An optimal scheduling method for data center workloads considering server temperature

Abstract

The present invention discloses a method for optimally scheduling data center workloads that considers server temperature. The method comprises the following steps: predicting the operating parameters of each server when processing a scheduled computing task based on the requirements of the scheduled computing task and the utilization rate of existing server resources; obtaining the heat generation power consumption of the server chips based on the predicted parameters; constructing a chip heat transfer model; obtaining the server chip temperature based on the chip heat transfer model and the server chip heat generation power consumption; obtaining the total server power consumption based on the server chip heat generation power consumption and the server chip temperature in combination with other server power consumption; and performing scheduling based on the total server power consumption. The present invention ultimately achieves two functions: first, optimizing the total server energy consumption; and second, maximizing the renewable energy absorption rate of the data center.

Description

Technical Field

[0001] The present invention belongs to the technical field of data center server power consumption scheduling, and in particular relates to an optimal scheduling method for data center work tasks taking server temperature into consideration. Background Art

[0002] Existing server energy consumption models, such as additive, baseline load, and regression models, can describe server power consumption at varying load rates, frequencies, and voltages. However, with the increasing demand for high-density computing and cooling, characterizing server operating temperature has become crucial for safe, reliable, economical operation, and intelligent regulation. Unlike the instantaneous response of computing power consumption to load, server operating temperature exhibits nonlinear, delayed variations due to the thermal inertia of server thermal resistance and heat capacity. This, in turn, impacts power optimization for upper-tier servers and operational regulation of lower-tier cooling systems. Furthermore, in the context of AI, technologies such as computing power transfer can lead to more frequent fluctuations in server utilization. Therefore, based on the heat generation and heat transfer mechanisms of server chips, it is necessary to clarify and reveal the patterns, mechanisms, and models of temporal variations in operating temperature and heat dissipation under high- and low-frequency step changes in server utilization, thereby accurately characterizing server power consumption and heat dissipation.

[0003] CN109189190A discloses a neural network-based data center temperature prediction algorithm. Based on the assumption of thermal locality, the temperature around a server is primarily affected by the configuration of its neighboring servers, while servers farther away have little impact. Due to this local nature, modeling tasks can be distributed among servers: each server uses its own data and the data from the two servers above and below it to build a model and predict its own inlet temperature. An artificial neural network prediction model is trained using real-time server node operation data, wireless sensor data, and CFD simulation data to predict the inlet temperature distribution of data center servers. This method does not reveal the temperature variation patterns caused by the thermal inertia of the server thermal resistance and heat capacity when server utilization changes. When a batch of workloads arrives at the data center, the data center uses load balancing technology to evenly distribute the tasks across multiple servers. Existing load balancing algorithms include round robin, least connections, source IP hashing, and URL hashing. The selection of these algorithms depends primarily on whether the performance of the servers is uniform, whether the request processing time is consistent, whether session state needs to be maintained, and whether the number of servers changes frequently. However, with the rapid increase in computing power demand and the development of high-density data centers, server energy consumption is rapidly increasing. Furthermore, server energy consumption is related to parameters such as voltage, frequency, and temperature, and the relationship is not a simple linear one. Therefore, the aforementioned load balancing algorithms can cause abnormal peaks on some servers, exacerbating the high energy consumption problem in high-density data centers.

[0004] In summary, server chip operating temperature significantly impacts its power consumption. When server utilization changes dramatically, existing server chip temperature acquisition methods cannot accurately characterize the chip's heat transfer properties, such as thermal capacitance and thermal resistance. Furthermore, existing data center energy-saving scheduling methods fail to account for changes in server power consumption caused by chip temperature. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes an optimal scheduling method for data center work tasks taking into account server temperature to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above objectives, the present invention provides a method for optimally scheduling data center workloads taking server temperature into consideration, comprising the following steps:

[0007] Based on the demand of the computing task to be scheduled and the utilization rate of existing server resources, the operating parameters of each server when processing the computing task to be scheduled are predicted, and the heat power consumption of the server chip is obtained based on the predicted parameters;

[0008] Constructing a chip heat transfer model, and obtaining the server chip temperature based on the chip heat transfer model and the heat generation power consumption of the server chip;

[0009] Obtaining the total power consumption of the server based on the heat generation power consumption of the server chip and the server chip temperature in combination with other power consumptions of the server;

[0010] Scheduling is performed based on the total power consumption of the server.

