Computing power center load optimization control method based on multi-agent distributed control

Abstract

The invention discloses a computing power center load optimization control method based on multi-agent distributed control, and the method comprises the steps: 1, building physical parameter models of all equipment of an actual computing power center, including building a processor load model, a storage equipment load model, a switch equipment load model and a refrigeration equipment load model, simulating the load condition of the computing power center; 2, establishing a power constraint, a calorific value constraint and a computing power constraint; step 3, based on the physical parameter model of each device, allocating a corresponding agent to each device to form a multi-agent system, and establishing an agent model; and 4, under the constraint condition, realizing real-time optimization and distribution of each equipment load by adopting a distributed algorithm so as to realize energy consumption optimization.

Description

Technical Field

[0001] The present invention belongs to the field of information technology and relates to a computing center load optimization control method based on multi-agent distributed control. Background Art

[0002] With the rapid popularization of technologies such as generative AI and cloud servers, society's demand for data information processing, storage, transmission, exchange and management has increased dramatically. In this context, the number and scale of computing centers have continued to expand. However, computing centers consume a lot of energy. Since the current power system mainly generates electricity by burning fossil fuels such as natural gas and coal, the huge power consumption of computing centers will inevitably lead to a large amount of carbon emissions. Therefore, computing centers are not only the pillar of future economic and social development, but also the most critical link in energy conservation and consumption reduction of new infrastructure.

[0003] Energy consumption optimization of computing centers is an inevitable requirement for reducing the operating costs of computing centers and their carbon emissions. Existing research on energy consumption optimization of computing centers can be roughly divided into three categories: the first category is to optimize the energy supply of computing centers, mainly focusing on exploring how to integrate renewable energy into the energy supply system of computing centers; the second category is to manage the workload of computing centers, thereby changing the load of computing centers to match the output of renewable energy. The third category is to manage the energy consumption of computing centers, using prediction technology or uncertainty processing technology to deal with uncertainties in the operation of computing centers, optimize the scheduling of various equipment in computing centers and power transactions with large power grids, so as to achieve the optimization of energy consumption, including minimizing the power purchase expenditure of the power grid, minimizing operating costs, maximizing the utilization rate of renewable energy and other goals.

[0004] In the current computing center load control problem, commonly used algorithms include genetic algorithm (GA), particle swarm optimization (PSO), ant colony algorithm (ACO) and reinforcement learning (RL). These algorithms are all aimed at optimizing the power consumption, load balancing, cooling efficiency and other aspects of the computing center. However, these algorithms have the problem of slow convergence or easy to fall into the local optimal solution. In the computing center, the energy consumption of computing equipment (CPU / GPU), storage equipment, air conditioning equipment and network equipment affects each other and has strong instability. Different computing equipment performs different computing tasks, and the computing power consumption, cooling capacity and upstream and downstream data volume are all different. Using the same control algorithm will inevitably lead to energy loss and waste. Summary of the invention

[0005] In order to solve the above technical problems, the purpose of the present invention is to provide a computing center load optimization control method based on multi-agent distributed control, so that each group of equipment can make real-time load control strategies for different computing tasks in a collaborative manner, thereby improving the overall energy efficiency of the computing center.

[0006] The present invention provides a computing center load optimization control method based on multi-agent distributed control, comprising:

[0007] Step 1: Establish the physical parameter model of each device in the actual computing center, including the processor load model, storage device load model, switch device load model and refrigeration device load model, to simulate the load of the computing center;

[0008] Step 2: Establish power constraints, heat constraints, and computing power constraints;

[0009] Step 3: Based on the physical parameter models of each device, assign a corresponding agent to each device to form a multi-agent system and establish an agent model;

[0010] Step 4: Under the constraints, a distributed algorithm is used to achieve real-time optimization and allocation of the load of each device to achieve optimal energy consumption.

[0011] The present invention aims at the problem that various devices in the computing center cannot operate in the optimal energy consumption state, causing unnecessary power loss, and proposes a computing center load optimization control method based on multi-agent distributed control, which can make computing devices, storage devices, switch devices and refrigeration equipment in a collaborative way, according to the working state and heating conditions of different devices, and make real-time power consumption adjustment strategies for different devices, thereby improving the energy utilization efficiency of the computing center as a whole; and the problem that traditional genetic algorithms, particle swarm algorithms and other algorithms converge slowly or easily fall into local optimal solutions. In the data center system, the present invention uses multiple independent agents to work together to complete the task of optimal working conditions. The most prominent feature is that through local autonomy and collaboration between agents, efficient decision-making and system optimization can be achieved in a complex and dynamic environment. Its advantages are high scalability and robustness, and it can flexibly respond to changes such as load fluctuations and equipment failures in large-scale computing centers, while improving energy efficiency, reducing the impact of system failures, and ensuring stable operation. The distributed algorithm effectively reduces the demand for communication bandwidth, effectively improves the utilization rate of communication equipment and the requirements for equipment performance, and reduces the deployment cost of control equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a flow chart of the computing center load optimization control method based on multi-agent distributed control of the present invention;

