energy consumption Nu ij , calculate the computing weight of each heterogeneous computing power for different power management tasks α ij :

[0020]

[0021] where j represents the sequential number of each power management task, j = 1, 2, …, m, and m represents the total number of all power management tasks.

[0022] Further, obtain the computing power evaluation coefficient Spg i of the heterogeneous computing power and the computing weight of each heterogeneous computing power for different power management tasks α ij , calculate the regular weight of each heterogeneous computing power for different power management tasks μ ij :

[0023]

[0024] Further, assign the to-be-assigned power management task to the idle heterogeneous computing power with the largest regular weight μ ij , and record the real-time processing time. The difference between the average processing time of the current heterogeneous computing power for the same kind of power management task and the real-time processing time is recorded as the current heterogeneous computing power waiting time Da i .

[0025] Further, if there is no idle heterogeneous computing power, the waiting time Da of each heterogeneous computing power is obtained i and the real-time demand level Ja of the power management task to be allocated, the real-time weight p of each heterogeneous computing power for the power management task to be allocated is calculated i :

[0026]

[0027] The real-time demand level of the power management task is divided according to the type of the power management task, and the specific division is as follows:

[0028] The planning task is of low real-time demand, and the real-time demand level is recorded as 1. The running monitoring is of medium real-time demand, and the real-time demand level is recorded as 2. The real-time scheduling is of high real-time demand, and the real-time demand level is recorded as 3. The fault handling is of extremely high real-time demand, and the real-time demand level is recorded as 4.

[0029] Further, the real-time weight p of each heterogeneous computing power for the power management task to be allocated is obtained i and the conventional weight m of each heterogeneous computing power for the power management task to be allocated i , the comprehensive weight x of each heterogeneous computing power for the power management task to be allocated is calculated i :

[0030] x i =p i * m i

[0031] When there is no idle heterogeneous computing power, the power management task to be allocated is allocated to the heterogeneous computing power with the maximum comprehensive weight x i , and the waiting time of the current heterogeneous computing power is updated, that is, the original waiting time of the heterogeneous computing power + the average processing time of the same type of power management task.

[0032] Further, the average waiting time De j , the processing time Ci j , the energy consumption Xh j and the total number per unit time Sg j of each power management task are obtained from the running record of the heterogeneous computing power, and the busy coefficient Fm of each power management task is calculated j :

[0033]

[0034] max(De j ) represents the maximum value of the average waiting time of all power management tasks, and max(Ci j) represents the maximum value of processing time of all power management tasks, max(Xh j ) represents the maximum value of energy consumption of all power management tasks, max(Sg j ) represents the maximum value of total number of unit time of all power management tasks.

[0035] Further, the busy factor Fm j of each power management task is obtained, the maximum difference of busy factors Fm j between each power management task is calculated, denoted as busy factor difference Cz, when the busy factor difference exceeds 2ΔCz, an heterogeneous computing power optimization instruction is sent out;

[0036] Wherein, ΔCz represents the mean value of busy factor difference of all heterogeneous computing power systems.

[0037] The power management device based on heterogeneous computing power scheduling comprises:

[0038] The conventional weight calculation module obtains the computing power Sl i , bandwidth Dk i and computing density Sm i of each heterogeneous computing power, calculates the computing power evaluation coefficient Spg i of the heterogeneous computing power, and makes each heterogeneous computing power execute the same kind of power management task, uses a monitoring tool to record the processing time Cs ij and energy consumption Nu ij of the same power management task in different heterogeneous computing powers, calculates the computing weight α ij of each heterogeneous computing power to different power management tasks, calculates the conventional weight μ i of each heterogeneous computing power to different power management tasks according to the computing power evaluation coefficient Spg ij of the heterogeneous computing power and the computing weight α ij of each heterogeneous computing power to different power management tasks;

