An AI server optimization method, system, device and readable storage medium
By classifying AI servers and building a computing power-energy consumption model, and combining task prioritization and reinforcement learning, the problems of computing power decay and energy consumption increase caused by equipment aging were solved, and the long-term stable scheduling of AI servers and the improvement of cluster reliability were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI SHUNYUAN COMPUTER TECH CO LTD
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-31
AI Technical Summary
Existing AI server resource scheduling and optimization methods cannot effectively address the problems of computing power decay, energy consumption increase and stability decline caused by equipment aging. In particular, the aging effect of high-performance computing units is significant and cannot achieve long-term stable scheduling.
By acquiring basic operational data of each core computing unit and auxiliary computing component of the AI server, the data is classified and a basic computing power-energy consumption model is constructed. A computing power redundancy threshold and an energy consumption overload threshold are set. Dynamic scheduling is performed in combination with task priority and server level. Reinforcement learning is used to optimize parameters to minimize deviation and achieve adaptive adjustment.
It has achieved long-term stable scheduling of AI servers, significantly improved the reliability and resource utilization efficiency of the cluster, reduced energy consumption and extended the service life of equipment.
Smart Images
Figure CN122489274A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of AI server optimization technology, and in particular to an AI server optimization method, system, apparatus, and readable storage medium. Background Technology
[0002] With the rapid development of artificial intelligence technology, AI servers, as the core carrier of computing infrastructure, undertake large-scale training and inference tasks. AI servers typically integrate multiple core computing units such as CPUs and GPUs, as well as auxiliary components such as high-bandwidth memory. In actual long-term operation, AI server devices may experience problems such as computing power decay, increased energy consumption, and decreased stability due to factors such as electromigration, thermal stress, and silicon aging. The aging effect of high-performance computing units such as GPUs is particularly significant.
[0003] Existing AI server resource scheduling and optimization methods are mainly geared towards new devices or ideal conditions, and typically employ static load balancing strategies or dynamic scheduling algorithms based on real-time load, which cannot achieve stable scheduling. Summary of the Invention
[0004] The purpose of this invention is to provide an optimization method, system, apparatus, and readable storage medium for AI servers to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides an optimization method for an AI server, comprising the following steps: S1. Obtain basic operational data of each core computing unit and auxiliary computing component of the AI server, and classify the AI server based on the basic operational data; S2. Obtain the computing power output value and energy consumption value corresponding to different computing power loads of AI servers at different levels, construct the AI server computing power-energy consumption basic model based on the obtained data, and determine the computing power redundancy threshold and energy consumption overload threshold. S3. Prioritize tasks based on their computing power requirements and latency requirements, determine the optimal computing power supply required for each task based on the AI server level, and match tasks with AI servers based on priority and optimal computing power supply. S4. Set a fixed scheduling cycle, combine task priority and AI server level to implement computing power scheduling, and implement energy consumption control based on AI server level and current load; S5. Based on different AI server levels, set differentiated optimization goals, calculate the optimization deviation of each goal, and dynamically adjust the baseline model, scheduling strategy and energy consumption management parameters based on reinforcement learning with the goal of minimizing the total deviation.
[0006] Preferably, step S1 specifically includes: S11. Collect basic operational data for each core computing unit and auxiliary computing component. The core computing units include CPU and GPU, and the auxiliary computing components include memory. Basic operational data includes: component runtime. Standard computing power output value of new equipment Current actual computing power output value New equipment standard energy consumption value Current actual energy consumption value ; S12. Calculate the computing power attenuation rate based on the basic operating data. and energy consumption growth rate The formula is: ; ; S13. Based on runtime, computing power decay rate, and energy consumption increase rate, the aging level of AI server equipment is classified into new equipment, slightly aged, moderately aged, and severely aged, using the following formula: ; In the formula, , , Indicates weight, Indicates the maximum computing power. This indicates the maximum energy consumption. This indicates the maximum runtime.