[0013] Figure 2 It is a distributed control structure diagram of the present invention. DETAILED DESCRIPTION

[0014] like Figure 1As shown, the computing center load optimization control method based on multi-agent distributed control of the present invention includes:

[0015] Step 1: Establish the physical parameter model of each device in the actual computing center, including the processor load model, storage device load model, switch device load model and refrigeration device load model, to simulate the load of the computing center, specifically:

[0016] Step 1.1: Create the following processor load model:

[0017] P C =p C (f) = k 1 ×C×V 2 ×f+p 0

[0018] Q C =q C (P C ) = k′×C×V 2 ×f+p 0

[0019] CP=cp(f)=M×f×ECP

[0020] Among them, P C is the processor power consumption, which is a function of the CPU frequency f C (f); Q C The heat generated by the processor is about the CPU power consumption P C The function q C (P C );CP is the processor computing power, which is a function of CPU frequency f cp(f); M is the number of processor cores, f is the processor single-core frequency, ECP is the CPU single-core computing power parameter, which is related to the processor process and architecture; k 1 is the power consumption ratio coefficient of the processor, which needs to be obtained through actual testing; C is the processor capacitance load, which depends on the chip design and process; V is the processor operating voltage, p 0 is the processor static power consumption, and different models of processors have different static power consumption; k′ is a 1 The close coefficient means that the processor's electrical energy is mainly consumed in transistor switches, and the switching energy and leakage current energy are eventually dissipated in the form of heat energy, resulting in the heat generation being close to the processor's power consumption.

[0021] Step 1.2: Create the following storage device load model:

[0022] P H =p H (cp) = P 0 +α 1 ×cp

[0023] Q H =q H (p H ) = k 2 ×p H

[0024] Among them, P H is the power consumption of the storage device, through the function p of the real-time computing power cp H (cp) definition; Q H The heat generated by the storage device is expressed as a function of power consumption q H (P H ) definition; 0 Indicates the normal power consumption of the solid state drive, α 1 is the power consumption ratio coefficient of the storage device. After testing, the heat generation of the solid-state hard disk is linearly related to the data reading and writing speed; the power consumption of the solid-state hard disk is almost completely converted into heat. The heat generation ratio coefficient k of the storage device is used. 2 Make an approximation.

[0025] Step 1.3: Create the following switch equipment load model:

[0026] P S =P static +P dynamic

[0027] P dynamic =P port +P processing =α 2 ×cp+β 2 ×cp

[0028] Q S =k 3 ×P S

[0029] Among them, P S Represents the power consumption of each switch device, which is composed of static power consumption P static and dynamic power consumption P dynamic ;P port Indicates the dynamic power consumption of the port, P processing Represents the power consumption of data parsing and encryption inside the switch; they are fitted into functions of data throughput cp, respectively, and the dynamic power consumption ratio coefficient α of the switch port 2 The power consumption ratio coefficient β of the switch's internal data parsing and encryption 2 OK; Q S is the heat generated by the switch, k 3 is the heat generation ratio coefficient of the switch.

[0030] Step 1.4: Establish the following refrigeration equipment load model:

[0031] Q A =q A (P A )=P A ×(1+α 3 ×T β )

[0032] P A =p A (Q C +Q H +Q S +Q A )=-k 4 (Q C +Q H +Q S +Q A )

[0033] Among them, Q A is the heat generated by each refrigeration device, defined as a function of the refrigeration device power P A The function q A (P A );and the refrigeration equipment power P A The size of is related to the heat generated by the device it regulates, and is represented by a function p A (·) definition; α 3 is a coefficient related to air conditioning design and internal heat loss, usually a positive value; T is the temperature of the air conditioning external environment, β is an index of the effect of ambient temperature on air conditioning heat loss, usually a positive number less than , used to indicate the effect of temperature on heat dissipation; k 4 is the proportionality coefficient, which needs to be fitted and calculated based on the on-site equipment.