[0011] Preferably, the expression for obtaining the heat consumption of the server chip is:

[0012]

[0013] Where: E SW Indicates the number of logic gate state transitions per clock cycle, V dd represents the terminal voltage, f represents the crystal oscillator frequency, I leak represents leakage current, and C represents load capacitance.

[0014] Preferably, the chip heat transfer model is expressed as:

[0015]

[0016] Where C th Represents the chip thermal capacity, R th1 Represents the first conduction thermal resistance, R th2 Represents the second thermal conduction resistance, R th3 Represents the third conduction thermal resistance, R th4represents the convective heat transfer resistance, and T represents the transmission temperature.

[0017] Preferably, the expressions of the conductive thermal resistance and the convective heat transfer resistance are:

[0018]

[0019] Where δ1, δ2, and δ3 represent the thickness of the chip, upper packaging material, and lower packaging material, respectively; λ1, λ2, and λ3 represent the thermal conductivity of the chip, upper packaging material, and lower packaging material, respectively; A s1 、A s2 、A s3 、A s4 They represent the areas of the chip, upper packaging material, lower packaging material and PCB board respectively, and h represents the convection heat transfer coefficient.

[0020] Preferably, the chip heat capacity is expressed as:

[0021] C th =ρVc;

[0022] Where ρ is the specific heat capacity of the chip material, V is the density of the chip material, and c is the volume of the chip material.

[0023] Preferably, the total power consumption of the server includes the power consumption of the server chip and other power consumption;

[0024] Server chip power consumption includes idle power consumption, dynamic power consumption and leakage power consumption;

[0025] Dynamic power consumption includes state transition power consumption and unit internal power consumption;

[0026] The leakage power consumption is obtained based on the server chip temperature.

[0027] Preferably, the method for scheduling based on the total power consumption of the server includes:

[0028] Taking the real-time utilization of each server in the data center as the decision variable, the server chip operating voltage, current, heat generation and surface temperature as intermediate variables, and the minimum power consumption of the entire stack of the data center as the objective function, the computing task scheduling situation that minimizes the total power consumption of the data center servers is obtained, and scheduling is performed based on the computing task scheduling situation that minimizes the total power consumption of the servers.

[0029] Preferably, the method for scheduling based on the total power consumption of the server further includes:

[0030] Taking the renewable energy consumption rate of the data center as the objective function, server load balancing is performed under the conditions of satisfying the electricity market constraints, operating cost constraints, peak demand constraints and operating temperature range constraints. The computing task scheduling situation that maximizes the renewable energy consumption rate of the data center is iteratively screened out, and scheduling is performed based on the computing task scheduling situation with the maximum renewable energy consumption rate.

[0031] Compared with the prior art, the present invention has the following advantages and technical effects:

[0032] First, this invention accurately characterizes the dynamic changes in server chip temperature caused by server thermal capacitance and thermal resistance as server utilization fluctuates, ensuring the chip operates within a safe and stable temperature range. Second, based on a temperature-sensitive server power consumption model, it utilizes an intelligent algorithm to iteratively optimize the allocation of computing tasks. Ultimately, this approach achieves two functions: optimizing total server energy consumption and maximizing the renewable energy consumption rate in the data center. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0034] Figure 1 This is a flow chart of a scheduling method according to an embodiment of the present invention;

[0035] Figure 2 This is a device for obtaining heat power consumption of a server chip according to an embodiment of the present invention;

[0036] Figure 3 A chip heat transfer model according to an embodiment of the present invention;

[0037] Figure 4 A device for obtaining server chip temperature according to an embodiment of the present invention;

[0038] Figure 5 A device for obtaining server power consumption according to an embodiment of the present invention;

