Data center energy consumption and response time joint optimization method based on MOFA

Abstract

The invention provides a joint optimization method of data center energy consumption and response time based on MOFA, which realizes normal form transformation of data center resource scheduling from extensive guarantee to intelligent equilibrium through innovation in three levels of modeling, optimization and decision making. Data center operation is converted into an intelligent system capable of sensing task content and dynamically adjusting resources; the joint optimization method comprises the following steps: constructing a data center energy consumption cost model and a service quality model; establishing a joint optimization objective function and constraint conditions; establishing an optimization model of MOFA, and carrying out multiple times of position updating to obtain a Pareto solution set; and a dynamic decision-making mechanism based on task type matching.

Description

Technical Field

[0001] This invention relates to the field of data center response strategy technology, and specifically to a joint optimization method for data center energy consumption and response time based on MOFA. Background Technology

[0002] With the rapid development of information and communication technologies, the number and scale of data centers worldwide continue to grow, leading to a significant increase in energy consumption due to their massive load. Data centers are typical "constantly high-load" users, and their 24 / 7 uninterrupted operation puts continuous pressure on the baseload supply of the power grid. During summer or peak electricity consumption periods, the cooling demand of data centers surges, and their peak energy consumption may overlap with the peak electricity consumption of urban residential and commercial users, which can easily lead to local power grid overload and trigger risks to power supply reliability.

[0003] Existing solutions often focus on the "facilities" and "hardware" levels, failing to dynamically and meticulously optimize from the fundamental source of "IT load scheduling." Traditional data center operations often prioritize service level agreements (SLAs) and tend to maintain an excessive number of active servers to cope with potential load peaks, resulting in a large number of servers "idling" at low utilization rates and causing huge waste of "idle power consumption."

[0004] The IT load of a data center is not inherently unadjustable; there is an optimizable dynamic balance between its energy consumption and service quality (such as task processing latency and response time).

[0005] To address this challenge, this invention proposes a joint optimization strategy. By adjusting the number of active servers, the energy consumption and service quality of a data center can be balanced. This strategy ensures that data center energy consumption is reduced without impacting user experience, thereby alleviating pressure on the power grid. Summary of the Invention

[0006] This invention aims to address the technical deficiencies of existing technologies by providing a joint optimization method for data center energy consumption and response time based on MOFA. Through innovations in modeling, optimization, and decision-making, it realizes a paradigm shift in data center resource scheduling from "extensive guarantee" to "intelligent balancing," transforming data center operation into an intelligent system capable of sensing task content and dynamically adjusting resources.

[0007] This invention discloses a joint optimization method for data center energy consumption and response time based on MOFA, comprising the following steps:

[0008] Step 1: Construct a data center energy consumption cost model and a service quality model;

[0009] Step 2: Establish the joint optimization objective function and constraints;

[0010] Step 3: Build an optimization model of MOFA and perform multiple position updates to obtain the Pareto solution set;

[0011] Step 4: Dynamic decision-making mechanism based on task type matching.

[0012] Furthermore,

[0013] In step 1, the energy consumption cost model of the data center is constructed. The energy consumption cost of the data center directly depends on the power consumption of the data center. It is assumed that the servers inside the data center are all the same equipment, and the power and performance of each server are exactly the same. The power consumption of the data center at a certain time period is mainly divided into three parts: the power of IT equipment, the power required for the cooling equipment to work, and other power.

[0014] The power consumption of the data center during a certain period is shown below:

[0015] ,

[0016] In the formula: It is the total power consumed by the data center during a certain period of time. It is the power consumed by IT equipment in a data center at a certain time. This is the power consumed by the cooling equipment in the data center at a certain time period.

[0017] This refers to the power consumed by other devices in the data center at a certain time.

[0018] Power of IT equipment It mainly consists of server power consumption, and the power consumption of a single server is shown below:

[0019] ,

[0020] In the formula: This refers to the power consumption of a single server. It is the idle power consumption of the server during a certain period of time. It represents the server's full-load power consumption at a certain time. It represents the CPU utilization rate during a certain period of time;

[0021] Assume the number of servers running in the data center at a certain time period is To handle data stream load,

[0022] As shown below:

[0023] ,

[0024] In the formula: It refers to the data flow load processed by the data center at a certain time period. It is the processing speed of data streams handled by servers in a data center;

[0025] Therefore, the server power consumption of the data center at a certain time period It can be represented as:

[0026] ;

[0027] The power consumption of data center cooling equipment during a certain period is expressed as follows:

[0028] ,

[0029] In the formula: This refers to the cooling capacity of a data center at a certain time period. It is the cooling energy efficiency ratio of a data center at a certain time period;

[0030] The cooling capacity of the data center is shown below:

[0031] ,

[0032] In the formula: It is the indoor temperature. It's the outdoor temperature. It is the equivalent thermal resistance.