[0007] Preferably, in step S2, the construction of the AI server computing power-energy consumption basic model based on the acquired data, and the determination of the computing power redundancy threshold and energy consumption overload threshold, include: A polynomial fitting algorithm is used to construct basic models of AI server computing power and energy consumption at different levels. The formula is as follows: ; In the formula, Indicates different aging levels, Indicates energy consumption value. This represents the computing power output value. , , The fitting coefficients are represented by the least squares method. Determine the computing power redundancy threshold and energy overload threshold The formula is: ; In the formula, This indicates the minimum percentage of idle computing power. This represents the maximum sustainable computing power, when the actual computing power is allocated. Less than When this occurs, a computing power redundancy alarm is triggered; ; In the formula, Indicates the rated power consumption of the device. This represents the overload tolerance coefficient.
[0008] Preferably, step S3 specifically includes: Obtain the computing power requirements for each AI task And latency requirement L, based on computing power demand. Construct a priority index with delay requirement L The formula is: ; In the formula, To meet the system's maximum computing power requirements, , These are the maximum and minimum acceptable delays, respectively. , For the weighting coefficients, satisfying ,according to The value categorizes tasks into three priorities: high, medium, and low. Based on the aging level of the AI server, the optimal computing power supply for different levels of AI servers is calculated using the following formula: ; ; In the formula, This represents the average computing power attenuation rate of the corresponding AI server level. To maximize the sustainable computing power of AI servers Indicates the first The computing power decay rate of the AI server Indicates belonging to The number of AI servers of varying levels; Construct a matching function to match tasks with AI servers; the formula is as follows: In the formula, For the task Priority index, AI server The average priority index of the currently assigned tasks will be used to prioritize the allocation of tasks to other tasks. The largest AI server, and must meet the following requirements. ≥ , Indicates task The computing power requirements, if all AI servers < If so, the task will be placed in a high-priority waiting queue.
[0009] Preferably, step S4 specifically includes: Set a scheduling cycle, and at the beginning of each cycle, collect the current load, task queue status and current actual energy consumption of all AI servers at different levels; Based on task priority The AI server is classified according to its level, employing a scheduling strategy that combines weighted round-robin and priority scheduling. The formula for the computing power allocation coefficient within a cycle is: ; In the formula, AI server The average priority index of all tasks. AI server The degree of aging indicators This represents the health status of the nth AI server. If the total number of AI servers is [number], then the number of AI servers is [number]. The obtained scheduling computing power is: ; In the formula, Total available computing power for AI servers; Calculate the difference between the current actual computing power used by each AI server and the total available computing power. If the current computing power is lower than the total available computing power, schedule according to task priority and AI server matching degree. Stop allocation when it is close to the total available computing power. After allocation, the actual energy consumption and load of the AI server are collected in real time. Combined with the AI server's energy overload threshold, the current energy consumption status is determined. Based on idle energy consumption, the operating frequency and voltage of the AI server are dynamically adjusted, using the following formula: ; In the formula, Indicates the energy consumption control factor. For the first Energy overload threshold for an AI server This represents the current actual energy consumption. Idle energy consumption, if If <0, then a frequency reduction operation is triggered; if If the value is greater than 0.9, then a frequency increase is allowed, and the actual frequency adjustment amount is: ; In the formula, This is the preset frequency step size.
[0010] Preferably, step S5 specifically includes: Based on the aging level of the AI server, set targets for computing power utilization, energy efficiency, and task latency. Among these, the AI server with a higher aging level has a higher weight for the energy efficiency target. At the end of each scheduling cycle, the deviation between the actual value and the target value of each objective is calculated; The Q-learning framework is adopted to update the baseline model coefficients, computing power redundancy threshold coefficients, energy consumption overload tolerance coefficients, and scheduling weights with the goal of minimizing the total deviation.