[0039] Figure 6 A server load balancing device is provided as a function of an embodiment of the present invention;

[0040] Figure 7 This is a server load balancing device for function 2 according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0042] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0043] Example 1

[0044] like Figure 1 As shown, this embodiment provides an optimal scheduling method for data center work tasks considering server temperature, including the following steps:

[0045] Based on the demand of the computing task to be scheduled and the utilization rate of existing server resources, the operating parameters of each server when processing the computing task to be scheduled are predicted, and the heat power consumption of the server chip is obtained based on the predicted parameters;

[0046] Constructing a chip heat transfer model, and obtaining the server chip temperature based on the chip heat transfer model and the heat generation power consumption of the server chip;

[0047] Obtaining the total power consumption of the server based on the heat generation power consumption of the server chip and the server chip temperature in combination with other power consumptions of the server;

[0048] Scheduling is performed based on the total power consumption of the server.

[0049] The specific steps are as follows:

[0050] Step 1: Obtain the heat dissipation of the server chip. When a batch of computing tasks arrives at the data center, the system can predict the operating voltage, current, utilization rate and other parameters of each server when processing the scheduled computing tasks based on the demand of the computing tasks to be scheduled and the utilization rate of existing server resources, based on the relationship obtained by experimental fitting, and obtain the heat dissipation based on the server chip heat generation model. Specifically:

[0051] When a data center server is running, the task load is compiled into binary machine code that accesses various components in the CPU. Therefore, the 0-1 transitions of the chip's underlying physical transistors are accompanied by the transitions of high- and low-level pulse currents, resulting in thermal power consumption. Resistive components such as resistors, capacitors, and inductors cause heat dissipation, so the chip's underlying physical circuit layer determines the chip's thermal power consumption. Based on the circuit-level power consumption model P = UI, and taking into account the chip's manufacturing process and operating conditions, the thermal power consumption of a server chip is calculated as follows:

[0052] Q=0.5·C·V d 2 d ·f·E sw +Q sc ·V dd ·f·E sw +I leak ·Vdd

[0053] The first item on the right represents the state transition power consumption, the second item is the unit internal power consumption, these two items are the energy consumption of the chip when it is in working state, collectively referred to as dynamic power consumption; the last item is the static power consumption. The parameters in the above formula are as follows: E SW It is called the conversion coefficient, which is the number of times the logic gate state changes in each clock cycle, V dd is the terminal voltage, f is the crystal oscillator frequency, I leak represents leakage current, and C represents load capacitance. During the heat transfer process of the chip, thermal power consumption is the heat transfer source term. Generally, for different servers, the functional relationship between the above parameters and server utilization can be obtained by experimental fitting. Figure 2 As shown in FIG, when a batch of computing tasks arrives at the data center, server voltage prediction, server current prediction, and server frequency prediction are performed based on server utilization to obtain the heat generation power of the relevant servers.

[0054] Step 2: Obtain the server chip temperature. Based on the server chip heat generation parameter Q transmitted by the heat generation model, the chip heat transfer model can be used to obtain the changes in the real-time server surface temperature T. Specifically:

[0055] After determining the heat source term, it is necessary to mathematically model the heat transfer path and heat dissipation conditions of the chip package, complete the construction of the thermal parameter network of thermal resistance-heat capacity, and calculate the real-time loss of the thermal network response results to obtain the real-time change of the chip temperature. In chip heat dissipation, there are three basic forms of heat transfer: heat conduction, heat convection, and heat radiation. In actual engineering applications, these three basic forms of heat transfer generally exist at the same time, and are even accompanied by a series of complex physical processes such as phase change of matter (such as condensers). However, for different heat dissipation systems, generally only one or several heat dissipation methods play a dominant role. When conducting system analysis, a reasonable selection should be made based on the actual situation of the system. In the present invention, the server adopts a pump-driven two-phase liquid cooling system for cooling. For this system, only heat conduction and heat convection need to be considered, and the heat dissipation of heat radiation can be ignored.

[0056] When modeling the chip heat transfer model, such as Figure 3 As shown, the heat transfer model is simplified.