[0039] The computing power allocation module allocates the power management task to be allocated to the idle heterogeneous computing power with the maximum conventional weight μ ij , if there is no idle heterogeneous computing power, obtains the waiting time Da i of each heterogeneous computing power and the real-time requirement level Ja of the power management task to be allocated, calculates the real-time weight ρ i of each heterogeneous computing power to the power management task to be allocated, calculates the comprehensive weight x i of each heterogeneous computing power to the power management task to be allocated according to the real-time weight ρ i of each heterogeneous computing power to the power management task to be allocated and the conventional weight μ i of each heterogeneous computing power to the power management task to be allocated, and allocates the power management task to be allocated to the comprehensive weight xi The largest heterogeneous computing power;

[0040] The heterogeneous computing power optimization module obtains the average waiting time (De) for each power management task from the operation records of heterogeneous computing power. j Processing time Ci j Energy consumption Xh j and total quantity per unit time Sg j Calculate the busy factor Fm for each type of power management task. j And calculate the busy factor Fm between each type of power management task. j The maximum difference is denoted as the busy coefficient difference Cz. When the busy coefficient difference exceeds 2ΔCz, a heterogeneous computing power optimization instruction is issued.

[0041] (III) Beneficial Effects

[0042] This invention provides a power management method and apparatus based on heterogeneous computing power scheduling, which has the following beneficial effects:

[0043] 1. Obtain the computing power Sl of each heterogeneous computing power i Bandwidth Dk i and calculate density Sm i Calculate the computing power evaluation coefficient Spg for heterogeneous computing power i Furthermore, each heterogeneous computing power performs the same type of power management task, and monitoring tools are used to record the processing time (Cs) of the same power management task on different heterogeneous computing power systems. ij and energy consumption Nu ij Calculate the computational weight α of each heterogeneous computing power for different power management tasks. ij Based on the computing power evaluation coefficient Spg of heterogeneous computing power i And the computational weight α of each heterogeneous computing power for different power management tasks ij Calculate the conventional weight μ of each heterogeneous computing power for different power management tasks. ij This allows for a better understanding of how each type of heterogeneous computing power performs on different tasks, thereby supporting more diverse task requirements. It also helps optimize resource allocation when deploying tasks, ensuring that tasks are assigned to the most suitable heterogeneous computing power for execution.

[0044] 2. Assign the pending power management tasks to the regular weight μ ij The maximum available heterogeneous computing power; if no available heterogeneous computing power is available, then obtain the waiting time (Da) for each heterogeneous computing power. i Based on the real-time requirement level Ja of the power management tasks to be assigned, calculate the real-time weight ρ of each heterogeneous computing power for the power management tasks to be assigned. i Based on the real-time weight ρ of each heterogeneous computing power to be allocated power management tasks iAnd the regular weight μ for each heterogeneous computing power to be assigned to power management tasks i Calculate the comprehensive weight x for each type of heterogeneous computing power to be assigned power management tasks. i The pending power management tasks will be assigned to a comprehensive weight x. i The maximum heterogeneous computing power takes into account the current load and task urgency of heterogeneous computing power, which helps to balance resource allocation, avoid the situation where some computing resources are overloaded while other resources are idle, and can flexibly cope with tasks with different real-time requirements, ensuring that the system can maintain efficient operation under different load conditions.

[0045] 3. Obtain the average waiting time (De) for each power management task from the operation records of heterogeneous computing power. j Processing time Ci j Energy consumption Xh j and total quantity per unit time Sg j Calculate the busy factor Fm for each type of power management task. j And calculate the busy factor Fm between each type of power management task. j The maximum difference is denoted as the busy coefficient difference Cz. When the busy coefficient difference exceeds 2ΔCz, a heterogeneous computing power optimization instruction is issued. This instruction can comprehensively reflect the task load and the tightness of computing resources, revealing the degree of load imbalance between tasks. The issuance of the optimization instruction helps to redistribute computing resources, enabling high-load tasks to receive more computing power support, while low-load tasks release computing resources, thereby improving the overall utilization rate of computing resources. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the power management method based on heterogeneous computing power scheduling according to the present invention.