[0011] The preferred formula for total deviation is: ; In the formula, , , This represents the weighting coefficient for each deviation. This indicates the deviation in computing power utilization. Indicates the deviation in energy efficiency. This indicates the deviation in the rate of delayed compliance.
[0012] An optimization system for AI servers is also provided, including: The data acquisition module obtains basic operational data from each core computing unit and auxiliary computing component of the AI server; The aging classification module is used to calculate the computing power decay rate and energy consumption increase rate of each AI server based on the data provided by the data acquisition module, and to calculate the comprehensive aging index. Based on the comprehensive aging index, the AI servers are divided into four levels: new equipment, mild aging, moderate aging, and severe aging. The computing power-energy consumption modeling module is used to obtain the computing power output value and energy consumption value corresponding to different computing power loads of AI servers at different levels, construct the basic model of AI server computing power-energy consumption based on the obtained data, and determine the computing power redundancy threshold and energy consumption overload threshold. The task matching module is used to prioritize tasks based on their computing power requirements and latency requirements, determine the optimal computing power supply required for each task based on the AI server level, and match tasks with AI servers based on priority and optimal computing power supply. The periodic scheduling module is used to schedule computing power according to task priority and AI server level within a fixed scheduling period, and to manage energy consumption according to AI server level and current load. The reinforcement learning optimization module is used to set differentiated optimization targets based on the aging level of each AI server, calculate the optimization deviation of each target, and dynamically adjust the baseline model, scheduling strategy and energy consumption management parameters based on reinforcement learning with the goal of minimizing the total deviation.
[0013] An apparatus is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes an optimized method for an AI server when it invokes the computer program in the memory.
[0014] A readable storage medium is also provided, which stores computer-executable instructions. When the computer-executable instructions are executed, an optimization method for an AI server is implemented.
[0015] Therefore, the present invention adopts the above-mentioned optimization method, system, device and readable storage medium for an AI server. By introducing aging classification and combining periodic scheduling and reinforcement learning to adaptively adjust parameters, long-term stable scheduling is achieved, which significantly improves the reliability of the cluster.
[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating an embodiment of an optimization method for an AI server according to the present invention. Figure 2 This is a structural framework diagram of an embodiment of an optimized AI server system according to the present invention. Detailed Implementation
[0018] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0020] Example like Figure 1 As shown, this invention provides an optimization method for an AI server, comprising the following steps: S1. Obtain basic operational data for each core computing unit and auxiliary computing component of the AI server, and classify the AI server based on this basic operational data. Specifically, this includes: S11. Periodically collect basic operational data for each core computing unit and auxiliary computing component. Core computing units include CPUs and GPUs, while auxiliary computing components include memory. Basic operational data includes: component runtime. Standard computing power output value of new equipment Current actual computing power output value New equipment standard energy consumption value Current actual energy consumption value .
[0021] S12. Calculate the computing power attenuation rate based on the basic operating data. and energy consumption growth rate The formula is: ; .
[0022] S13. Based on runtime, computing power decay rate, and energy consumption increase rate, the aging level of AI server equipment is classified into new equipment, slightly aged, moderately aged, and severely aged, using the following formula: ; In the formula, , , Indicates weight, Indicates the maximum computing power. This indicates the maximum energy consumption. This represents the maximum runtime. The specific division method is as follows: When γ≤0.2, it indicates new equipment; when 0.2<γ≤0.5, it indicates mild aging; when 0.5<γ≤0.8, it indicates moderate aging; and when γ>0.8, it indicates severe aging.
[0023] S2. Obtain the computing power output and energy consumption values corresponding to different computing power loads of AI servers at different levels. Based on the obtained data, construct a basic AI server computing power-energy consumption model and determine the computing power redundancy threshold and energy consumption overload threshold. Specifically, this includes: For each aging level k, k=1,2,3,4, within the low-load to high-load range, the computing power load is gradually increased, and the corresponding computing power output values are recorded. The energy consumption value E was measured three times at each load point and the average value was taken.