[0057] The chip heat transfer model expression is as follows:

[0058]

[0059] In this model, the calculation method for the conductive thermal resistance of the chip and package and the convective heat transfer thermal resistance between the chip substrate and the coolant can be expressed as:

[0060]

[0061]

[0062] Where R th1 、R th2 、R th3 Represents the conduction thermal resistance; R th4 represents the convection heat transfer thermal resistance, δ1, δ2, and δ3 represent the thickness of the chip, upper packaging material, and lower packaging material, respectively; λ1, λ2, and λ3 represent the thermal conductivity of the chip, upper packaging material, and lower packaging material, respectively; A s1 、A s2 、A s3 、A s4 They represent the areas of the chip, upper packaging material, lower packaging material and PCB board respectively, and h represents the convection heat transfer coefficient.

[0063] Chip thermal capacitance C th for:

[0064] C th =ρVc

[0065] Where ρ is the specific heat capacity of the chip material, V is the density of the chip material, and c is the volume of the chip material.

[0066] The above method can be used to obtain the chip thermal resistance and chip thermal capacity. After establishing the unsteady-state heat conduction equation, the boundary conditions must be obtained before the solution can be found. For the initial conditions, assuming that the chip surface has reached the steady-state temperature T0 at the previous moment before the utilization rate changes, then:

[0067] T(x,0)=T0

[0068] As for the boundary conditions, the coolant heat flux density at different times is controlled by controlling the coolant temperature, flow rate and other parameters in the cooling system microchannel. Therefore, this is the second type of boundary condition, which is:

[0069] When τ>0,

[0070] Where τ is time, λ is thermal conductivity, T is temperature, and x is any x distance in the cross section.

[0071] It can be seen that the chip temperature is obtained by coupling its heat generation, heat transfer and heat dissipation by the auxiliary cooling system.

[0072] like Figure 4 As shown, by combining the above steps, the server chip temperature can be obtained.

[0073] Step 3: Obtain the server power consumption. Based on the above parameters such as server utilization and temperature, the change of the server's real-time total power consumption P can be accurately predicted. Specifically:

[0074] After calculating the server chip surface temperature using the above method, its energy consumption can be calculated using the following server energy consumption additive model:

[0075] P server =P CPU +P other

[0076] Where, P CPU P is the CPU power consumption, which fluctuates greatly due to utilization and temperature. other The power consumption of other parts such as memory and disk is small and does not fluctuate much, so it is usually regarded as a constant value.

[0077] Among them, CPU power consumption consists of three parts:

[0078] P CPU =P idle +P dyn +P leak

[0079] Where, P idle Indicates idle power consumption, which is a fixed constant. Dynamic power consumption P dyn It is generated by the switching process of the transistors that make up the CPU and is related to the CPU's execution of computing tasks. It can be expressed as the CPU utilization u:

[0080] P dyn =au

[0081] Where a is the coefficient obtained by fitting the measurement results.

[0082] Idle power consumption can be directly obtained according to different server models, while dynamic power consumption can be obtained through coefficients and chip utilization obtained by chip fitting.

[0083] CPU leakage power consumption P leak Generated by transistor leakage current, it is related to the internal carrier movement and is affected by temperature. The leakage power consumption can be obtained by comparing it with the chip temperature T CPU The relevant exponential form is:

[0084]

[0085] Therefore, the power consumption of the CPU chip is:

[0086]

[0087] Then, if Figure 5 As shown, the total power consumption of the server can be obtained by combining the power consumption of other parts of the server.