[0047] Figure 2 This is a schematic diagram of the power management device based on heterogeneous computing power scheduling according to the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Please see Figure 1 This invention provides a power management method based on heterogeneous computing power scheduling, comprising the following steps:

[0050] Step 1: Obtain the computing power Sl for each heterogeneous computing power i Bandwidth Dki and the computing density Sm i , the power evaluation coefficient Spg of the heterogeneous computing power is calculated i , and the same power management task is performed by each heterogeneous computing power, the processing time Cs of the same power management task in different heterogeneous computing powers is recorded using a monitoring tool ij and the energy consumption Nu ij , the computing weight a of each heterogeneous computing power to different power management tasks is calculated ij , the power evaluation coefficient Spg of the heterogeneous computing power is calculated i and the computing weight a of each heterogeneous computing power to different power management tasks ij , the general weight m of each heterogeneous computing power to different power management tasks is calculated ij .

[0051] The step one includes the following contents:

[0052] Step 101, the computing power Sl of each heterogeneous computing power is obtained by consulting the hardware specification or the performance parameters provided by the manufacturer i , the bandwidth Dk i and the computing density Sm i , the power evaluation coefficient Spg of the heterogeneous computing power is calculated i :

[0053]

[0054] wherein i represents the sequential number of each heterogeneous computing power, i = 1, 2, …, n, n represents the total number of all heterogeneous computing powers. max(Sl i ) represents the maximum value of the computing power of all heterogeneous computing powers, min(Sl i ) represents the minimum value of the computing power of all heterogeneous computing powers, max(Dk i ) represents the maximum value of the bandwidth of all heterogeneous computing powers, min(Dk i ) represents the minimum value of the bandwidth of all heterogeneous computing powers, max(Sm i ) represents the maximum value of the computing density of all heterogeneous computing powers, min(Sm i ) represents the minimum value of the computing density of all heterogeneous computing powers.

[0055] The computing power refers to the number of operations that can be performed by a computing unit per second, usually represented as FLOPs (Floating Point Operations Per Second). In heterogeneous computing, the computing power of different computing units may differ greatly, for example, GPUs usually have higher parallel computing capabilities than CPUs, so they perform higher computing power when processing a large amount of parallel data.

[0056] Bandwidth refers to the rate of data transmission between the computing unit and the memory or external devices, usually expressed as the number of bytes transmitted per second (Byte / s). In heterogeneous computing, bandwidth is crucial for data transmission speed and overall system performance. High bandwidth can ensure that the computing unit can efficiently access and process data, while low bandwidth may cause performance bottlenecks.

[0057] Computing density in heterogeneous computing refers to the computing power that can be provided within a given space or power consumption. It is usually related to the ratio of computing power and power consumption, but may also involve other factors such as heat dissipation performance, physical size, etc. A high computing density computing unit can provide higher computing power within limited resources.

[0058] Step 102, make each heterogeneous computing power perform the same kind of power management task, use monitoring tools to record the processing time Cs of the same power management task on different heterogeneous computing powers ij and energy consumption Nu ij , calculate the computing weight α of each heterogeneous computing power for different power management tasks ij :

[0059]

[0060] where j represents the sequential number of each power management task, j = 1, 2, …, m, m represents the total number of all power management tasks.