[0033] Furthermore,

[0034] In step 1, the data center service quality model is constructed. The service quality of a data center is usually measured by its average service response time.

[0035] The task response time is shown below:

[0036] ,

[0037] In the formula: It is the task response time. It is the probability of waiting;

[0038] It can be calculated using Erlang-C formulas:

[0039] ,

[0040] In the formula: This represents the number of servers currently in operation.

[0041] Furthermore,

[0042] In step 2, the joint optimization objective function F is established as follows:

[0043] ,

[0044] In the formula: It is the objective function for minimizing the energy consumption cost of data centers. It is the objective function for minimizing the data center response time;

[0045] ,

[0046] In the formula: It refers to the electricity price for a specific period of time;

[0047] .

[0048] Furthermore,

[0049] The constraints in step 2 include indoor temperature constraints, workload capacity constraints, response time constraints, cooling system power constraints, and upper limit constraints on server waiting probability.

[0050] Indoor temperature constraints: ,

[0051] In the formula: It is the minimum temperature required to maintain the normal operation of a data center; This is the maximum temperature that can sustain the normal operation of a data center;

[0052] Workload capacity handling constraints: ;

[0053] Response time constraints: ,

[0054] In the formula: This is the maximum response time that does not affect the user experience.

[0055] Refrigeration system power constraints: ,

[0056] In the formula: This is the upper limit of the power of the data center cooling system;

[0057] Upper limit constraint on server wait probability: .

[0058] Furthermore,

[0059] Step 3 involves establishing an optimization model for MOFA and performing multiple position updates to obtain the Pareto solution set, including the following steps:

[0060] Step 3.1: Randomly generate the initial population;

[0061] Random population generation: Generate N fireflies, each with its own position vector. for:

[0062] in, ,

[0063] In the formula: N is the population size, i.e., the number of fireflies, usually set to 50-100; D is the dimension of the decision variable. The j-th decision variable for the i-th firefly is randomly generated within the feasible range;

[0064] Step 3.2: Calculate the fitness value of each firefly;

[0065] The objective function value is to evaluate the brightness of each firefly, where brightness is determined by energy consumption cost. and response time In multi-objective optimization, brightness is defined based on the Pareto dominance relationship.

[0066] Pareto Dominance Rule: Solution Dominate If and only if and And at least one inequality holds strictly, F1(X) i (to solve) The corresponding energy consumption cost, F2(X) i (to solve) The corresponding response time, F1(X) j (to solve) The corresponding energy consumption cost, F2(X) j (to solve) The corresponding response time;

[0067] Step 3.3: Update the firefly positions;

[0068] Brightness can be defined as:

[0069] ,

[0070] For each pair of fireflies and ,if Then fireflies Attracted to move;

[0071] The attraction coefficient can be expressed as:

[0072] ,

[0073] In the formula: This is the initial attraction constant, usually set to 1.0. The light absorption coefficient controls the rate at which the attractive force decreases with distance, and is typically set to 0.01-1.0. firefly and The distance between them, where,

[0074] ,

[0075] The position update formula can be expressed as:

[0076] ,

[0077] In the formula: For the first Only fireflies are iterating The position at that time For the first Only fireflies are iterating The position at that time For the first Only fireflies are iterating The position at that time The random step size factor, It is a random vector whose elements follow a uniform or normal distribution;

[0078] Step 3.4: Update the fitness values ​​of each firefly.

[0079] After the location is updated, the objective function value and brightness of each firefly are recalculated to reflect the advantages and disadvantages of the new location;

[0080] Step 3.5: Pass the non-dominated solution set to the next generation;

[0081] Preserve the non-dominated solutions in the current iteration, i.e., the Pareto optimal solutions, to avoid losing high-quality solutions;

[0082] Step 3.6: Determine whether the maximum number of iterations has been reached. Repeat steps 3.1 to 3.6 to obtain a Pareto solution set.

[0083] Furthermore,

[0084] The dynamic decision-making mechanism based on task type matching in step 4 includes the following steps:

[0085] Step 4.1: Task Feature Identification: Determine the task feature vector by parsing the task request in real time or based on the task queue information;

[0086] Step 4.2: Matching Weight Calculation: Calculate the target preference weight W for this scheduling decision based on the task type;

[0087] ,

[0088] In the formula: , These are adjustable weighting coefficients used to control the influence of latency sensitivity and computational intensity on decision-making. S is the latency sensitivity coefficient, and C is the computational intensity coefficient. A larger W value indicates...