[0034] Step 1.5: Each device is deployed with a temperature sensor and a communication control device. Each device communicates with the other in real time about its power status. The computing power of the processor is communicated to other devices in one direction through the processor, and the heat is communicated to the adjacent cooling device in one direction through other devices.

[0035] Step 2: Establish power constraints, heat generation constraints, and computing power constraints, specifically:

[0036] P CMin ≤P C ≤P CMax ,Q C ≤Q CMax

[0037] P HMin ≤P H ≤P HMax ,Q H ≤Q HMax

[0038] PSMin ≤P S ≤P SMax ,Q S ≤Q SMax

[0039] P AMin ≤P A ≤P AMax ,Q A ≤Q AMax

[0040] Q C +Q H +Q S +Q A ≤Q Limit

[0041] f=p C -1 (P C )

[0042] ∑cp(f)≥cp Min

[0043] Among them, P CMin and P CMax is the lower and upper limits of the processor power consumption, Q CMax is the upper limit of the processor's heat generation; P HMin and P HMax is the lower and upper limits of the power consumption of the storage device, Q HMax is the upper limit of the heat generated by the storage device; P SMin and P SMax is the lower and upper limits of the switch power consumption, Q SMax P is the upper limit of the heat generated by the switch; AMin and P AMax is the lower and upper limits of the power of the refrigeration equipment; Q AMax Upper limit of heat generation of refrigeration equipment, Q Limit The maximum cooling capacity of the refrigeration equipment; cp Min The minimum computing power required to handle the current working conditions.

[0044] Step 3: Based on the physical parameter models of each device, assign a corresponding agent to each device to form a multi-agent system and establish an agent model. Specifically:

[0045] Step 3.1: Assign a corresponding agent to each device:

[0046] Processor Load Agent:

[0047] C=[C 1 ,C 2 ,…,C n1 ]

[0048] C k = {P Ck =p C (f),Q Ck =q C (P Ck ),CP k =cp(f)}

[0049] Where n1 is the number of processors, C k represents the kth processor, P Ck is the power consumption of the kth processor, Q Ck is the heat generated by the kth processor, CP k is the computing power of the kth processor.

[0050] Storage device load agent:

[0051] H=[H 1 ,H 2 ,…,H n2 ]

[0052] H o = {P Ho =p H (cp),Q Ho =q H (P Ho )}

[0053] Among them, n2 is the number of storage devices, H o Indicates the working status of the oth storage device, including the power consumption P of the storage device Ho and storage device heat Q Ho .

[0054] Switch equipment load agent:

[0055] S=[S 1 ,S 2 ,…,S n3 ]

[0056] S w = {P Sw =p S (cp),Q Sw =q S (P Sw )}

[0057] Among them, n3 is the number of switch devices, S w represents the working status of the wth switch, P Sw represents the power consumption of the wth switch device; Q Sw is the heat generated by the wth switch.

[0058] Refrigeration Equipment Load Agent:

[0059] A=[A 1 ,A 2 ,…,A n4 ]

[0060]

[0061] Among them, n4 is the number of refrigeration equipment, A v represents the working state of the vth refrigeration equipment, Q Av is the calorific value of the vth refrigeration equipment; and the refrigeration equipment power P A The size of is related to the heat generated by the device it regulates, and is represented by a function p A (·)definition, is the maximum cooling capacity of the vth refrigeration equipment.

[0062] Step 3.2: Abstract the above device into a multi-agent system consisting of four agents. The communication graph between them is G = (V, E, A), where V = I [1, N] represents the node set, E ∈ I [1, N] × I [1, N] represents the edge set, and A = A [a ij ] N×N represents the adjacency matrix, a ij Indicates the connection state, 1 indicates connection, 0 indicates non-connection, and defines the Laplace matrix of the adjacency matrix L = L[l ij ] N×N , N=4 is the number of intelligent agents.

[0063] Step 3.3: Model the communication graph between agents as a random switching graph: G σ(t) =(V,E σ(t) ,A σ(t) ), in several different graphs G 1 ,G 2 …G s Switch between; G if and only if σ(t)=q σ(t) =G q ; The switching signal σ(t) obeys a Markov process that takes values ​​in a finite state space S = {1,…,s}, where s is the number of state spaces, and Λ = [λ mq ] s×s is the conversion rate matrix, where λ mq satisfy:

[0064]

[0065] λ mm =-Σ q≠m λ mq ,λ mq >0

[0066] Among them, o(Δt) represents the higher-order infinitesimal of time t.