[0061] Step 103, obtain the computing power evaluation coefficient Spg of the heterogeneous computing power i and the computing weight α of each heterogeneous computing power for different power management tasks ij , calculate the regular weight μ of each heterogeneous computing power for different power management tasks ij :

[0062]

[0063] In use, in combination with the contents in steps 101 to 103:

[0064] obtain the computing power Sl of each heterogeneous computing power i , bandwidth Dk i and computing density Sm i , calculate the computing power evaluation coefficient Spg of the heterogeneous computing power i , make each heterogeneous computing power perform the same kind of power management task, use monitoring tools to record the processing time Cs of the same power management task on different heterogeneous computing powers ij and energy consumption Nu ij , calculate the computing weight α of each heterogeneous computing power for different power management tasks ij , according to the computing power evaluation coefficient Spg of the heterogeneous computing poweri and the calculation weight a of each heterogeneous computing power on different power management tasks ij , calculating the regular weight m of each heterogeneous computing power on different power management tasks ij It can better understand the performance of each heterogeneous computing power on different tasks, support more diverse task requirements, help optimize resource allocation when deploying tasks, and ensure that tasks are assigned to the most suitable heterogeneous computing power for execution.

[0065] Step two, allocate the power management task to be allocated to the regular weight m ij The largest idle heterogeneous computing power, if there is no idle heterogeneous computing power, get the waiting time Da of each heterogeneous computing power i and the real-time demand level Ja of the power management task to be allocated, calculate the real-time weight p of each heterogeneous computing power on the power management task to be allocated i , according to the real-time weight p of each heterogeneous computing power on the power management task to be allocated i and the regular weight m of each heterogeneous computing power on the power management task to be allocated i , calculate the comprehensive weight x of each heterogeneous computing power on the power management task to be allocated i , allocate the power management task to be allocated to the comprehensive weight x i The largest heterogeneous computing power.

[0066] The step two includes the following contents:

[0067] Step 201, allocate the power management task to be allocated to the regular weight m ij The largest idle heterogeneous computing power, and record the real-time processing time in real time, and record the difference between the average processing time of the current heterogeneous computing power processing the same kind of power management task and the real-time processing time as the current heterogeneous computing power waiting time Da i .

[0068] Step 202, if there is no idle heterogeneous computing power, get the waiting time Da of each heterogeneous computing power i and the real-time demand level Ja of the power management task to be allocated, calculate the real-time weight p of each heterogeneous computing power on the power management task to be allocated i :

[0069]

[0070] The real-time demand level of the power management task is divided according to the type of the power management task, and the specific division is as follows:

[0071] The planning task has low real-time requirements, with a real-time requirement level of 1; the operation monitoring has medium real-time requirements, with a real-time requirement level of 2; the real-time scheduling has high real-time requirements, with a real-time requirement level of 3; and the fault handling has extremely high real-time requirements, with a real-time requirement level of 4.

[0072] Step 203: Obtain the real-time weight ρ of each heterogeneous computing power for the power management task to be allocated. i And the regular weight μ for each heterogeneous computing power to be assigned to power management tasks i Calculate the comprehensive weight x for each type of heterogeneous computing power to be assigned power management tasks. i :

[0073] x i =ρ i *μ i

[0074] When there is no available heterogeneous computing power, the power management tasks to be allocated will be assigned to the comprehensive weight x. i The maximum heterogeneous computing power is determined, and the current heterogeneous computing power waiting time is updated, which is the original heterogeneous computing power waiting time plus the average processing time of the same type of power management tasks to be assigned.

[0075] When using this method, refer to steps 201 to 203:

[0076] Assign the pending power management tasks to the regular weight μ ij The maximum available heterogeneous computing power; if no available heterogeneous computing power is available, then obtain the waiting time (Da) for each heterogeneous computing power. i Based on the real-time requirement level Ja of the power management tasks to be assigned, calculate the real-time weight ρ of each heterogeneous computing power for the power management tasks to be assigned. i Based on the real-time weight ρ of each heterogeneous computing power to be allocated power management tasks i And the regular weight μ for each heterogeneous computing power to be assigned to power management tasks i Calculate the comprehensive weight x for each type of heterogeneous computing power to be assigned power management tasks. i The pending power management tasks will be assigned to a comprehensive weight x. i The maximum heterogeneous computing power takes into account the current load and task urgency of heterogeneous computing power, which helps to balance resource allocation, avoid the situation where some computing resources are overloaded while other resources are idle, and can flexibly cope with tasks with different real-time requirements, ensuring that the system can maintain efficient operation under different load conditions.

