A collaborative control method for intelligent computing center servers and cold plate liquid cooling systems

Abstract

The present invention provides a collaborative control method applicable to servers and cold plate liquid cooling systems in an intelligent computing center, comprising the following steps: Step 10, obtaining IT system scheduling data and liquid cooling system data of the intelligent computing center within a preset time period, and establishing a server real-time power consumption model, a cooling tower power consumption model, a water pump power consumption model, and a server temperature rapid prediction model respectively; Step 20, based on the server real-time power consumption model, the cooling tower power consumption model, the water pump power consumption model, and the server temperature rapid prediction model, adopting an optimization method combining real-time joint and timing sequence to collaboratively control virtual machine deployment and the cold plate liquid cooling system. The collaborative control method applicable to servers and cold plate liquid cooling systems in an intelligent computing center provided by the present invention can solve the overcooling technical problem that is prone to occur in energy-saving scheduling of the intelligent computing center, so as to improve the overall energy efficiency and operational stability of the intelligent computing center.

Description

Technical Field

[0001] The present invention belongs to the technical field of liquid cooling control of intelligent computing centers, and specifically relates to a collaborative control method suitable for intelligent computing center servers and cold plate liquid cooling systems. Background Art

[0002] As the infrastructure for artificial intelligence (AI) technology, intelligent computing centers have experienced rapid development in recent years. Compared to traditional data centers, intelligent computing centers have higher power density, with single racks capable of exceeding 40kW. Therefore, cold plate liquid cooling systems are often used for cooling. Compared to servers in general computing centers, the power consumption of servers in intelligent computing centers fluctuates extremely rapidly. However, existing control strategies fail to effectively coordinate servers and cooling systems, resulting in severe overcooling problems in intelligent computing centers and significant energy waste.

[0003] In order to improve the precise cooling capacity of the cold plate liquid cooling system, an effective method currently used is to add electric valves to the inlet / outlet water pipes of the end rack / server of the secondary loop (micromodule loop), and based on the server temperature prediction, the precise control of the cooling capacity is achieved through the linkage of the liquid cooling distribution unit (CDU) and the valve. However, since the length of the secondary loop pipeline is usually short and its hydraulic characteristics are relatively complex, the valve control of a single branch can easily cause drastic changes in the flow of each branch of the entire pipeline network, thus affecting the safety and stability of the intelligent computing center. In addition, due to the limitations of fluid distribution and heat capacity characteristics, the cooling capacity supply of the liquid cooling system is difficult to dynamically match the heat dissipation changes of each intelligent computing server (equipped with GPU). To this end, it is necessary to explore the collaborative control technology of the server and the cold plate liquid cooling system. Summary of the Invention

[0004] The present invention provides a collaborative control method applicable to the server and cold plate liquid cooling system of the intelligent computing center, so as to solve the technical problem of overcooling that is prone to occur in the energy-saving scheduling of the intelligent computing center, thereby improving the overall energy efficiency and operation stability of the intelligent computing center.

[0005] To solve the above technical problems, the embodiments of the present invention adopt the following technical solutions:

[0006] An embodiment of the present invention provides a collaborative control method applicable to an intelligent computing center server and a cold plate liquid cooling system, comprising the following steps:

[0007] Step 10: Obtain IT system scheduling data and liquid cooling system data from the intelligent computing center within a preset time period, and establish a real-time server power consumption model, a cooling tower power consumption model, a water pump power consumption model, and a server temperature rapid prediction model.

[0008] Step 20, based on the server real-time power consumption model, cooling tower power consumption model, water pump power consumption model and server temperature rapid prediction model, adopts the optimization method combining real-time joint and timed sequence to coordinately control the virtual machine deployment and cold plate liquid cooling system.

[0009] As a further improvement of the embodiment of the present invention, the IT system scheduling data includes the CPU resource requirements of the virtual machine, the GPU resource requirements of the virtual machine, the mapping relationship between the virtual machine and the server, the CPU utilization rate of each server, the GPU utilization rate of each server, the real-time power consumption of each server and the chip temperature of each server; the liquid cooling system data includes the flow rate of each water pump on the primary side, the power consumption of each water pump on the primary side, the opening degree of each valve on the primary side, the flow rate of each CDU water pump, the power consumption of each CDU water pump, the outdoor wet-bulb temperature, the inlet and outlet water temperature difference of each cooling tower, the approach degree of each cooling tower, the power consumption of each cooling tower and the flow rate of each cooling tower.

[0010] As a further improvement of the embodiment of the present invention, in step 10, a real-time power consumption model of the server of the intelligent computing center is established based on the IT system scheduling data;

[0011] Specifically include:

[0012] Step 101: Filter the power consumption data of each server at different CPU and GPU usage rates from IT system scheduling data;

[0013] Step 102: Establish a server static power consumption matrix based on the power consumption data of each server when the CPU and GPU usage rates are zero.

[0014] Step 103: Based on the data of the change of each server's real-time power consumption as a function of CPU usage, a multiple linear regression method is used to fit the conversion coefficient between each server's CPU usage and dynamic power consumption, thereby forming a conversion coefficient matrix between CPU usage and dynamic power consumption. Based on the data of the change of each server's real-time power consumption as a function of GPU usage, a multiple linear regression method is used to fit the conversion coefficient between each server's GPU usage and dynamic power consumption, thereby forming a conversion coefficient matrix between GPU usage and dynamic power consumption.