[0089] The smaller the W value, the more the decision-making process favors solutions with low latency; the smaller the W value, the more it favors solutions with low energy consumption.

[0090] Step 4.3: Compare the preference weights W calculated by the task with the attribute scores A of the Pareto solution set. k conduct

[0091] Matching, selecting solutions that satisfy the following conditions As the final scheduling strategy ;

[0092] ,

[0093] You can then select the attribute rating A. k The Pareto solution that is closest to the current task preference weight W. To solve It belongs to the Pareto solution set.

[0094] Furthermore,

[0095] The task feature vector in step 4.1 can be represented as:

[0096] ,

[0097] In the formula: This is the delay sensitivity coefficient, with a value range of [0, 1]. The closer the value is to 1, the more sensitive it is to low latency.

[0098] The higher the latency requirement; The intensity coefficient is calculated with a value range of [0, 1]. The closer the value is to 1, the more computing resources are consumed per unit of task.

[0099] Furthermore,

[0100] In step 4.3, the attribute score of the Pareto solution set is calculated as a comprehensive attribute score A for each solution. k This is used to characterize the tendency of the solution to lie on the "energy consumption-response time" trade-off curve, and can be expressed as:

[0101] ,

[0102] In the formula: To solve The corresponding energy consumption cost, To solve The corresponding response time ,

[0103] These represent the maximum and minimum energy consumption costs in the current Pareto solution set; , These are the maximum and minimum response times in the current Pareto solution set;

[0104] A k The value range of A is [-1, 1]. k The closer the value is to 1, the more the solution tends to optimize response time.

[0105] Suitable for latency-sensitive tasks; A k The closer the solution is to -1, the more it tends to optimize energy consumption, making it suitable for computationally intensive or ordinary tasks; A k A value close to 0 indicates that the solution achieves a good balance between the two.

[0106] Compared with the prior art, the beneficial effects of the present invention are:

[0107] This invention not only provides an advanced multi-objective optimization algorithm application, but also constructs a complete, closed-loop, and intelligent data center energy efficiency management methodology. Through innovations in modeling, optimization, and decision-making, it achieves a paradigm shift in data center resource scheduling from "extensive guarantee" to "intelligent balancing." Its beneficial effects are specifically reflected in the following aspects:

[0108] 1. Achieve a precise balance between energy consumption and performance, breaking through the traditional trade-off dilemma.

[0109] In traditional operations and maintenance, in order to ensure service quality (response time), over-configuration strategies are often adopted, leading to energy waste; conversely, blindly reducing resources in order to save energy will cause performance degradation. The multi-objective joint modeling of the effect of this solution, by constructing an accurate energy consumption cost model and service quality model, puts the two originally conflicting objectives (low energy consumption and low latency) under a unified mathematical framework for quantitative analysis.

[0110] By using the multi-objective firefly algorithm, a series of Pareto optimal solutions are obtained at once. Each solution represents the best balance between "energy consumption" and "response time" under the current load. This provides managers with a wealth of non-inferior decision options, rather than a single extreme solution.

[0111] This strategy allows data centers to reduce energy consumption without affecting service quality, and shorten task processing response time within the acceptable range of data center economic consumption.

[0112] 2. Introduce task-aware dynamic decision-making to enhance the intelligence and precision of resource scheduling.

[0113] Most optimization methods treat tasks as homogeneous queues and adopt a "one-size-fits-all" resource allocation strategy, failing to distinguish the different service level requirements of critical tasks and background tasks. This solution employs a dynamic decision-making mechanism that matches tasks based on task type and invokes different optimization preference strategies. By introducing this dynamic decision-making mechanism, this invention combines offline optimization with online decision-making, transforming data center operations into an intelligent system capable of sensing task content and dynamically adjusting resources. This not only ensures the service quality of critical tasks but also maximizes energy efficiency at the macro level, significantly enhancing commercial value and practicality. Attached Figure Description

[0114] Figure 1 This is a schematic diagram illustrating the specific process of the joint optimization strategy for data center energy consumption and response time based on MOFA of the present invention.