[0077] Step 3: Obtain the average waiting time (De) for each power management task from the operation records of heterogeneous computing power. j Processing time Ci j Energy consumption Xh jand total quantity per unit time Sg j Calculate the busy factor Fm for each type of power management task. j And calculate the busy factor Fm between each type of power management task. j The maximum difference is denoted as the busy coefficient difference Cz. When the busy coefficient difference exceeds 2ΔCz, a heterogeneous computing power optimization instruction is issued.

[0078] Step three includes the following:

[0079] Step 301: Obtain the average waiting time (De) for each power management task from the operation records of heterogeneous computing power. j Processing time Ci j Energy consumption Xh j and total quantity per unit time Sg j Calculate the busy factor Fm for each type of power management task. j :

[0080]

[0081] Where max(De) j ) represents the maximum average waiting time for all power management tasks, max(Ci j ) represents the maximum processing time for all power management tasks, max(Xh) j ) represents the maximum energy consumption of all power management tasks, max(Sg) j This represents the maximum total number of all power management tasks per unit of time.

[0082] Step 302: Obtain the busy coefficient Fm for each type of power management task. j Calculate the busy factor Fm between each type of power management task. j The maximum difference is denoted as the busy coefficient difference Cz. When the busy coefficient difference exceeds 2ΔCz, a heterogeneous computing power optimization instruction is issued.

[0083] Where ΔCz represents the average of the busy coefficient differences among all heterogeneous computing power systems.

[0084] When using this method, refer to steps 301 and 302:

[0085] The average waiting time for each power management task is obtained from the operation records of heterogeneous computing power. j Processing time Ci j Energy consumption Xh j and total quantity per unit time Sg j Calculate the busy factor Fm for each type of power management task. j And calculate the busy factor Fm between each type of power management task. jThe maximum difference is denoted as the busy coefficient difference Cz. When the busy coefficient difference exceeds 2ΔCz, a heterogeneous computing power optimization instruction is issued. This instruction can comprehensively reflect the task load and the tightness of computing resources, revealing the degree of load imbalance between tasks. The issuance of the optimization instruction helps to redistribute computing resources, enabling high-load tasks to receive more computing power support, while low-load tasks release computing resources, thereby improving the overall utilization rate of computing resources.

[0086] Please see Figure 2 This invention provides a power management device based on heterogeneous computing power scheduling, comprising:

[0087] The standard weight calculation module obtains the computing power Sl of each heterogeneous computing power. i Bandwidth Dk i and calculate density Sm i Calculate the computing power evaluation coefficient Spg for heterogeneous computing power i Furthermore, each heterogeneous computing power performs the same type of power management task, and monitoring tools are used to record the processing time (Cs) of the same power management task on different heterogeneous computing power systems. ij and energy consumption Nu ij Calculate the computational weight α of each heterogeneous computing power for different power management tasks. ij Based on the computing power evaluation coefficient Spg of heterogeneous computing power i And the computational weight α of each heterogeneous computing power for different power management tasks ij Calculate the conventional weight μ for each type of heterogeneous computing power for different power management tasks. ij .

[0088] The computing power allocation module assigns the power management tasks to be allocated to the regular weights μ. ij The maximum available heterogeneous computing power; if no available heterogeneous computing power is available, then obtain the waiting time (Da) for each heterogeneous computing power. i Based on the real-time requirement level Ja of the power management tasks to be assigned, calculate the real-time weight ρ of each heterogeneous computing power for the power management tasks to be assigned. i Based on the real-time weight ρ of each heterogeneous computing power to be allocated power management tasks i And the regular weight μ for each heterogeneous computing power to be allocated power management tasks i Calculate the comprehensive weight x for each type of heterogeneous computing power to be allocated power management tasks. i The pending power management tasks will be assigned to a comprehensive weight x. i The largest heterogeneous computing power.