[0015] Step 104: Combine the server static power consumption matrix, the conversion coefficient matrix between CPU usage and running power consumption, and the conversion coefficient matrix between GPU usage and running power consumption to obtain the real-time power consumption model of the intelligent computing center server shown in formula (1):

[0016] P PM (t) = P SC (t)+u VM,CPU (t)*MM VM→PM (A(t),F(t))*f h,CPU +u VM,GPU (t)*MMVM→PM (A(t),F(t))*f h,GPU

[0017] Formula (1)

[0018] Where, P PM (t) represents the server real-time power consumption matrix, P SC (t) represents the server static power consumption matrix, u VM,CPU (t) represents the CPU resource requirement matrix of the virtual machine, MM VM→PM (A(t), F(t)) represents the mapping matrix between virtual machines and servers, A(t) represents the virtual machine real-time deployment matrix, F(t) represents the virtual machine migration matrix, u VM,CPU (t)*MM VM→PM (A(t), F(t)) represents the CPU usage of each server, f h,CPU Represents the conversion coefficient matrix between CPU utilization and dynamic power consumption, u VM,GPU (t) represents the GPU resource requirement matrix of the virtual machine, u VM,GPU (t)*MM VM→PM (A(t), F(t)) represents the GPU utilization rate of each server, f h,GPU A conversion coefficient matrix representing GPU utilization and dynamic power consumption.

[0019] As a further improvement of the embodiment of the present invention, in step 10, a cooling tower power consumption model shown in formula (2) is established based on the liquid cooling system data:

[0020]

[0021] Where, P T (t) represents the real-time power consumption of the cooling tower, P T,e Indicates the rated operating power consumption of the cooling tower, ΔT T (t) represents the real-time temperature difference between the inlet and outlet water of the cooling tower, ΔT T,e Indicates the rated operating temperature difference of the cooling tower inlet and outlet water, T app (t) represents the real-time approximation of the cooling tower, T app,e Indicates the degree of approximation to the rated operating conditions of the cooling tower, T wb (t) represents the outdoor wet-bulb real-time temperature, T wb,e Indicates the outdoor wet-bulb rated operating temperature, m cw (t) represents the real-time flow rate of the cooling tower, m cw,e represents the rated operating flow of the cooling tower, and a, b, c, d, e, f, g, h, i, j, k, l, m, n, and o represent the fitting coefficients of each order term respectively.

[0022] As a further improvement of the embodiment of the present invention, in step 10, based on the liquid cooling system data, a water pump power consumption model shown in formula (3) is established for the primary side water pump and each CDU water pump:

[0023]

[0024] Where, P pump (t) represents the real-time power consumption of the water pump, m(t) represents the real-time flow rate of the water pump, m des Indicates the rated operating flow of the pump, P pump,des represents the rated power consumption of the pump, b1, b2, and n represent the fitting coefficients respectively.

[0025] As a further improvement of the embodiment of the present invention, in step 10, a server temperature rapid prediction model is established based on IT system scheduling data and liquid cooling system data;

[0026] Specifically include:

[0027] Step 131: Filter out the mapping relationship between virtual machines and servers, virtual machine CPU resource requirements, virtual machine GPU resource requirements, and the maximum temperature of each server chip from the IT system scheduling data; and filter out the primary side water pump flow, primary side water pump power consumption, primary side valve opening, each CDU water pump flow, each CDU water pump power consumption, outdoor wet-bulb temperature, each cooling tower inlet and outlet water temperature difference, each cooling tower approximation, and each cooling tower power consumption from the liquid cooling system data to form a modeling data set;

[0028] Step 132: Based on the modeling data set, the water pump flow of each CDU, the outdoor wet-bulb temperature, the approximation degree of each cooling tower, the water pump flow of each primary side, the opening degree of each valve on the primary side, the CPU utilization rate of each server, and the GPU utilization rate of each server are used as input variables of the convolutional neural network model. The maximum chip temperature of each server is used as the output variable of the convolutional neural network model. The convolutional neural network model is trained to obtain a server temperature rapid prediction model.

[0029] As a further improvement of the embodiment of the present invention, in step 20, a real-time joint optimization method and a timed sequence optimization method are used to coordinately control virtual machine deployment and the cold plate liquid cooling system, including:

[0030] Step 21: When the cloud platform receives a new virtual machine deployment request, it randomly selects a deployable intelligent computing micro-module as the deployment micro-module, and jointly optimizes the deployment position of the new virtual machine in the deployment micro-module and the flow rate of each CDU water pump corresponding to the deployment micro-module to minimize the sum of the power consumption of the server deploying the micro-module and the corresponding CDU water pumps.

[0031] Step 22: At each given moment, first migrate the virtual machines running on each server as a whole to achieve the minimum number of intelligent computing micro-modules turned on; then optimize the water pump flow of each CDU, the water pump flow of each primary loop, the opening of each valve on the primary side, and the approach degree of each cooling tower to achieve the minimum total energy consumption of the cold plate liquid cooling system.

[0032] As a further improvement of the embodiment of the present invention, step 21 specifically includes:

[0033] Step 211: When the cloud platform receives a deployment request for a new virtual machine, it searches the remaining CPU and GPU resources of all active servers based on the CPU and GPU resource requirements of the new virtual machine to find a server available for deployment.

[0034] In step 212, if there are servers available for deployment, the intelligent computing micro-module where the servers available for deployment are located is used as the deployment micro-module for deploying the new virtual machine, the servers available for deployment are used as the optimization variables for deploying the new virtual machine, and step 213 is executed. If the remaining CPU resources and GPU resources of the active servers in all intelligent computing micro-modules are insufficient to deploy the new virtual machine, a intelligent computing micro-module with a dormant server is randomly selected as the deployment micro-module, a dormant server in the deployment micro-module is randomly switched to an active state, the new virtual machine is deployed, and step 214 is executed.