[0115] Figure 2 This constitutes the energy consumption of a data center;

[0116] Figure 3 This is a flowchart illustrating the specific process of MOFA used in this invention. Detailed Implementation

[0117] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0118] like Figures 1-3 As shown, it illustrates a specific embodiment of the present invention:

[0119] This invention discloses a joint optimization strategy for data center energy consumption and response time based on MOFA, comprising:

[0120] Step 1: Construct a data center energy consumption cost model and a service quality model;

[0121] Step 2: Establish the joint optimization objective function and constraints;

[0122] Step 3: Build an optimization model of MOFA and perform multiple position updates to obtain the Pareto solution set;

[0123] Step 4: Dynamic decision-making mechanism based on task type matching.

[0124] Step 1: Construct a data center energy consumption cost model and a service quality model.

[0125] In data center operations, energy cost and service quality are often two conflicting optimization goals. Reducing energy costs typically requires decreasing the number of running servers or improving energy efficiency, but this can increase task latency and response time, impacting user experience. Conversely, ensuring task availability usually requires increasing server processing power to reduce response time, but this leads to increased energy consumption. Therefore, finding a balance between these two goals and developing a joint optimization strategy is a key issue in data center energy management.

[0126] The model proposed in this invention takes energy consumption cost and service quality as two optimization objectives and adopts a joint optimization method to obtain multiple optimal solution sets. Different solutions can be selected according to the needs of different users to achieve energy saving without affecting the user experience.

[0127] First, we need to build a model for the energy consumption cost of data centers. The energy consumption cost of data centers directly depends on the power consumption of data centers. Therefore, optimizing the energy consumption cost of data centers is equivalent to optimizing the power consumption of data centers.

[0128] Assume that all servers within the data center are identical devices, and that each server has exactly the same power and performance.

[0129] The power consumption of a data center at a given time period can be mainly divided into three parts: the power of IT equipment, the power required for cooling equipment to operate, and other power.

[0130] The power consumption of the data center during a certain period is shown below:

[0131] ,

[0132] In the formula: It is the total power consumed by the data center during a certain period of time; This refers to the power consumed by IT equipment in a data center during a certain period of time. This is the power consumed by the cooling equipment in the data center during a certain period of time; It represents the power consumed by other devices in the data center at a certain time.

[0133] First, we need to study the IT equipment, including its power. It mainly consists of server power consumption. The power consumption of a single server is related to CPU utilization.

[0134] The power consumption of a single server is shown below:

[0135] ,

[0136] In the formula: This refers to the power consumption of a single server; It represents the idle power consumption of the server during a certain period of time; It represents the server's full-load power consumption during a certain period of time; It represents the CPU utilization rate at a certain time period.

[0137] Assume the number of servers running in the data center at a certain time period is To handle data stream load,

[0138] As shown below:

[0139] ,

[0140] In the formula: It refers to the data stream load processed by the data center during a certain period of time; It is the processing speed of data streams handled by servers in a data center.

[0141] Therefore, the server power consumption of the data center at a certain time period It can be represented as:

[0142] ,

[0143] Secondly, there's the power consumption of data center cooling equipment. The energy consumption of a cooling system is typically closely related to the external temperature of the data center and the heat generated by IT equipment. The efficiency of a cooling system is usually determined by its energy efficiency ratio (EER), which plays a crucial role in the power consumption of cooling systems.

[0144] The power consumption of data center cooling equipment during a certain period is expressed as follows:

[0145] ,

[0146] In the formula: This represents the cooling capacity of the data center at a given time period. It is the cooling energy efficiency ratio of a data center at a certain time period.

[0147] According to the principle of energy conservation, the thermodynamic model of a room requires that the cooling power of the refrigeration system be matched with the external environmental conditions, the power consumption of IT equipment, and the power of other systems in order to achieve energy balance.

[0148] The cooling capacity of the data center is shown below:

[0149] ,

[0150] In the formula: It is the indoor temperature; It is the outdoor temperature; It is the equivalent thermal resistance.

[0151] Secondly, there is the data center service quality optimization model.

[0152] The quality of service in a data center is typically measured by its average service response time; a lower response time means that user requests are answered more quickly. Therefore, improving the quality of service in a data center can essentially be seen as reducing the response time of tasks within the system.

[0153] The task response time is shown below:

[0154] ,

[0155] In the formula: It is the response time of the task; It represents the probability of waiting.

[0156] It can be calculated using Erlang-C formulas:

[0157] ,

[0158] In the formula: This represents the number of servers currently in operation.

[0159] Step 2: Establish the joint optimization objective function and constraints.

[0160] The proposed model takes energy cost and task latency as two optimization objectives and employs a joint optimization method. By dynamically adjusting factors such as server resources, load allocation, and cooling system power, it simultaneously optimizes these two objectives. By balancing the relationship between the two, the model can generate multiple Pareto optimal solutions, providing data centers with flexible scheduling schemes that satisfy task service quality while minimizing energy costs.