[0067] Step 3.4: Create the following agent model:

[0068]

[0069] in, represents the state of the ith agent, represents the power control input of the ith agent, and They represent the system matrices respectively.

[0070] Step 4: Under the constraints, a distributed algorithm is used to achieve real-time optimization and allocation of the load of each device to achieve optimal energy consumption, specifically:

[0071] Step 4.1: Establish the following optimization objective function:

[0072]

[0073] Among them, P i (x Pi ) is the function of each agent to calculate the internal power consumption, cp(x Pi ) is the computing power function of the processor load agent, and its ratio f i (x Pi ) is a cost function known only to each agent, which is differentiable, emergent and satisfies:

[0074]

[0075] Where R i is a negative definite matrix, satisfy:

[0076]

[0077]

[0078] The CPU will cp(x Pi ) reports to each agent, the processor load agent, storage device load agent and switch device load agent report the heat to the cooling device load agent.

[0079] Step 4.2: Through synchronous communication between neighbors, design a fully distributed event-triggered control scheme for the i-th agent as follows:

[0080]

[0081]

[0082]

[0083] in, For x Pi Open-loop estimation of μ ij , ij is the adaptive gain, F is the feedback gain matrix to be specified; v ij , o ij Is a normal number.

[0084] Defining State Estimates

[0085]

[0086] Define sensor error:

[0087] The k+1th trigger time of the information transmission from the i-th agent to its neighboring agents is determined by the following trigger mechanism:

[0088]

[0089] Among them, the trigger function h i (e i ,π i ), Designed to:

[0090]

[0091]

[0092]

[0093]

[0094] Among them, δ,δ 1 ,δ 2 ,γ,γ 1 ,γ 2 ,γ 3 , are all positive numbers; π i , is the internal dynamic variable of the ith agent and is greater than zero; γ 2 ∈(0,1),γ 3 >(1-γ 2 )γ 1 ,

[0095] When the i-th agent is triggered, keep h i (e i ,πi )>0 or Reset i =0, and then the i-th agent transmits the current information to all external neighbors, thereby saving computing resources and reducing the demand for communication bandwidth.

[0096] The computing power cp only communicates outward from the processor load agent without establishing two-way communication. Similarly, the heat Q of the agent only needs to establish one-way communication with the refrigeration equipment load agent without establishing two-way communication. Of course, in order to establish a distributed structure, two-way communication about cp and Q must be established between agents that are not related to computing power or heat to ensure the stability of data transmission. Of course, the above directional description is only based on the fact that the two are neighbors, and global knowledge sharing is achieved through changes in neighbors. Thus, the following is established: Figure 2 The distributed control structure shown.

[0097] Step 4.3: By iterating and converging the power consumption reference values ​​of each intelligent agent, the power consumption requirements of each intelligent agent are inversely calculated, and the power is allocated to the response of each device so that the device can operate in the optimal energy consumption state that can meet the computing requirements.

[0098] This method adopts multi-agent distributed control technology, treating each device in the computing power center as an independent agent, allowing them to work together to optimize the overall load control strategy while ensuring their respective goals. Each group of equipment dynamically adjusts the load according to real-time task requirements, thereby effectively responding to changes in different computing power tasks and improving the system's responsiveness and adaptability. This method can not only improve the energy utilization efficiency of the computing power center and reduce energy waste, but also control heat accumulation, avoid overheating, ensure stable operation of equipment, and reduce the risk of failure. At the same time, the optimized load control strategy can effectively reduce power consumption while ensuring computing power output, thereby achieving economical operation of the computing power center.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the concept of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A computing center load optimization control method based on multi-agent distributed control, characterized in that: include: Step 1: Establish the physical parameter model of each device in the actual computing center, including the processor load model, storage device load model, switch device load model and refrigeration device load model, to simulate the load of the computing center; Step 2: Establish power constraints, heat constraints, and computing power constraints; Step 3: Based on the physical parameter models of each device, assign a corresponding agent to each device to form a multi-agent system and establish an agent model; Step 4: Under the constraints, a distributed algorithm is used to achieve real-time optimization and allocation of the load of each device to achieve optimal energy consumption.