[0035] Step 213: With the optimization goal of minimizing the total power consumption of the deployed micromodule after the new virtual machine is deployed, and with the optimization constraints that the sum of the resources used by all virtual machines running on each server does not exceed the total rated resources of the server, and that the maximum temperature of each server chip does not exceed a preset temperature threshold, a genetic algorithm is used to jointly optimize the deployment position of the new virtual machine deployed in the micromodule and the water pump flow in each CDU corresponding to the deployed micromodule, to obtain the optimal real-time deployment position of the virtual machine and the optimal water pump flow of each CDU. The cloud platform implements the deployment of the new virtual machine based on the optimal real-time deployment position of the virtual machine, and the cold plate liquid cooling system controls the flow of each CDU water pump corresponding to the deployed micromodule based on the optimal water pump flow of each CDU.

[0036] In step 214, a genetic algorithm is used to optimize the flow rates of the CDU water pumps corresponding to the deployed micromodules, with the optimization goal of minimizing the sum of the power consumption of the CDU water pumps corresponding to the deployed micromodules after the new virtual machine is deployed, and with the constraint that the maximum temperature of each server chip cannot exceed a preset temperature threshold, to obtain the optimal flow rates of the CDU water pumps. The cold plate liquid cooling system controls the flow rates of the CDU water pumps corresponding to the deployed micromodules based on the optimal flow rates of the CDU water pumps.

[0037] As a further improvement of the embodiment of the present invention, step 22 specifically includes:

[0038] Step 221 , at a given moment, calculate the minimum number of active servers required based on the sum of the number of all virtual machines running on all servers in the intelligent computing center and their resource requirements;

[0039] Step 222: Based on the required minimum number of active servers, select servers as active servers for the next period according to the order of the intelligent computing micro-module numbers, the order of the rack numbers in each intelligent computing micro-module, and the order of the server numbers on each rack, until the number of selected servers reaches the minimum number of active servers.

[0040] Step 223, according to the load balancing strategy, the existing virtual machines are evenly migrated to the selected active servers; the idle servers are hibernated;

[0041] Step 224 , with minimizing the total power consumption of the cold plate liquid cooling system as the optimization goal and with the maximum temperature of each server chip not exceeding a preset temperature threshold as the constraint, a genetic algorithm is used to optimize the CDU water pump flow rate, the primary side water pump flow rate, the primary side valve opening, and the cooling tower proximity. The optimal CDU water pump flow rate, the optimal primary side water pump flow rate, the optimal primary side valve opening, and the optimal cooling tower proximity are obtained.

[0042] In step 225, the cold plate liquid cooling system regulates the flow of each CDU water pump according to the optimal flow of each CDU water pump, regulates the flow of each primary side water pump according to the optimal flow of each primary side water pump, regulates the opening of each primary side valve according to the optimal opening of each primary side valve, and regulates the approach degree of each cooling tower according to the optimal approach degree of each cooling tower.

[0043] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0044] The present invention provides a collaborative control method for the server and cold plate liquid cooling system of the intelligent computing center. First, the IT system scheduling data and liquid cooling system data of the intelligent computing center within a preset time period are obtained, and a server real-time power consumption model, a cooling tower power consumption model, a water pump power consumption model and a server temperature rapid prediction model are established respectively; then, in the control stage, when real-time deployment optimization is adopted, the deployment position of the new virtual machine in the deployment micromodule and the corresponding CDU water pump flow are jointly optimized to improve the real-time matching between the server heat dissipation and the secondary loop cooling capacity distribution in the deployment micromodule, so as to reduce the overcooling in the deployment micromodule, thereby realizing the deployment of a single virtual machine. The sum of the power consumption of the servers and CDU water pumps is the lowest; when using timed sequence optimization, the virtual machines running on each server are first migrated as a whole at a given time node to minimize the number of intelligent computing micro-modules turned on. On this basis, the flow rate of each CDU water pump, the flow rate of each water pump in the primary loop, the opening degree of each valve on the primary side and the approach degree of each cooling tower are optimized to achieve the lowest total energy consumption of the cold plate liquid cooling system, and promote the timing high matching between the heat dissipation of the server and the cooling distribution of the secondary loop, thereby further improving the cooling utilization efficiency of the intelligent computing center; through the coordination of real-time joint optimization and timed sequence optimization, the overall energy efficiency and operational stability of the intelligent computing center can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0046] Figure 1 This is a structural module diagram of an intelligent computing center to which the method according to the embodiment of the present invention is applicable;

[0047] Figure 2 The embodiment of the present invention is applicable to the collaborative control method of the intelligent computing center server and the cold plate liquid cooling system;

[0048] Figure 3 is a flowchart of real-time joint optimization in the method of an embodiment of the present invention;

[0049] Figure 4 This is a flowchart of timing sequence optimization in the method of an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0051] The embodiment of the present invention provides a collaborative control method for an intelligent computing center server and a cold plate liquid cooling system, which is applied to Figure 1 The intelligent computing center shown in the figure. The intelligent computing center includes intelligent computing micro-modules and a cold plate liquid cooling system. The cold plate liquid cooling system is divided into a primary loop (covering the loop from the cooling tower to each CDU) and a secondary loop (the loop from each CDU to each server). Among them, the primary loop includes a cooling tower and a primary water pump, and there is a one-to-one correspondence between the cooling tower and the primary water pump (by adjusting each primary water pump, the flow rate of the corresponding cooling tower can be controlled). In addition, each branch return pipe of the primary loop is equipped with an electric valve, and the opening of each valve can be adjusted as needed. The secondary loop is the loop from each CDU to each server, which is composed of multiple intelligent computing micro-module loops. Due to the limited length of the intelligent computing micro-module loop, hydraulic imbalance cannot be avoided. Therefore, electric regulating valves are usually not installed in the loop. Therefore, the loop flow can only be adjusted by the water pump in the CDU, making it difficult to achieve accurate distribution of cooling capacity.