[0161] The multi-objective joint optimization model F for mission assurance and energy consumption cost is as follows:

[0162] ,

[0163] In the formula: It is the objective function for minimizing the energy consumption cost of data centers; It is the objective function for minimizing the response time of the data center.

[0164] ,

[0165] In the formula: It refers to the electricity price for a specific period.

[0166] ,

[0167] Secondly, some conditions are subject to constraints.

[0168] 1) Indoor Temperature Constraints. Data center temperature management needs to be maintained within a reasonable range to ensure the normal operation of IT equipment. Excessively high or low temperatures may affect equipment stability and performance. As shown in the following formula:

[0169] ,

[0170] In the formula: It is the minimum temperature required to maintain the normal operation of a data center; This is the maximum temperature that can sustain the normal operation of a data center.

[0171] 2) Workload capacity constraints. It is essential to ensure that the server's processing capacity exceeds the incoming workload to prevent overload. As shown in the following formula:

[0172] .

[0173] 3) Response time constraints. To ensure business continuity in the data center, the response time for task processing must not exceed the specified maximum response time limit, thereby ensuring that the data center meets business continuity and user experience requirements when processing tasks.

[0174] ,

[0175] In the formula: This is the maximum response time limit that does not affect the user experience.

[0176] 4) Cooling system power constraints. This constraint ensures that the cooling system does not operate under overload and guarantees that the system operates within a safe range. If the cooling system power is too low, the data center will overheat, while if the power is too high, it may lead to energy waste or system overload.

[0177] ,

[0178] In the formula: This is the upper limit of the power of the data center cooling system.

[0179] 5) Upper limit constraint on server waiting probability. This constraint limits the probability of the server waiting to process tasks, preventing too many waiting tasks from causing the system response time to become excessively long and exceed acceptable limits. This ensures the stable operation of the data center and the quality of service.

[0180] .

[0181] Step 3: Build an optimization model of MOFA and perform multiple position updates to obtain the Pareto solution set.

[0182] MOFA is a swarm intelligence optimization algorithm based on the bioluminescence behavior of fireflies in nature, suitable for solving multi-objective optimization problems. In this scheme, MOFA is used to balance two conflicting objectives: data center energy consumption cost F1 and response time F2. It searches for a Pareto optimal solution set by simulating the movement of fireflies. The Pareto solution set represents a set of solutions where improvement in one objective leads to a deterioration in the other, thus providing decision-makers with multiple trade-offs.

[0183] The core idea of ​​MOFA is to treat each candidate solution, such as the number of servers or indoor temperature, as a firefly, with its position representing the decision vector and its brightness determined by the objective function value. The algorithm updates the positions using the attraction between the fireflies, eventually converging to the Pareto front. Its flowchart is shown below. Figure 3 As shown:

[0184] Step 3.1: Randomly generate the initial population

[0185] Initialize the firefly population. The location of each firefly represents a set of decision variables, such as the number of servers and indoor temperature. These variables are used to optimize energy consumption and response time.

[0186] Random population generation: Generate N fireflies, each with its own position vector. for:

[0187] in, ,

[0188] In the formula: N is the population size, i.e. the number of fireflies, which is usually set to 50-100; D is the dimension of the decision variable; The j-th decision variable for the i-th firefly is randomly generated within the feasible range.

[0189] Step 3.2: Calculate the fitness value of each firefly.

[0190] The brightness of each firefly is evaluated (i.e., the objective function value), which is determined by energy consumption cost. and response time The determination is made jointly. In multi-objective optimization, brightness is defined based on the Pareto dominance relationship.

[0191] Pareto Dominance Rule: Solution Dominate If and only if and And at least one inequality holds strictly, F1(X) i (to solve) The corresponding energy consumption cost, F2(X) i (to solve) The corresponding response time, F1(X) j(to solve) The corresponding energy consumption cost, F2(X) j (to solve) The corresponding response time;

[0192] Step 3.3: Update the firefly positions

[0193] Fireflies are attracted to each other based on brightness; those with weaker light move towards those with stronger light, thus updating their positions. This step is the core of MOFA and is used to explore the solution space.

[0194] In MOFA, brightness It is usually related to the objective function value. For a minimization problem, brightness can be defined as:

[0195] ,

[0196] For each pair of fireflies and ,if Then fireflies Attracted to move.