[0088] Step 4: Server load balancing redistributes utilization. Server load balancing is performed under constraints such as network cost, queuing delay, peak demand, and operating temperature range. The computing task scheduling (i.e., the utilization of each server) that minimizes the total power consumption of the data center servers is iteratively selected. Specifically:

[0089] Based on the predicted power consumption of each server, the present invention can achieve two scheduling goals. Using intelligent algorithms, the real-time utilization of each server in the data center is used as the decision variable, the server chip operating voltage, current, heat generation, surface temperature and other parameters are used as intermediate variables, and the minimum power consumption of the entire data center stack is used as the objective function. Figure 6 As shown, server load balancing is performed under relevant constraints such as network cost constraints, queuing delay constraints, peak demand constraints, and operating temperature range constraints, and the computing task scheduling situation that minimizes the total power consumption of the data center server is iteratively screened, which is the invention function 1. The invention function 2 takes into account the existence of a certain safe operating temperature range and thermal inertia of the data center server, and takes the renewable energy consumption rate of the data center as an example. Figure 7 As shown, server load balancing is performed under relevant constraints such as electricity market constraints, operating cost constraints, peak demand constraints and operating temperature range constraints, and the computing task scheduling situation that maximizes the renewable energy absorption rate of the data center is iteratively screened, which is function 2 of the invention.

[0090] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for optimal scheduling of data center workloads taking into account server temperature, characterized in that: The following steps are involved: Based on the demand of the computing task to be scheduled and the utilization rate of existing server resources, the operating parameters of each server when processing the computing task to be scheduled are predicted, and the heat power consumption of the server chip is obtained based on the predicted parameters; The expression for obtaining the heat consumption of the server chip is: Where: E SW Indicates the number of logic gate state transitions per clock cycle, V dd represents the terminal voltage, f represents the crystal oscillator frequency, I leak Represents leakage current, C represents load capacitance; Constructing a chip heat transfer model, and obtaining the server chip temperature based on the chip heat transfer model and the heat generation power consumption of the server chip; The expression of the chip heat transfer model is: Where C th Represents the chip thermal capacity, R th1 Represents the first conduction thermal resistance, R th2 Represents the second thermal conduction resistance, R th3 Represents the third conduction thermal resistance, R th4 represents the convective heat transfer resistance, T represents the transmission temperature; The expressions of the conduction thermal resistance and convection heat transfer thermal resistance are: Where δ1, δ2, and δ3 represent the thickness of the chip, upper packaging material, and lower packaging material, respectively; λ1, λ2, and λ3 represent the thermal conductivity of the chip, upper packaging material, and lower packaging material, respectively; A s1 、A s2 、A s3 、A s4 They represent the areas of the chip, upper packaging material, lower packaging material and PCB board respectively, and h represents the convection heat transfer coefficient; The expression of the chip heat capacity is: C th =ρVc; Where ρ is the specific heat capacity of the chip material, V is the density of the chip material, and c is the volume of the chip material; Obtaining the total power consumption of the server based on the heat generation power consumption of the server chip and the server chip temperature in combination with other power consumptions of the server; Scheduling is performed based on the total power consumption of the server.

2. The optimal scheduling method for data center workloads considering server temperature according to claim 1, characterized in that: The total power consumption of the server includes the power consumption of the server chip and other power consumption; Server chip power consumption includes idle power consumption, dynamic power consumption and leakage power consumption; Dynamic power consumption includes state transition power consumption and unit internal power consumption; The leakage power consumption is obtained based on the server chip temperature.

3. The optimal scheduling method for data center workloads considering server temperature according to claim 1, characterized in that: The method for scheduling based on the total power consumption of the server includes: Taking the real-time utilization of each server in the data center as the decision variable, the server chip operating voltage, current, heat generation and surface temperature as intermediate variables, and the minimum power consumption of the entire stack of the data center as the objective function, the computing task scheduling situation that minimizes the total power consumption of the data center servers is obtained, and scheduling is performed based on the computing task scheduling situation that minimizes the total power consumption of the servers.

4. The optimal scheduling method for data center workloads considering server temperature according to claim 1, characterized in that: The method for scheduling based on the total power consumption of the server further includes: Taking the renewable energy consumption rate of the data center as the objective function, server load balancing is performed under the conditions of satisfying the electricity market constraints, operating cost constraints, peak demand constraints and operating temperature range constraints. The computing task scheduling situation that maximizes the renewable energy consumption rate of the data center is iteratively screened out, and scheduling is performed based on the computing task scheduling situation with the maximum renewable energy consumption rate.

Images