[0052] The embodiment of the present invention provides a collaborative control method for an intelligent computing center server and a cold plate liquid cooling system, such as Figure 2 As shown, the following steps are included:

[0053] Step 10: Obtain the IT system scheduling data and liquid cooling system data of the intelligent computing center within a preset time period; establish a real-time power consumption model of the server of the intelligent computing center based on the IT system scheduling data; establish a cooling tower power consumption model and a water pump power consumption model of the intelligent computing center based on the liquid cooling system data; establish a rapid temperature prediction model of the server of the intelligent computing center based on the IT system scheduling data and the liquid cooling system data.

[0054] Step 20, based on the server real-time power consumption model, cooling tower power consumption model, water pump power consumption model and server temperature rapid prediction model, adopts the optimization method combining real-time joint and timed sequence to coordinately control the virtual machine deployment and cold plate liquid cooling system.

[0055] In step 10 above, IT system scheduling data is obtained from the scheduling data stored in the cloud platform. The IT system scheduling data includes the CPU resource requirements of the virtual machine, the GPU resource requirements of the virtual machine, the mapping relationship between the virtual machine and the server, the CPU utilization rate of each server, the GPU utilization rate of each server, the real-time power consumption of each server, and the chip temperature of each server. The liquid cooling system data is obtained from the management system of the cold plate liquid cooling system. The liquid cooling system data includes the flow rate of each water pump on the primary side, the power consumption of each water pump on the primary side, the opening degree of each valve on the primary side, the flow rate of each CDU water pump, the power consumption of each CDU water pump, the outdoor wet-bulb temperature, the inlet and outlet water temperature difference of each cooling tower, the proximity of each cooling tower, the power consumption of each cooling tower, and the flow rate of each cooling tower. Preferably, the preset time period is not less than 24 hours.

[0056] In the above step 10, a real-time power consumption model of the server is established based on the IT system scheduling data using a matrix modeling method.

[0057] Specifically include:

[0058] Step 101 : Filter out the power consumption data of each server at different CPU and GPU usage rates from IT system scheduling data.

[0059] Step 102 : establishing a server static power consumption matrix based on the power consumption data of each server when the CPU and GPU usage rates are 0.

[0060] Step 103: Based on the data on the variation of each server's real-time power consumption as a function of CPU usage, a multivariate linear regression method is used to fit the conversion coefficient between each server's CPU usage and dynamic power consumption (dynamic power consumption is calculated by subtracting static power consumption from real-time power consumption under standard server operating conditions), thereby forming a conversion coefficient matrix between CPU usage and dynamic power consumption. Based on the data on the variation of each server's real-time power consumption as a function of GPU usage, a multivariate linear regression method is used to fit the conversion coefficient between each server's GPU usage and dynamic power consumption, thereby forming a conversion coefficient matrix between GPU usage and dynamic power consumption.

[0061] Step 104: Combine the server static power consumption matrix, the conversion coefficient matrix between CPU usage and running power consumption, and the conversion coefficient matrix between GPU usage and running power consumption to obtain the real-time power consumption model of the intelligent computing center server shown in formula (1):

[0062] P PM (t) = P SC (t)+u VM,CPU (t)*MM VM→PM (A(t),F(t))*f h,CPU +u VM,GPU (t)*MM VM→PM (A(t),F(t))*f h,GPU

[0063] Formula (1)

[0064] Where, P PM (t) represents the server real-time power consumption matrix, matrix P PM Each element in (t) represents the real-time power consumption model of a server, and each element corresponds to each server one by one. SC (t) represents the server static power consumption matrix. VM,CPU (t) represents the CPU resource requirement matrix of the virtual machine. VM→PM(A(t), F(t)) represents the mapping matrix between virtual machines and servers. It is a function of the virtual machine real-time deployment matrix A(t) and the virtual machine migration matrix F(t) and can be expressed as a (0,1) matrix. For example, Indicates that VM 1 is deployed on server 1, and VM 2 is deployed on server 2.

[0065] u VM,CPU (t)*MM VM→PM (A(t), F(t)) represents the CPU usage of each server, f h,CPU Represents the conversion coefficient matrix between CPU utilization and dynamic power consumption, u VM,GPU (t) represents the GPU resource requirement matrix of the virtual machine, u VM,GPU (t)*MM VM→PM (A(t), F(t)) represents the GPU utilization rate of each server, f h,GPU A conversion coefficient matrix representing GPU utilization and dynamic power consumption.

[0066] In the established real-time power consumption model of the intelligent computing center server, the server real-time power consumption consists of static power consumption, CPU dynamic power consumption and GPU dynamic power consumption. Among them, static power consumption is a fixed value related to server performance, CPU dynamic power consumption is a function of the server CPU utilization rate, and GPU dynamic power consumption is a function of the server GPU utilization rate.

[0067] In the above step 10, based on the liquid cooling system data, the cooling tower power consumption model shown in formula (2) is established:

[0068]

[0069] Where, P T (t) represents the real-time power consumption of the cooling tower, P T,e Indicates the rated operating power consumption of the cooling tower, ΔT T (t) represents the real-time temperature difference between the inlet and outlet water of the cooling tower, ΔT T,e Indicates the rated operating temperature difference of the cooling tower inlet and outlet water, T app (t) represents the real-time approximation of the cooling tower, T app,e Indicates the degree of approximation to the rated operating conditions of the cooling tower, T wb (t) represents the outdoor wet-bulb real-time temperature, T wb,e Indicates the outdoor wet-bulb rated operating temperature, m cw (t) represents the real-time flow rate of the cooling tower, m cw,e represents the rated operating flow of the cooling tower, and a, b, c, d, e, f, g, h, i, j, k, l, m, n, and o represent the fitting coefficients of each order term respectively.