2. The computing center load optimization control method based on multi-agent distributed control according to claim 1 is characterized in that: The step 1 is specifically as follows: Step 1.1: Create the following processor load model: P C =p C (f)=k1×C×V 2 ×f+p0 Q C =q C (P C )=k′×C×V 2 ×f+p0 CP=cp(f)=M×f×ECP Among them, P C is the processor power consumption, which is a function of the CPU frequency f C (f); Q C The heat generated by the processor is about the CPU power consumption P C The function q C (P C );CP is the processor computing power, which is a function of CPU frequency f cp(f); M is the number of processor cores, f is the processor single-core frequency, ECP is the CPU single-core computing power parameter, which is related to the processor process and architecture; k1 is the power consumption ratio coefficient of the processor, which needs to be obtained through actual testing; C is the processor capacitance load, which depends on the chip design and process; V is the processor operating voltage, p0 is the processor static power consumption, and different models of processors have different static power consumption; k′ is a coefficient close to k1. The processor's power is mainly consumed in transistor switches. The switching energy and leakage current energy are eventually dissipated in the form of heat energy, resulting in the heat generation being close to the processor's power consumption; Step 1.2: Create the following storage device load model: P H =p H (cp)=P0+α1×cp Q H =q H (p H )=k2×p H Among them, P H is the power consumption of the storage device, through the function p of the real-time computing power cp H (cp) definition; Q H The heat generated by the storage device is expressed as a function of power consumption q H (P H )definition; P0 represents the normal power consumption of the SSD, α1 is the power consumption ratio coefficient of the storage device. After testing, the heat generation of the SSD is linearly related to the data reading and writing speed; the power consumption of the SSD is almost completely converted into heat, which is approximated by the heat generation ratio coefficient k2 of the storage device; Step 1.3: Create the following switch equipment load model: P S =P static +P dynamic P dynamic =P port +P processing =α2×cp+β2×cp Q S =k3×P S Among them, P S Represents the power consumption of each switch device, which is composed of static power consumption P static and dynamic power consumption P dynamic ;P port Indicates the dynamic power consumption of the port, P processing represents the power consumption of data parsing and encryption inside the switch; they are respectively fitted into functions of data throughput cp, which are determined by the switch port dynamic power consumption ratio coefficient α2 and the switch internal data parsing and encryption power consumption ratio coefficient β2; Q S is the heat generated by the switch, and k3 is the heat generation ratio coefficient of the switch; Step 1.4: Establish the following refrigeration equipment load model: Q A =q A (P A )=P A ×(1+α3×T β ) P A =p A (Q C +Q H +Q S +Q A )=-k4(Q C +Q H +Q S +Q A ) Among them, Q A is the heat generated by each refrigeration device, defined as a function of the refrigeration device power P A The function q A (P A );and the refrigeration equipment power P A The size of is related to the heat generated by the device it regulates, and is represented by a function p A (·) Definition; α3 is a coefficient related to air conditioning design and internal heat loss, usually a positive value; T is the temperature of the air conditioning external environment, β is an index of the effect of ambient temperature on air conditioning heat loss, usually a positive number less than , used to represent the effect of temperature on heat dissipation; k4 is a proportionality coefficient, which needs to be fitted and calculated according to the on-site equipment; Step 1.5: Each device is deployed with a temperature sensor and a communication control device. Each device communicates with the other in real time about its power status. The computing power of the processor is communicated to other devices in one direction through the processor, and the heat is communicated to the adjacent cooling device in one direction through other devices.

3. The computing center load optimization control method based on multi-agent distributed control according to claim 1 is characterized in that: The step 2 establishes the following constraints: P CMin ≤P C ≤P CMax ,Q C ≤Q CMax P HMin ≤P H ≤P HMax ,Q H ≤Q HMax P SMin ≤P S ≤P SMax ,Q S ≤Q SMax P AMin ≤P A ≤P AMax ,Q A ≤Q AMax Q C +Q H +Q S +Q A ≤Q Limit ∑cp(f)≥cp Min Among them, P CMin and P CMax is the lower and upper limits of the processor power consumption, Q CMax is the upper limit of the processor's heat generation; P HMin and P HMax is the lower and upper limits of the power consumption of the storage device, Q HMax is the upper limit of the heat generated by the storage device; P SMin and P SMax is the lower and upper limits of the switch power consumption, Q SMax P is the upper limit of the heat generated by the switch; AMin and P AMax is the lower and upper limits of the power of the refrigeration equipment; Q AMax Upper limit of heat generation of refrigeration equipment, Q Limit The maximum cooling capacity of the refrigeration equipment; cp Min The minimum computing power required to handle the current working conditions.