[0024] A polynomial fitting algorithm is used to construct basic models of AI server computing power and energy consumption at different levels. The formula is as follows: ; In the formula, Indicates different aging levels, Indicates energy consumption value. This represents the computing power output value. , , The coefficients are represented by , which are solved using the least squares method.
[0025] Determine the computing power redundancy threshold and energy overload threshold The formula is: ; In the formula, This indicates the minimum percentage of idle computing power. This represents the maximum sustainable computing power, when the actual computing power is allocated. Less than When this occurs, a computing power redundancy alarm is triggered; ; In the formula, Indicates the rated power consumption of the device. This represents the overload tolerance coefficient.
[0026] S3. Prioritize tasks based on their computing power and latency requirements, determine the optimal computing power supply for each task based on the AI server's level, and match tasks with AI servers according to priority and optimal computing power supply. Specifically, this includes: Obtain the computing power requirements for each AI task And latency requirement L, based on computing power demand. Construct a priority index with delay requirement L The formula is: ; In the formula, To meet the system's maximum computing power requirements, , These are the maximum and minimum acceptable delays, respectively. , For the weighting coefficients, satisfying ,according to The value categorizes tasks into three priorities: high, medium, and low. When When the value is ≥0.7, it has high priority; when 0.3 ≤ When <0.7, it is a medium priority; when When the value is less than 0.3, it is considered a low priority.
[0027] Based on the aging level of the AI server, the optimal computing power supply for different levels of AI servers is calculated using the following formula: ; ; In the formula, This represents the average computing power attenuation rate of the corresponding AI server level. To maximize the sustainable computing power of AI servers Indicates the first The computing power decay rate of the AI server Indicates belonging to The number of AI servers of varying levels; Construct a matching function to match tasks with AI servers; the formula is as follows: In the formula, For the task Priority index, AI server The average priority index of the currently assigned tasks will be used to prioritize the allocation of tasks to other tasks. The largest AI server, and must meet the following requirements. ≥ , Indicates task The computing power requirements, if all AI servers < If so, the task will be placed in a high-priority waiting queue.
[0028] S4. Set a fixed scheduling cycle, implement computing power scheduling based on task priority and AI server level, and implement energy consumption management according to AI server level and current load. Specifically, this includes: Set a scheduling cycle, such as 5 seconds, and at the beginning of each cycle, collect the current load, task queue status and current actual energy consumption of all AI servers at different levels; Based on task priority The AI server is classified according to its level, employing a scheduling strategy that combines weighted round-robin and priority scheduling. The formula for the computing power allocation coefficient within a cycle is: ; In the formula, AI server The average priority index of all tasks. AI server The degree of aging indicators This represents the health status of the nth AI server. If the total number of AI servers is [number], then the number of AI servers is [number]. The obtained scheduling computing power is: ; In the formula, Total available computing power for AI servers; Calculate the difference between the current actual computing power used by each AI server and the total available computing power. If the current computing power is lower than the total available computing power, scheduling is performed according to task priority and AI server matching degree. When it approaches the total available computing power, allocation stops. If the current computing power is higher than the total available computing power, the lowest priority task on that AI server is migrated to other AI servers with spare computing power, or the task is put back into the waiting queue until the current computing power drops below the total available computing power. If all AI servers are unable to satisfy a high-priority task, the task remains in the waiting queue, and a resource alarm is triggered.
[0029] After allocation, the actual energy consumption and load of the AI server are collected in real time. Combined with the AI server's energy overload threshold, the current energy consumption status is determined. Based on idle energy consumption, the operating frequency and voltage of the AI server are dynamically adjusted, using the following formula: ; In the formula, Indicates the energy consumption control factor. For the first Energy overload threshold for an AI server This represents the current actual energy consumption. Idle energy consumption, if If <0, then a frequency reduction operation is triggered; if If the value is greater than 0.9, then a frequency increase is allowed, and the actual frequency adjustment amount is: ; In the formula, This is the preset frequency step size.