[0070] In step 10 above, based on the liquid cooling system data, the water pump power consumption model shown in formula (3) is established for the primary side water pump and each CDU water pump:

[0071]

[0072] Where, P pump (t) represents the real-time power consumption of the water pump, m(t) represents the real-time flow rate of the water pump, m des Indicates the rated operating flow of the pump, P pump,des represents the rated power consumption of the pump, b1, b2, and n represent the fitting coefficients respectively.

[0073] In the above step 10, a convolutional neural network model is established, with the water pump flow of each CDU, the outdoor wet-bulb temperature, the approximation of each cooling tower, the water pump flow of each primary side, the opening of each valve on the primary side, the CPU utilization rate of each server, and the GPU utilization rate of each server as the input variables of the convolutional neural network model, and the maximum chip temperature of each server as the output variable of the convolutional neural network model, to construct a dynamic mapping relationship between the water pump flow of each CDU, the outdoor wet-bulb temperature, the approximation of each cooling tower, the water pump flow of each primary side, the opening of each valve on the primary side, the CPU utilization rate of each server, the GPU utilization rate of each server and the maximum chip temperature of each server.

[0074] Specifically include:

[0075] Step 131: Filter the mapping relationship between virtual machines and servers, virtual machine CPU resource requirements, virtual machine GPU resource requirements, and the maximum temperature of each server chip from the IT system scheduling data. Filter the primary water pump flow rate, primary water pump power consumption, primary valve opening, CDU water pump flow rate, CDU water pump power consumption, outdoor wet-bulb temperature, cooling tower inlet and outlet water temperature difference, cooling tower approximation, and cooling tower power consumption from the liquid cooling system data to form a modeling dataset. The modeling dataset is divided into a training set and a test set with a 4:1 ratio.

[0076] In step 132, the water pump flow of each CDU, the outdoor wet-bulb temperature, the approximation degree of each cooling tower, the water pump flow of each primary side, the opening degree of each valve on the primary side, the CPU utilization rate of each server (calculated based on the mapping relationship between the virtual machine and the server and the CPU resource demand of the virtual machine), and the GPU utilization rate of each server (calculated based on the mapping relationship between the virtual machine and the server and the GPU resource demand of the virtual machine) are used as input variables of the convolutional neural network model, and the maximum chip temperature of each server is used as the output variable of the convolutional neural network model. The convolutional neural network model is trained using the training set, and hyperparameters such as the number of layers of the model and the number of neurons in each layer are tuned during the training process.

[0077] The accuracy of the trained convolutional neural network model is tested using a test set. If the model error meets the requirements (for example, less than 1%), the model is considered qualified and a rapid server temperature prediction model is obtained.

[0078] In step 20 above, a real-time joint and timed sequence optimization method is used to coordinately control virtual machine deployment and the cold plate liquid cooling system, including:

[0079] Step 21: When the cloud platform receives a new virtual machine deployment request, it randomly selects a deployable intelligent computing micro-module as the deployment micro-module, and jointly optimizes the deployment position of the new virtual machine in the deployment micro-module and the flow rate of each CDU water pump corresponding to the deployment module to achieve the lowest sum of the power consumption of the server of the deployment micro-module and the corresponding CDU water pumps.

[0080] At each given time, in step 22, all virtual machines running on each server are migrated to minimize the number of active intelligent computing modules. Next, the CDU pump flow rate, primary loop pump flow rate, primary side valve opening, and cooling tower proximity are optimized to minimize the total energy consumption of the cold plate liquid cooling system. Preferably, the time interval between two consecutive given times is 3 to 15 minutes.

[0081] Preferably, Figure 3 As shown, step 21 specifically includes:

[0082] In step 211 , after receiving the deployment request of the new virtual machine, the cloud platform searches the remaining CPU resources of all active servers according to the CPU resource requirements and GPU resource requirements of the new virtual machine to find a server available for deployment.

[0083] In step 212, if there are servers available for deployment, the intelligent computing micro-module where the servers available for deployment are located is used as the deployment micro-module for the new virtual machine deployment, the servers available for deployment are used as the optimization variables for the new virtual machine deployment, and step 213 is executed; if the remaining CPU and GPU resources of the active servers in all intelligent computing micro-modules cannot deploy new virtual machines, then a smart computing micro-module with a dormant server is randomly selected as the deployment micro-module, a dormant server in the deployment micro-module is randomly turned into an active state, a new virtual machine is deployed, and step 214 is executed.

[0084] In step 213, with the optimization goal of minimizing the total power consumption of the deployed micro-module after the new virtual machine is deployed, and with the optimization constraints that the sum of the resources used by all virtual machines running on each server does not exceed the total rated resources of the server, and that the maximum temperature of the chip on each server does not exceed the preset temperature threshold, a genetic algorithm is used to jointly optimize the deployment position of the new virtual machine in the deployed micro-module and the water pump flow in each CDU corresponding to the deployed micro-module, to obtain the optimal real-time deployment position of the virtual machine and the optimal water pump flow of each CDU.

[0085] As follows:

[0086] Optimization goal: minP total,PM+CDU,mc (t) = ∑P PM,mc (t)+∑P pump,CDU,mc (t)

[0087] Optimization constraint: T PM (t)≤T max,PM and∑u VMiin PMk(t) ≤C PM

[0088] Optimization variables: mc (t),m CDU,mc (t)}

[0089] Where, ∑P PM,mc (t) represents the sum of the real-time power consumption of each server in the deployed micromodule, ∑P pump,CDU,mc (t) represents the sum of the power consumption of each CDU pump corresponding to the deployed micromodule, T PM (t) represents the maximum temperature of each server chip, T max,PM Indicates the preset temperature threshold, u VMiin PMk(t) represents the sum of resources used by all virtual machines running on each server, C PM Indicates the total rated resources of the server, A mc (t) represents the real-time deployment matrix of virtual machines deploying micromodules, m CDU,mc (t) represents the flow matrix of each CDU pump corresponding to the deployed micromodule.