4. The computing center load optimization control method based on multi-agent distributed control according to claim 1 is characterized in that: The step 3 is specifically as follows: Step 3.1: Assign a corresponding agent to each device: Processor Load Agent: C=[C1,C2,…,C n1 ] C k ={P Ck =p C (f),Q Ck =q C (P Ck ),CP k =cp(f)} Where n1 is the number of processors, C k represents the kth processor, P Ck is the power consumption of the kth processor, Q Ck is the heat generated by the kth processor, CP k is the computing power of the kth processor; Storage device load agent: H=[H1,H2,…,H n2 ] H o ={P Ho =p H (cp),Q Ho =q H (P Ho )} Among them, n2 is the number of storage devices, H o Indicates the working status of the oth storage device, including the power consumption P of the storage device Ho and storage device heat Q Ho ; Switch equipment load agent: S=[S1,S2,…,S n3 ] S w ={P Sw =p S (cp),Q Sw =q S (P Sw )} Among them, n3 is the number of switch devices, S w represents the working status of the wth switch, P Sw represents the power consumption of the wth switch device; Q Sw is the heat generated by the wth switch; Refrigeration Equipment Load Agent: <h2 style=";text-align:left;direction:ltr">A=[A1,A2,…,A<h2 style=";text-align:left;direction:ltr"> n4 <h2 style=";text-align:left;direction:ltr"> ] Among them, n4 is the number of refrigeration equipment, A v represents the working state of the vth refrigeration equipment, Q Av is the calorific value of the vth refrigeration equipment; and the refrigeration equipment power P A The size of is related to the heat generated by the device it regulates, and is represented by a function p A (·)definition, is the maximum cooling capacity of the vth refrigeration equipment; Step 3.2: Abstract the above device into a multi-agent system consisting of four agents. The communication graph between them is G = (V, E, A), where V = I [1, N] represents the node set, E ∈ I [1, N] × I [1, N] represents the edge set, and A = A [a ij ] N×N represents the adjacency matrix, a ij Indicates the connection state, 1 indicates connection, 0 indicates non-connection, and defines the Laplace matrix of the adjacency matrix L = L[l ij ] N×N , N = 4 is the number of agents; Step 3.3: Model the communication graph between agents as a random switching graph: G σ(t) =(V,E σ(t) ,A σ(t) ), in several different graphs G 1 ,G 2 …G s Switch between; G if and only if σ(t)=q σ(t) =G q ; The switching signal σ(t) obeys a Markov process that takes values ​​in a finite state space S = {1,…,s}, where s is the number of state spaces, and Λ = [λ mq ] s×s is the conversion rate matrix, where λ mq satisfy: l mm =-∑ q≠m l mq ,l mq >0 Among them, o(Δt) represents the higher-order infinitesimal of time t; Step 3.4: Create the following agent model: in, represents the state of the ith agent, represents the power control input of the ith agent, and They represent the system matrices respectively.

5. The computing center load optimization control method based on multi-agent distributed control as claimed in claim 4 is characterized in that: The step 4 is specifically as follows: Step 4.1: Establish the following optimization objective function: Among them, P i (x Pi ) is the function of each agent to calculate the internal power consumption, cp(x Pi ) is the computing power function of the processor load agent, and its ratio f i (x Pi ) is a cost function known only to each agent, which is differentiable, emergent and satisfies: Where R i is a negative definite matrix, satisfy: The CPU will cp(x Pi ) reports to each intelligent agent, the processor load intelligent agent, the storage device load intelligent agent and the switch device load intelligent agent report the heat to the cooling device load intelligent agent; Step 4.2: Through synchronous communication between neighbors, design a fully distributed event-triggered control scheme for the i-th agent as follows: in, For x Pi The open-loop estimate of μ ij , ij is the adaptive gain, F is the feedback gain matrix to be specified; v ij , o ij is a positive constant; Defining State Estimates Define sensor error: The k+1th trigger time of the information transmission from the i-th agent to its neighboring agents is determined by the following trigger mechanism: Among them, the trigger function h i (e i ,π i ), Designed to: Among them, δ, δ1, δ2, γ, γ1, γ2, γ3, are all positive numbers; π i , is the internal dynamic variable of the ith agent and is greater than zero; When the i-th agent is triggered, keep h i (e i ,π i )>0 or Reset i = 0, then the i-th agent transmits the current information to all external neighbors, saving computing resources and reducing the demand for communication bandwidth; Step 4.3: By iterating and converging the power consumption reference values ​​of each intelligent agent, the power consumption requirements of each intelligent agent are inversely calculated, and the power is allocated to the response of each device so that the device can operate in the optimal energy consumption state that can meet the computing requirements.

Images