[0030] If the current energy consumption is close to or exceeds the overload threshold, it indicates an overload risk, and the system needs to reduce the frequency and voltage to decrease power consumption. If the current energy consumption is far below the overload threshold and the AI server load is high, it indicates room for performance improvement, and the system can appropriately increase the frequency to speed up task processing. If the current energy consumption is in the intermediate range, the current frequency should be kept unchanged.
[0031] S5. Based on different AI server levels, set differentiated optimization objectives, calculate the optimization deviation of each objective, and dynamically adjust the baseline model, scheduling strategy, and energy consumption management parameters based on reinforcement learning, with the goal of minimizing the total deviation. Specifically, this includes: Based on the aging level of the AI server, set targets for computing power utilization, energy efficiency, and task latency. Among these, the AI server with a higher aging level has a higher weight for the energy efficiency target. At the end of each scheduling cycle, the deviation between the actual value and the target value of each objective is calculated; Using a Q-learning framework, with the goal of minimizing the total deviation, the baseline model coefficients, computing power redundancy threshold coefficients, energy consumption overload tolerance coefficients, and scheduling weights are updated. The formula for the total deviation is: ; In the formula, , , This represents the weighting coefficient for each deviation. This indicates the deviation in computing power utilization. Indicates the deviation in energy efficiency. This indicates the deviation in the rate of delayed compliance.
[0032] Reinforcement learning is set as follows: Status s: The current status of AI servers at each level. , , ).
[0033] Action a: Adjustment amount of adjustable parameters, such as baseline model coefficients, computing power redundancy threshold coefficients, energy consumption overload tolerance coefficients, and scheduling weights; Reward r: .
[0034] Q value update: ; In the formula, For learning rate, This is the discount factor. Through continuous training, the action that maximizes the Q-value is selected in each state, thereby minimizing the total bias. The adjusted parameters are updated in real time in the model of steps S2-S4, forming a closed-loop optimization.
[0035] like Figure 2 As shown, the present invention also provides an optimization system for an AI server, comprising: The data acquisition module obtains basic operational data from each core computing unit and auxiliary computing component of the AI server; The aging classification module is used to calculate the computing power decay rate and energy consumption increase rate of each AI server based on the data provided by the data acquisition module, and to calculate the comprehensive aging index. Based on the comprehensive aging index, the AI servers are divided into four levels: new equipment, mild aging, moderate aging, and severe aging. The computing power-energy consumption modeling module is used to obtain the computing power output value and energy consumption value corresponding to different computing power loads of AI servers at different levels, construct the basic model of AI server computing power-energy consumption based on the obtained data, and determine the computing power redundancy threshold and energy consumption overload threshold. The task matching module is used to prioritize tasks based on their computing power requirements and latency requirements, determine the optimal computing power supply required for each task based on the AI server level, and match tasks with AI servers based on priority and optimal computing power supply. The periodic scheduling module is used to schedule computing power according to task priority and AI server level within a fixed scheduling period, and to manage energy consumption according to AI server level and current load. The reinforcement learning optimization module is used to set differentiated optimization targets based on the aging level of each AI server, calculate the optimization deviation of each target, and dynamically adjust the baseline model, scheduling strategy and energy consumption management parameters based on reinforcement learning with the goal of minimizing the total deviation.
[0036] The present invention also provides an apparatus including a memory and a processor, wherein the memory stores a computer program, and the processor executes the above-described optimization method for an AI server when it calls the computer program in the memory.
[0037] The present invention also provides a readable storage medium storing computer-executable instructions, which, when executed, implement the aforementioned optimization method for an AI server.
[0038] Therefore, the present invention adopts the above-mentioned optimization method, system, device and readable storage medium for an AI server. By introducing aging classification and combining periodic scheduling and reinforcement learning to adaptively adjust parameters, long-term stable scheduling is achieved, which significantly improves the reliability of the cluster.