[0197] The attraction coefficient can be expressed by the formula:

[0198] ,

[0199] In the formula: This is the initial attraction constant, usually set to 1.0; The light absorption coefficient controls the rate at which the attractive force decreases with distance, and is usually set to 0.01-1.0; The distance between fireflies i and j;

[0200] ,

[0201] The position update formula can be expressed as:

[0202] ,

[0203] In the formula: For the first Only fireflies are iterating The position at that time For the first Only fireflies are iterating The position at that time For the first Only fireflies are iterating The position at that time The random step size factor, It is a random vector whose elements follow a uniform or normal distribution.

[0204] Step 3.4: Update the fitness values ​​of each firefly.

[0205] After the location is updated, the objective function value and brightness of each firefly are recalculated to reflect the advantages and disadvantages of the new location.

[0206] Step 3.5: Pass the non-dominated solution set to the next generation.

[0207] Retain the non-dominated solutions, i.e. Pareto optimal solutions, in the current iteration to avoid losing high-quality solutions.

[0208] Step 3.6: Determine if the maximum number of iterations has been reached, and repeat the above operation continuously.

[0209] Ultimately, a set of Pareto solutions will be obtained. First, the trade-off between data center energy costs and task latency is not uniform; energy costs change more significantly during some periods and more slowly during others. Second, the Pareto front indicates that there is no single "optimal solution" in the multi-objective optimization model established in this paper, but rather a set of Pareto optimal solutions. Each Pareto front solution represents a different trade-off between energy costs and task latency. Therefore, decision-makers can choose the most suitable solution from the Pareto front solutions based on different application scenarios and needs; for example, data centers tend to prefer low energy consumption or low latency.

[0210] Step 4: Dynamic decision-making mechanism based on task type matching.

[0211] After obtaining the Pareto optimal solution set through the MOFA algorithm, this invention further proposes a dynamic decision-making mechanism, which aims to automatically and intelligently select the most suitable scheduling scheme from the solution set based on the real-time arrival of task types.

[0212] 4.1 Task Type Classification and Quantification

[0213] First, the tasks processed by the data center need to be categorized and their characteristics quantified. Based on their sensitivity to service quality and energy consumption, they are mainly divided into three categories:

[0214] 1) Latency-sensitive tasks: such as online transactions, real-time rendering, and interactive services. These tasks have extremely high requirements for response time and low tolerance.

[0215] 2) Computationally intensive tasks: such as scientific computing, batch data processing, model training, etc. These tasks allow for some latency, but involve large amounts of computation and may result in high energy consumption.

[0216] 3) Routine tasks: such as email service, file backup, etc. These have relatively low requirements for latency and computing resources.

[0217] To quantify task types, a task feature vector is defined, which can be represented as:

[0218] ,

[0219] In the formula: This is the latency sensitivity coefficient, with a value range of [0, 1]. The closer the value is to 1, the higher the requirement for low latency. This is used to calculate the intensity coefficient. The value ranges from [0, 1]. The closer the value is to 1, the more computing resources are consumed per unit of task.

[0220] 4.2 Evaluation and Labeling of Pareto Solution Sets

[0221] For each Pareto solution obtained (in , (The total number of solutions), which itself contains a set of decision variables and corresponding two objective function values.

[0222] Calculate a comprehensive attribute score A for each solution. k This is used to characterize the tendency of the solution to lie on the "energy consumption-response time" trade-off curve. It can be expressed as:

[0223] ,

[0224] In the formula: To solve The corresponding energy consumption cost, To solve The corresponding response time , These represent the maximum and minimum energy consumption costs in the current Pareto solution set; , These represent the maximum and minimum response times in the current Pareto solution set.

[0225] The rating is A. k The value range of A is [-1, 1]. Its significance lies in: k The closer the value is to 1, the more the solution tends to optimize response time (i.e., the response time is close to the minimum, but energy consumption may be higher), making it suitable for latency-sensitive tasks. A k The closer the value is to -1, the more the solution tends to optimize energy consumption (i.e., energy consumption is close to the minimum, but the response time may be longer), making it suitable for computationally intensive or ordinary tasks. k A value close to 0 indicates that the solution achieves a good balance between the two.

[0226] 4.3 Dynamic Matching Strategy

[0227] When a new task arrives, the system executes the following matching process:

[0228] 1) Task Feature Recognition: Determine the task feature vector by parsing task requests in real time or based on task queue information. .

[0229] 2) Matching weight calculation: Calculate the target preference weight W for this scheduling decision based on the task type.

[0230] ,

[0231] In the formula: , These are adjustable weighting coefficients used to control the impact of latency sensitivity and computational intensity on decision-making.

[0232] A larger W value indicates a greater preference for solutions with low latency; a smaller W value indicates a greater preference for solutions with low energy consumption.