[0090] The cloud platform implements the deployment of new virtual machines based on the optimal real-time deployment location of the virtual machines, and the cold plate liquid cooling system controls the flow of each CDU water pump corresponding to the deployed micromodules based on the optimal flow of each CDU water pump.

[0091] In step 214, the optimization goal is to minimize the sum of the power consumption of each CDU water pump corresponding to the deployed micromodule after the new virtual machine is deployed. Under the constraint that the maximum temperature of each server chip cannot exceed the preset temperature threshold, a genetic algorithm is used to optimize the flow rate of each CDU water pump corresponding to the deployed micromodule to obtain the optimal flow rate of each CDU water pump.

[0092] As follows:

[0093] Optimization goal: minP total,CDU,mc (t) = ∑P pump,CDU,mc (t)

[0094] Optimization constraint: T PM (t)≤T max,PM

[0095] Optimization variables: {m CDU,mc (t)}.

[0096] The cold plate liquid cooling system controls the flow of each CDU water pump corresponding to the deployed micromodule according to the optimal flow of each CDU water pump.

[0097] In the method of this embodiment, during real-time optimization, a server available for deployment of a new virtual machine is first found from active servers. If so, the intelligent computing micromodule where the deployable server is located is used as the deployment micromodule, and the new virtual machine deployment location of the deployment micromodule and the flow rates of the CDU water pumps corresponding to the deployment micromodule are jointly optimized, so that the sum of the power consumption of the server where the micromodule is deployed and the CDU water pumps corresponding to the deployment micromodule are minimized. If not, a dormant server is randomly started for deployment, and the flow rates of the CDU water pumps corresponding to the deployment micromodule are optimized, so that the sum of the power consumption of the CDU water pumps corresponding to the deployment micromodule is minimized. This improves the real-time matching between server heat dissipation and secondary loop cooling capacity distribution in the deployment micromodule, thereby reducing the amount of overcooling within the deployment micromodule.

[0098] Preferably, Figure 4 As shown, step 22 specifically includes:

[0099] Step 221, at a given moment, calculate the minimum number of active servers required based on the total number of virtual machines running on all servers in the intelligent computing center and the sum of their resource requirements (calculate the sum of CPU resource requirements and the sum of GPU resource requirements separately).

[0100] Step 222: Based on the required minimum number of active servers, select servers as active servers for the next period according to the order of the intelligent computing micro-module numbers, the order of the rack numbers in each intelligent computing micro-module, and the order of the server numbers on each rack, until the number of selected servers reaches the minimum number of active servers.

[0101] Step 223: According to the load balancing strategy, the existing virtual machines are evenly migrated to the selected active servers and the idle servers are hibernated.

[0102] In step 224, with the goal of minimizing the total power consumption of the cold plate liquid cooling system and the constraint that the maximum temperature of each server chip does not exceed a preset temperature threshold, a genetic algorithm is used to optimize the CDU water pump flow rate, the primary side water pump flow rate, the primary side valve opening, and the cooling tower approximation degree to obtain the optimal CDU water pump flow rate, the optimal primary side water pump flow rate, the optimal primary side valve opening, and the optimal cooling tower approximation degree.

[0103] As follows:

[0104] Optimization goal: minP total,cooling (t) = ∑P pump,CDU (t)+∑P pump,cw (t)+∑P T (t)

[0105] Optimization constraint: T PM (t)≤T max,PM

[0106] Optimization variables: {m CDU (t),m cw (t),V av (t),ΔT app (t)}.

[0107] Where, ∑P pump,CDU (t) represents the sum of the water pump power consumption of each CDU, ∑P pump,cw (t) represents the sum of the power consumption of the primary side water pump, ∑P T (t) represents the sum of the power consumption of each cooling tower.

[0108] In step 225, the cold plate liquid cooling system regulates the flow of each CDU water pump according to the optimal flow of each CDU water pump, regulates the flow of each primary side water pump according to the optimal flow of each primary side water pump, regulates the opening of each primary side valve according to the optimal opening of each primary side valve, and regulates the approach degree of each cooling tower according to the optimal approach degree of each cooling tower.

[0109] In the method of this embodiment, during the timing optimization, the virtual machines are first migrated as a whole, and the idle servers after the migration are hibernated. This can minimize the number of intelligent computing micro-modules that are turned on, and the location of active servers within a single intelligent computing micro-module is more concentrated, and the total power consumption of the servers is minimized. Then, on this basis, the flow rate of each CDU pump, the flow rate of each water pump in the primary loop, the opening degree of each valve on the primary side, and the proximity of each cooling tower are optimized to minimize the total energy consumption of the cold plate liquid cooling system. This promotes a high degree of timing matching between the heat dissipation of the servers in the intelligent computing center and the cooling supply of the cold plate liquid cooling system, thereby reducing the overall overcooling of the intelligent computing center.

[0110] The method of this embodiment, when there is a new virtual machine deployment request, performs real-time joint optimization of the placement of new virtual machines in a single deployment micromodule and the corresponding CDU water pump flow rate, so as to improve the real-time matching between the heat dissipation of the servers in the deployment micromodule and the cooling capacity distribution of the secondary loop, thereby reducing the amount of overcooling in the micromodule. However, the process of real-time deployment optimization is prone to cause the dispersion of active servers, that is, the deployment of virtual machines that can be accommodated by a single intelligent computing micromodule is dispersed across multiple intelligent computing micromodules. Although the total power consumption of the server has not changed, due to the lack of means for fine-grained control of the cooling capacity in the secondary loop, the dispersion of active servers is prone to cause a large amount of cooling capacity waste.