[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. An optimization method for an AI server, characterized in that, Includes the following steps: S1. Obtain basic operational data of each core computing unit and auxiliary computing component of the AI server, and classify the AI server based on the basic operational data; S2. Obtain the computing power output value and energy consumption value corresponding to different computing power loads of AI servers at different levels, construct the AI server computing power-energy consumption basic model based on the obtained data, and determine the computing power redundancy threshold and energy consumption overload threshold. S3. Prioritize tasks based on their computing power requirements and latency requirements, determine the optimal computing power supply required for each task based on the AI server level, and match tasks with AI servers based on priority and optimal computing power supply. S4. Set a fixed scheduling cycle, combine task priority and AI server level to implement computing power scheduling, and implement energy consumption control based on AI server level and current load; S5. Based on different AI server levels, set differentiated optimization goals, calculate the optimization deviation of each goal, and dynamically adjust the baseline model, scheduling strategy and energy consumption management parameters based on reinforcement learning with the goal of minimizing the total deviation.
2. The optimization method for an AI server according to claim 1, characterized in that, Step S1 specifically includes: S11. Collect basic operational data for each core computing unit and auxiliary computing component. The core computing units include CPU and GPU, and the auxiliary computing components include memory. Basic operational data includes: component runtime. Standard computing power output value of new equipment Current actual computing power output value New equipment standard energy consumption value Current actual energy consumption value ; S12. Calculate the computing power attenuation rate based on the basic operating data. and energy consumption growth rate The formula is: ； ； S13. Based on runtime, computing power decay rate, and energy consumption increase rate, the aging level of AI server equipment is classified into new equipment, slightly aged, moderately aged, and severely aged, using the following formula: ； In the formula, , , Indicates weight, Indicates the maximum computing power. This indicates the maximum energy consumption. This indicates the maximum runtime.
3. The optimization method for an AI server according to claim 2, characterized in that, In step S2, a basic AI server computing power-energy consumption model is constructed based on the acquired data, and the computing power redundancy threshold and energy consumption overload threshold are determined, including: A polynomial fitting algorithm is used to construct basic models of AI server computing power and energy consumption at different levels. The formula is as follows: ； In the formula, Indicates different aging levels, Indicates energy consumption value. This represents the computing power output value. , , The fitting coefficients are represented by the least squares method. Determine the computing power redundancy threshold and energy overload threshold The formula is: ； In the formula, This indicates the minimum percentage of idle computing power. This represents the maximum sustainable computing power, when the actual computing power is allocated. Less than When this occurs, a computing power redundancy alarm is triggered; ； In the formula, Indicates the rated power consumption of the device. This represents the overload tolerance coefficient.
4. The optimization method for an AI server according to claim 3, characterized in that, Step S3 specifically includes: Obtain the computing power requirements for each AI task And latency requirement L, based on computing power demand. Construct a priority index with delay requirement L The formula is: ； In the formula, To meet the system's maximum computing power requirements, , These are the maximum and minimum acceptable delays, respectively. , Let be the weighting coefficient, satisfying ,according to The value categorizes tasks into three priorities: high, medium, and low. Based on the aging level of the AI server, the optimal computing power supply for different levels of AI servers is calculated using the following formula: ； ； In the formula, This represents the average computing power attenuation rate of the corresponding AI server level. To maximize the sustainable computing power of AI servers Indicates the first The computing power decay rate of the AI server Indicates belonging to The number of AI servers of varying levels; Construct a matching function to match tasks with AI servers; the formula is as follows: In the formula, For the task Priority index, AI server The average priority index of the currently assigned tasks will be used to prioritize the allocation of tasks to other tasks. The largest AI server, and must meet the following requirements. ≥ , Indicates task The computing power requirements, if all AI servers < If so, the task will be placed in a high-priority waiting queue.