[0233] 3) Compare the calculated preference weights W of the task with the attribute scores A of the Pareto solution set. k Perform a matching process. Select a solution that meets the following conditions. As the final scheduling scheme:

[0234] ,

[0235] You can then select the attribute rating A. k The Pareto solution that is closest to the current task preference weight W. To solve It belongs to the Pareto solution set. This ensures that the selected solution is compatible with the current task.

[0236] It best matches the needs.

[0237] By introducing this dynamic decision-making mechanism, this invention combines offline optimization with online decision-making, transforming data center operations into an intelligent system capable of sensing task content and dynamically adjusting resources. This not only ensures the service quality for critical tasks but also maximizes energy efficiency at the macro level, significantly enhancing commercial value and practicality.

[0238] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention. These changes involve related technologies well known to those skilled in the art, and all of them fall within the protection scope of the present invention.

[0239] Many other changes and modifications can be made without departing from the concept and scope of this invention. It should be understood that this invention is not limited to the specific embodiments, and the scope of this invention is defined by the appended claims.

Claims

1. A joint optimization method for data center energy consumption and response time based on MOFA, characterized in that, Includes the following steps: Step 1: Construct a data center energy consumption cost model and a service quality model; Step 2: Establish the joint optimization objective function and constraints; Step 3: Build an optimization model of MOFA and perform multiple position updates to obtain the Pareto solution set; Step 4: Dynamic decision-making mechanism based on task type matching.

2. The method for joint optimization of data center energy consumption and response time based on MOFA according to claim 1, characterized in that, In step 1, the energy consumption cost model of the data center is constructed. The energy consumption cost of the data center directly depends on the power consumption of the data center. It is assumed that the servers inside the data center are all the same equipment, and the power and performance of each server are exactly the same. The power consumption of the data center at a certain time period is mainly divided into three parts: the power of IT equipment, the power required for the cooling equipment to work, and other power. The power consumption of the data center during a certain period is shown below: ， In the formula: It is the total power consumed by the data center during a certain period of time. It is the power consumed by IT equipment in a data center at a certain time. This is the power consumed by the cooling equipment in the data center at a certain time period. This refers to the power consumed by other devices in the data center at a certain time. Power of IT equipment It mainly consists of server power consumption, and the power consumption of a single server is shown below: ， In the formula: This refers to the power consumption of a single server. It is the idle power consumption of the server during a certain period of time. It represents the server's full-load power consumption at a certain time. It represents the CPU utilization rate during a certain period of time; Assume the number of servers running in the data center at a certain time period is To handle data stream load, As shown below: ， In the formula: It refers to the data flow load processed by the data center at a certain time period. It is the processing speed of data streams handled by servers in a data center; Therefore, the server power consumption of the data center at a certain time period It can be represented as: ； The power consumption of data center cooling equipment during a certain period is expressed as follows: ， In the formula: This refers to the cooling capacity of a data center at a certain time period. It is the cooling energy efficiency ratio of a data center at a certain time period; The cooling capacity of the data center is shown below: ， In the formula: It is the indoor temperature. It's the outdoor temperature. It is the equivalent thermal resistance.

3. The method for joint optimization of data center energy consumption and response time based on MOFA according to claim 2, characterized in that, In step 1, the data center service quality model is constructed. The service quality of a data center is usually measured by its average service response time. The task response time is shown below: ， In the formula: It is the task response time. It is the probability of waiting; It can be calculated using Erlang-C formulas: ， In the formula: This represents the number of servers currently in operation.

4. The method for joint optimization of data center energy consumption and response time based on MOFA according to claim 3, characterized in that, In step 2, the joint optimization objective function F is established as follows: ， In the formula: It is the objective function for minimizing the energy consumption cost of data centers. It is the objective function for minimizing the data center response time; ， In the formula: It refers to the electricity price for a specific period of time; 。 5. The method for joint optimization of data center energy consumption and response time based on MOFA according to claim 4, characterized in that, The constraints in step 2 include indoor temperature constraints, workload capacity constraints, response time constraints, cooling system power constraints, and upper limit constraints on server waiting probability. Indoor temperature constraints: , In the formula: It is the minimum temperature required to maintain the normal operation of a data center; This is the maximum temperature that can sustain the normal operation of a data center; Workload capacity handling constraints: ; Response time constraints: , In the formula: This is the maximum response time that does not affect the user experience. Refrigeration system power constraints: , In the formula: This is the upper limit of the power of the data center cooling system; Upper limit constraint on server wait probability: .