[0111] At each given moment, the virtual machine deployment and cold plate liquid cooling system are optimized and controlled in a timed sequence. First, the virtual machines running on each server are migrated as a whole to minimize the total power consumption of the servers and facilitate centralized adjustment of active servers. On this basis, the flow rate of each CDU pump, the flow rate of each primary loop pump, the opening degree of each valve on the primary side, and the proximity of each cooling tower are optimized to achieve the lowest overall power consumption of the cold plate liquid cooling system and facilitate a highly timed match between the heat dissipation of the servers in the intelligent computing center and the cooling supply of the cold plate liquid cooling system, thereby reducing the overall overcooling of the intelligent computing center. However, if only timed sequence optimization is performed, the frequent migration of virtual machines can easily affect their service quality, and the frequent adjustment of the opening degree of each valve on the primary side and the proximity of each cooling tower can easily cause hydraulic imbalance in the primary side pipe network and a significant reduction in the life of the cooling tower fan. Therefore, the time interval for timed sequence optimization should not be too short, and a long time interval obviously cannot meet the real-time deployment requirements of virtual machines.

[0112] The method of this embodiment combines real-time deployment optimization with timing sequence optimization. Through the timing coordination of the two, the cooling efficiency of the intelligent computing center can be significantly improved, and the overall energy efficiency and operational stability of the intelligent computing center can be improved.

[0113] In the method of the present invention, during the operation of the server, the existing virtual machines are automatically released (shut down in real time) after completing their tasks. If all virtual machines on a server are released, the server is considered an idle server and will automatically sleep (power consumption is reduced to 0 in real time).

[0114] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the specific embodiments described above. The specific embodiments and descriptions in the specification are only intended to further illustrate the principles of the present invention. Various changes and improvements are possible without departing from the spirit and scope of the present invention, and such changes and improvements fall within the scope of the invention as claimed.

Claims

1. A collaborative control method for an intelligent computing center server and a cold plate liquid cooling system, characterized in that: The following steps are involved: Step 10: Obtain IT system scheduling data and liquid cooling system data from the intelligent computing center within a preset time period, and establish a real-time server power consumption model, a cooling tower power consumption model, a water pump power consumption model, and a server temperature rapid prediction model. Step 20: Based on the server real-time power consumption model, the cooling tower power consumption model, the water pump power consumption model, and the server temperature rapid prediction model, a real-time joint and timed sequence optimization method is used to coordinately control virtual machine deployment and the cold plate liquid cooling system. The step 20 specifically includes: Step 211: When the cloud platform receives a deployment request for a new virtual machine, it searches the remaining CPU and GPU resources of all active servers based on the CPU and GPU resource requirements of the new virtual machine to find a server available for deployment. In step 212, if there are servers available for deployment, the intelligent computing micro-module where the servers available for deployment are located is used as the deployment micro-module for deploying the new virtual machine, the servers available for deployment are used as the optimization variables for deploying the new virtual machine, and step 213 is executed. If the remaining CPU resources and GPU resources of the active servers in all intelligent computing micro-modules are insufficient to deploy the new virtual machine, a intelligent computing micro-module with a dormant server is randomly selected as the deployment micro-module, a dormant server in the deployment micro-module is randomly switched to an active state, the new virtual machine is deployed, and step 214 is executed. Step 213, with minimizing the total power consumption of the deployed micromodule after the new virtual machine is deployed as the optimization goal, and with the sum of the resources used by all virtual machines running on each server not exceeding the total rated resources of the server, and the maximum temperature of each server chip not exceeding a preset temperature threshold as the optimization constraints, a genetic algorithm is used to jointly optimize the deployment position of the new virtual machine deployed in the micromodule and the water pump flow in each CDU corresponding to the deployed micromodule, to obtain the optimal real-time deployment position of the virtual machine and the optimal water pump flow of each CDU; the total power consumption of the deployed micromodule is the sum of the real-time power consumption of each server in the deployed micromodule and the sum of the water pump power consumption of each CDU corresponding to the deployed micromodule; Step 214: With the optimization goal of minimizing the sum of the power consumption of the CDU water pumps corresponding to the deployed micromodules after the new virtual machine is deployed, and with the constraint that the maximum temperature of each server chip cannot exceed a preset temperature threshold, a genetic algorithm is used to optimize the flow rate of the CDU water pumps corresponding to the deployed micromodules to obtain the optimal flow rate of each CDU water pump. Step 22: At each given moment, first migrate the virtual machines running on each server as a whole to achieve the minimum number of intelligent computing micro-modules turned on; then, with the lowest total power consumption of the cold plate liquid cooling system as the optimization goal, and with the maximum temperature of each server chip not exceeding the preset temperature threshold as the constraint, optimize the water pump flow of each CDU, the water pump flow of each primary loop, the opening of each valve on the primary side, and the approximation of each cooling tower to achieve the minimum total energy consumption of the cold plate liquid cooling system; the total energy consumption of the cold plate liquid cooling system is the sum of the water pump power consumption of each CDU, the sum of the water pump power consumption of the primary side, and the sum of the power consumption of each cooling tower.

2. The collaborative control method for intelligent computing center servers and cold plate liquid cooling systems according to claim 1 is characterized in that: The IT system scheduling data includes the CPU resource requirements of the virtual machine, the GPU resource requirements of the virtual machine, the mapping relationship between the virtual machine and the server, the CPU utilization rate of each server, the GPU utilization rate of each server, the real-time power consumption of each server and the chip temperature of each server; the liquid cooling system data includes the flow rate of each water pump on the primary side, the power consumption of each water pump on the primary side, the opening degree of each valve on the primary side, the flow rate of each CDU water pump, the power consumption of each CDU water pump, the outdoor wet-bulb temperature, the inlet and outlet water temperature difference of each cooling tower, the approach degree of each cooling tower, the power consumption of each cooling tower and the flow rate of each cooling tower.