[0089] The heterogeneous computing power optimization module obtains the average waiting time (De) for each power management task from the operation records of heterogeneous computing power. j Processing time Ci j Energy consumption Xhj and total quantity per unit time Sg j Calculate the busy factor Fm for each type of power management task. j And calculate the busy factor Fm between each type of power management task. j The maximum difference is denoted as the busy coefficient difference Cz. When the busy coefficient difference exceeds 2ΔCz, a heterogeneous computing power optimization instruction is issued.

[0090] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0091] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0092] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A power management method based on heterogeneous computing power scheduling, characterized in that: Comprising the following steps: acquire the computing power of each heterogeneous computing power , bandwidth and computing density , calculate the computing power evaluation coefficient of the computing power of the computing power , and make each heterogeneous computing power perform the same kind of power management task, record the processing time of the same power management task on different heterogeneous computing powers using a monitoring tool and energy consumption , calculate the computing weight of each heterogeneous computing power for different power management tasks , according to the computing power evaluation coefficient of the heterogeneous computing power and the computing weight of each heterogeneous computing power for different power management tasks , calculate the regular weight of each heterogeneous computing power for different power management tasks ; allocating the power management task to be distributed to the regular weight the largest idle heterogeneous computing power, if there is no idle heterogeneous computing power, obtaining the waiting time length of each heterogeneous computing power and the real-time requirement level Ja of the power management task to be distributed, calculating the real-time weight of each heterogeneous computing power for the power management task to be distributed , according to the real-time weight of each heterogeneous computing power for the power management task to be distributed and the regular weight of each heterogeneous computing power for the power management task to be distributed , calculating the comprehensive weight of each heterogeneous computing power for the power management task to be distributed allocating the power management task to be distributed to the comprehensive weight the largest heterogeneous computing power; The average waiting time for each power management task was obtained from the operation records of heterogeneous computing power. Processing time Energy consumption Total quantity per unit time Calculate the busy factor for each type of power management task. And calculate the busy factor between each type of power management task. The maximum difference is denoted as the busy coefficient difference Cz. When the busy coefficient difference exceeds 2ΔCz, a heterogeneous computing power optimization instruction is issued. Wherein, ΔCz represents the mean value of the busy coefficient difference of all heterogeneous computing systems.

2. The power management method based on heterogeneous computing scheduling according to claim 1, characterized in that: Obtain the computing power of each heterogeneous computing power , bandwidth and computing density , the computing power evaluation coefficient of computing heterogeneous computing power : Wherein i represents the sequential number of each isomerism computing power, i = 1, 2, …, n, n represents the total number of all isomerism computing power; The maximum computing power of all isomerism computing power, The minimum computing power of all isomerism computing power, The maximum bandwidth of all isomerism computing power, The minimum bandwidth of all isomerism computing power, The maximum computing density of all isomerism computing power, The minimum computing density of all isomerism computing power.

3. The power management method based on heterogeneous computing scheduling according to claim 1, characterized in that: causing each heterogeneous computing power to perform the same kind of power management task, recording processing time length of the same power management task at different heterogeneous computing powers using a monitoring tool and energy consumption , calculating computing weight of each heterogeneous computing power on different power management tasks : Wherein, j represents the sequential number of each power management task, j = 1, 2, …, m, m represents the total number of all power management tasks.

4. The power management method based on heterogeneous computing scheduling according to claim 1, characterized in that: Obtaining computing power evaluation coefficients of heterogeneous computing power and the calculation weight of each heterogeneous computing power for different power management tasks , calculating the regular weight of each heterogeneous computing power for different power management tasks :

5. The power management method based on heterogeneous computing scheduling according to claim 1, characterized in that: Assigning the power management task to be allocated to a regular weight The maximum idle heterogeneous computing power, and record the real-time processing time length in real time, record the difference between the current heterogeneous computing power processing the same kind of power management task processing time length and the real-time processing time length as the current heterogeneous computing power waiting time length .