6. The method for joint optimization of data center energy consumption and response time based on MOFA according to claim 1, characterized in that, Step 3 involves establishing an optimization model for MOFA and performing multiple position updates to obtain the Pareto solution set, including the following steps: Step 3.1: Randomly generate the initial population; Random population generation: Generate N fireflies, each with its own position vector. for: in, , In the formula: N is the population size, i.e., the number of fireflies, usually set to 50-100; D is the dimension of the decision variable. The j-th decision variable for the i-th firefly is randomly generated within the feasible range; Step 3.2: Calculate the fitness value of each firefly; The objective function value is to evaluate the brightness of each firefly, where brightness is determined by energy consumption cost. and response time In multi-objective optimization, brightness is defined based on the Pareto dominance relationship. Pareto Dominance Rule: Solution Dominate If and only if and And at least one inequality holds strictly, F1(X) i (to solve) The corresponding energy consumption cost, F2(X) i (to solve) The corresponding response time, F1(X) j (to solve) The corresponding energy consumption cost, F2(X) j (to solve) The corresponding response time; Step 3.3: Update the firefly positions; Brightness can be defined as: ， For each pair of fireflies and ,if Then fireflies Attracted to move; The attraction coefficient can be expressed as: ， In the formula: This is the initial attraction constant, usually set to 1.

0. The light absorption coefficient controls the rate at which the attractive force decreases with distance, and is typically set to 0.01-1.

0. firefly and The distance between them, where, ， The position update formula can be expressed as: ， In the formula: For the first Only fireflies are iterating The position at that time For the first Only fireflies are iterating The position at that time For the first Only fireflies are iterating The position at that time The random step size factor, It is a random vector whose elements follow a uniform or normal distribution; Step 3.4: Update the fitness values ​​of each firefly. After the location is updated, the objective function value and brightness of each firefly are recalculated to reflect the advantages and disadvantages of the new location; Step 3.5: Pass the non-dominated solution set to the next generation; Preserve the non-dominated solutions in the current iteration, i.e., the Pareto optimal solutions, to avoid losing high-quality solutions; Step 3.6: Determine whether the maximum number of iterations has been reached. Repeat steps 3.1 to 3.6 to obtain a Pareto solution set.

7. The method for joint optimization of data center energy consumption and response time based on MOFA according to claim 6, characterized in that, The dynamic decision-making mechanism based on task type matching in step 4 includes the following steps: Step 4.1: Task Feature Identification: Determine the task feature vector by parsing the task request in real time or based on the task queue information; Step 4.2: Matching Weight Calculation: Calculate the target preference weight W for this scheduling decision based on the task type; ， In the formula: , These are adjustable weighting coefficients used to control the influence of latency sensitivity and computational intensity on decision-making. S is the latency sensitivity coefficient, and C is the computational intensity coefficient. A larger W value indicates... The smaller the W value, the more the decision-maker prefers to choose the solution with low latency; the smaller the W value, the more the decision-maker prefers to choose the solution with low energy consumption. Step 4.3: Compare the preference weights W calculated by the task with the attribute scores A of the Pareto solution set. k conduct Matching, selecting solutions that satisfy the following conditions As the final scheduling strategy ; ， You can then select the attribute rating A. k The Pareto solution that is closest to the current task preference weight W. To solve It belongs to the Pareto solution set.

8. A data center energy consumption and response time method based on MOFA as described in claim 7. The joint optimization method is characterized by, The task feature vector in step 4.1 can be represented as: ， In the formula: This is the delay sensitivity coefficient, with a value range of [0, 1]. The closer the value is to 1, the more sensitive it is to low latency. The higher the latency requirement; The intensity coefficient is calculated with a value range of [0, 1]. The closer the value is to 1, the more computing resources are consumed per unit of task.

9. A data center energy consumption and response time method based on MOFA as described in claim 7. The joint optimization method is characterized by, In step 4.3, the attribute score of the Pareto solution set is calculated as a comprehensive attribute score A for each solution. k This is used to characterize the tendency of the solution to lie on the "energy consumption-response time" trade-off curve, and can be expressed as: ， In the formula: To solve The corresponding energy consumption cost, To solve The corresponding response time , These represent the maximum and minimum energy consumption costs in the current Pareto solution set; , These are the maximum and minimum response times in the current Pareto solution set; A k The value range of A is [-1, 1]. k The closer the value is to 1, the more the solution tends to optimize response time. Suitable for latency-sensitive tasks; A k The closer the solution is to -1, the more it tends to optimize energy consumption, making it suitable for computationally intensive or ordinary tasks; A k A value close to 0 indicates that the solution achieves a good balance between the two.

Images