3. The collaborative control method for intelligent computing center servers and cold plate liquid cooling systems according to claim 2, characterized in that: In step 10, a real-time power consumption model of the intelligent computing center server is established based on the IT system scheduling data; Specifically include: Step 101: Filter the power consumption data of each server at different CPU and GPU usage rates from IT system scheduling data; Step 102: Establish a server static power consumption matrix based on the power consumption data of each server when the CPU and GPU usage rates are zero. Step 103: Based on the data of the change of the real-time power consumption of each server with the CPU usage rate, a multiple linear regression method is used to fit the conversion coefficient of the CPU usage rate and dynamic power consumption of each server, thereby forming a conversion coefficient matrix of the CPU usage rate and dynamic power consumption; based on the data of the change of the real-time power consumption of each server with the GPU usage rate, a multiple linear regression method is used to fit the conversion coefficient of the GPU usage rate and dynamic power consumption of each server, thereby forming a conversion coefficient matrix of the GPU usage rate and dynamic power consumption; In step 104, the server static power consumption matrix, the conversion coefficient matrix between CPU usage and running power consumption, and the conversion coefficient matrix between GPU usage and running power consumption are combined to obtain the real-time power consumption model of the intelligent computing center server shown in formula (1): Formula (1) Where, represents the server real-time power consumption matrix, represents the server static power consumption matrix, Represents the CPU resource requirement matrix of the virtual machine, Represents the mapping matrix between virtual machines and servers, represents the real-time deployment matrix of virtual machines, represents the virtual machine migration matrix, Indicates the CPU usage of each server. Represents the conversion coefficient matrix between CPU usage and dynamic power consumption, Represents the GPU resource requirement matrix of the virtual machine, Indicates the GPU usage of each server, A conversion coefficient matrix representing GPU utilization and dynamic power consumption.

4. The collaborative control method for intelligent computing center servers and cold plate liquid cooling systems according to claim 2, characterized in that: In step 10, based on the liquid cooling system data, a cooling tower power consumption model as shown in formula (2) is established: Formula (2) Where, Indicates the real-time power consumption of the cooling tower. Indicates the cooling tower's rated operating power consumption. Indicates the real-time temperature difference between the inlet and outlet water of the cooling tower. Indicates the rated operating temperature difference of the cooling tower inlet and outlet water. Indicates the real-time approach of the cooling tower, Indicates the degree of approximation to the rated operating conditions of the cooling tower. Indicates the outdoor wet-bulb real-time temperature. Indicates the outdoor wet-bulb rated operating temperature. Indicates the real-time flow of the cooling tower. Indicates the rated operating flow of the cooling tower. Represent the fitting coefficients of each order term.

5. The collaborative control method for intelligent computing center servers and cold plate liquid cooling systems according to claim 2, characterized in that: In step 10, based on the liquid cooling system data, the water pump power consumption model shown in formula (3) is established for the primary side water pump and each CDU water pump: Formula (3) Where, Indicates the real-time power consumption of the water pump. Indicates the real-time flow of the water pump. Indicates the rated operating flow of the pump. Indicates the rated working power consumption of the pump. represent the fitting coefficients respectively.

6. The collaborative control method for intelligent computing center servers and cold plate liquid cooling systems according to claim 2, characterized in that: In step 10, a server temperature rapid prediction model is established based on IT system scheduling data and liquid cooling system data; Specifically include: Step 131: Filter out the mapping relationship between virtual machines and servers, virtual machine CPU resource requirements, virtual machine GPU resource requirements, and the maximum temperature of each server chip from the IT system scheduling data; and filter out the primary side water pump flow, primary side water pump power consumption, primary side valve opening, each CDU water pump flow, each CDU water pump power consumption, outdoor wet-bulb temperature, each cooling tower inlet and outlet water temperature difference, each cooling tower approximation, and each cooling tower power consumption from the liquid cooling system data to form a modeling data set; Step 132: Based on the modeling data set, the water pump flow of each CDU, the outdoor wet-bulb temperature, the approximation degree of each cooling tower, the water pump flow of each primary side, the opening degree of each valve on the primary side, the CPU utilization rate of each server, and the GPU utilization rate of each server are used as input variables of the convolutional neural network model. The maximum chip temperature of each server is used as the output variable of the convolutional neural network model. The convolutional neural network model is trained to obtain a server temperature rapid prediction model.

7. The collaborative control method for intelligent computing center servers and cold plate liquid cooling systems according to claim 1, characterized in that: The step 22 specifically includes: Step 221 , at a given moment, calculate the minimum number of active servers required based on the sum of the number of all virtual machines running on all servers in the intelligent computing center and their resource requirements; Step 222: Based on the required minimum number of active servers, select servers as active servers for the next period according to the order of the intelligent computing micro-module numbers, the order of the rack numbers in each intelligent computing micro-module, and the order of the server numbers on each rack, until the number of selected servers reaches the minimum number of active servers. Step 223, according to the load balancing strategy, the existing virtual machines are evenly migrated to the selected active servers; the idle servers are hibernated; Step 224 , with minimizing the total power consumption of the cold plate liquid cooling system as the optimization goal and with the maximum temperature of each server chip not exceeding a preset temperature threshold as the constraint, a genetic algorithm is used to optimize the CDU water pump flow rate, the primary side water pump flow rate, the primary side valve opening, and the cooling tower proximity. The optimal CDU water pump flow rate, the optimal primary side water pump flow rate, the optimal primary side valve opening, and the optimal cooling tower proximity are obtained. In step 225, the cold plate liquid cooling system regulates the flow of each CDU water pump according to the optimal flow of each CDU water pump, regulates the flow of each primary side water pump according to the optimal flow of each primary side water pump, regulates the opening of each primary side valve according to the optimal opening of each primary side valve, and regulates the approach degree of each cooling tower according to the optimal approach degree of each cooling tower.

Images