5. The optimization method for an AI server according to claim 4, characterized in that, Step S4 specifically includes: Set a scheduling cycle, and at the beginning of each cycle, collect the current load, task queue status and current actual energy consumption of all AI servers at different levels; Based on task priority The AI server is classified according to its level, employing a scheduling strategy that combines weighted round-robin and priority scheduling. The formula for the computing power allocation coefficient within a cycle is: ； In the formula, AI server The average priority index of all tasks. AI server The degree of aging indicators This represents the health status of the nth AI server. If the total number of AI servers is [number], then the number of AI servers is [number]. The obtained scheduling computing power is: ； In the formula, Total available computing power for AI servers; Calculate the difference between the current actual computing power used by each AI server and the total available computing power. If the current computing power is lower than the total available computing power, schedule according to task priority and AI server matching degree. Stop allocation when it is close to the total available computing power. After allocation, the actual energy consumption and load of the AI server are collected in real time. Combined with the AI server's energy overload threshold, the current energy consumption status is determined. Based on idle energy consumption, the operating frequency and voltage of the AI server are dynamically adjusted, using the following formula: ； In the formula, Indicates the energy consumption control factor. For the first Energy overload threshold for an AI server This represents the current actual energy consumption. Idle energy consumption, if If <0, then a frequency reduction operation is triggered; if If the value is greater than 0.9, then a frequency increase is allowed, and the actual frequency adjustment amount is: ； In the formula, This is the preset frequency step size.
6. The optimization method for an AI server according to claim 5, characterized in that, Step S5 specifically includes: Based on the aging level of the AI server, set targets for computing power utilization, energy efficiency, and task latency. Among these, the AI server with a higher aging level has a higher weight for the energy efficiency target. At the end of each scheduling cycle, the deviation between the actual value and the target value of each objective is calculated; The Q-learning framework is adopted to update the baseline model coefficients, computing power redundancy threshold coefficients, energy consumption overload tolerance coefficients, and scheduling weights with the goal of minimizing the total deviation.
7. The optimization method for an AI server according to claim 6, characterized in that, The total deviation formula is: ； In the formula, , , This represents the weighting coefficient for each deviation. This indicates the deviation in computing power utilization. Indicates the deviation in energy efficiency. This indicates the deviation in the rate of delayed compliance.
8. An optimization system for an AI server, comprising executing an optimization method for an AI server as described in any one of claims 1-7, characterized in that, include: The data acquisition module obtains basic operational data from each core computing unit and auxiliary computing component of the AI server; The aging classification module is used to calculate the computing power decay rate and energy consumption increase rate of each AI server based on the data provided by the data acquisition module, and to calculate the comprehensive aging index. Based on the comprehensive aging index, the AI servers are divided into four levels: new equipment, mild aging, moderate aging, and severe aging. The computing power-energy consumption modeling module is used to obtain the computing power output value and energy consumption value corresponding to different computing power loads of AI servers at different levels, construct the basic model of AI server computing power-energy consumption based on the obtained data, and determine the computing power redundancy threshold and energy consumption overload threshold. The task matching module is used to prioritize tasks based on their computing power requirements and latency requirements, determine the optimal computing power supply required for each task based on the AI server level, and match tasks with AI servers based on priority and optimal computing power supply. The periodic scheduling module is used to schedule computing power according to task priority and AI server level within a fixed scheduling period, and to manage energy consumption according to AI server level and current load. The reinforcement learning optimization module is used to set differentiated optimization targets based on the aging level of each AI server, calculate the optimization deviation of each target, and dynamically adjust the baseline model, scheduling strategy and energy consumption management parameters based on reinforcement learning with the goal of minimizing the total deviation.
9. An apparatus, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes an optimization method for an AI server as described in any one of claims 1 to 7 when it invokes the computer program in the memory.
10. A readable storage medium, characterized in that, The readable storage medium stores computer-executable instructions, which, when executed, implement an optimization method for an AI server as described in any one of claims 1 to 7.