6. The power management method based on heterogeneous computing scheduling according to claim 1, characterized in that: If there is no idle heterogeneous computing power, obtain the waiting time length of each heterogeneous computing power and the real-time requirement level Ja of the power management task to be allocated, calculate the real-time weight of each heterogeneous computing power on the power management task to be allocated : The real-time requirement level of the power management task is divided according to the type of the power management task, and the specific division is as follows: The planning task is low real-time requirement, and the real-time requirement level is recorded as 1; the operation monitoring is medium real-time requirement, and the real-time requirement level is recorded as 2; the real-time scheduling is high real-time requirement, and the real-time requirement level is recorded as 3; and the fault handling is extremely high real-time requirement, and the real-time requirement level is recorded as 4.

7. The power management method based on heterogeneous computing scheduling according to claim 1, characterized in that: obtaining a real-time weight of each heterogeneous computing power for the power management task to be allocated and a regular weight of each heterogeneous computing power for the power management task to be allocated , calculating a comprehensive weight of each heterogeneous computing power for the power management task to be allocated : when there is no idle heterogeneous computing power, the to-be-assigned power management task is assigned to the comprehensive weight the maximum heterogeneous computing power, and the current heterogeneous computing power waiting time is updated, i.e., the original heterogeneous computing power waiting time + the processing time of the to-be-assigned power management task of the same type of power management task 8. The power management method based on heterogeneous computing scheduling according to claim 1, characterized in that: from the run records of the heterogeneous computing power to obtain the average waiting time of each power management task , processing time , energy consumption and total number per unit time , calculate the busy factor of each power management task : wherein, represents a maximum value of the average waiting time length of all power management tasks, represents a maximum value of the processing time length of all power management tasks, represents a maximum value of the energy consumption of all power management tasks, represents a maximum value of the total number per unit time of all power management tasks.

9. A power management device based on heterogeneous computing power scheduling for implementing the method of any one of claims 1 to 8, characterized in that: Comprising: A conventional weight calculation module obtains the computing power of each heterogeneous computing power , bandwidth and computing density , calculates the computing power evaluation coefficient of the heterogeneous computing power , and makes each heterogeneous computing power perform the same kind of power management task, uses a monitoring tool to record the processing time of the same power management task on different heterogeneous computing powers and energy consumption , calculates the computing weight of each heterogeneous computing power for different power management tasks , according to the computing power evaluation coefficient of the heterogeneous computing power and the computing weight of each heterogeneous computing power for different power management tasks , calculates the conventional weight of each heterogeneous computing power for different power management tasks ; The computing power distribution module distributes the power management task to be distributed to the conventional weight The maximum idle heterogeneous computing power, if there is no idle heterogeneous computing power, the waiting time of each heterogeneous computing power is obtained And the real-time requirement level Ja of the power management task to be distributed, the real-time weight of each heterogeneous computing power for the power management task to be distributed is calculated According to the real-time weight of each heterogeneous computing power for the power management task to be distributed And the conventional weight of each heterogeneous computing power for the power management task to be distributed The comprehensive weight of each heterogeneous computing power for the power management task to be distributed is calculated The power management task to be distributed is distributed to the comprehensive weight The maximum heterogeneous computing power; The heterogeneous computing power optimization module obtains the average waiting time for each power management task from the operation records of heterogeneous computing power. Processing time Energy consumption Total quantity per unit time Calculate the busy factor for each type of power management task. And calculate the busy factor between each type of power management task. The maximum difference is denoted as the busy coefficient difference Cz. When the busy coefficient difference exceeds 2ΔCz, a heterogeneous computing power optimization instruction is issued.

Images