A data center optimized operation method considering computing load type and user satisfaction

Abstract

The present application relates to a kind of data center optimization operation method considering computing load type and user satisfaction, method includes the following steps: constructing data center flexible workload constraint;Data center rigid workload constraint is constructed;Data center user satisfaction modeling is carried out;Based on data center flexible workload constraint, data center rigid workload constraint and satisfaction, data center and user double-layer collaborative optimization model is constructed, and the optimal timing scheduling sequence of flexible load is obtained by solving model.Compared with prior art, the present application has the advantages of effectively reducing data center operating cost under the premise of guaranteeing user satisfaction, and improving renewable energy utilization level.

Description

Technical Field

[0001] This invention relates to the field of data center optimization, and in particular to a data center optimization method that takes into account computing load type and user satisfaction. Background Technology

[0002] With the rapid development of the digital economy and the continuous expansion of cloud computing, artificial intelligence, big data analytics, and internet applications, data centers have become one of the core infrastructures supporting the operation of the information society. At the same time, the energy consumption of data centers is also showing a rapid growth trend, and their electricity demand has become a significant source of power system load growth in some regions. How to achieve efficient and low-carbon operation of data centers while ensuring the quality of computing services has become a key technical issue in the integrated development of the energy internet and information infrastructure.

[0003] In the context of new power systems, the large-scale integration of renewable energy sources such as wind and solar power has led to more significant volatility and uncertainty in the power system's supply and demand balance. Traditional power systems primarily rely on power source-side regulation to achieve power balance. However, with the increasing proportion of renewable energy, relying solely on power source-side regulation is no longer sufficient to meet the system's flexibility requirements. Therefore, exploring the adjustable potential of the load side and constructing new models that allow flexible load resources to participate in power system regulation has become an important way to improve the operational stability of the power system.

[0004] Data centers, as typical large-scale power loads, possess a degree of flexibility in time and space scheduling for their internal computing tasks, thus having the potential to become an important flexible load resource in new power systems. Through reasonable scheduling strategies, some computing tasks can be executed at different times or migrated between different data centers, enabling dynamic adjustments to data center load based on electricity price changes or renewable energy output. This "computing power-electricity" collaborative scheduling model can improve energy efficiency and promote the integration of renewable energy while ensuring computing service capabilities.

[0005] However, optimizing data center operations is not merely an energy scheduling problem; it is essentially a multi-objective collaborative optimization problem balancing computing service quality and energy costs. Different types of computing tasks vary significantly in terms of latency sensitivity, service quality requirements, and spatial migration capabilities. For example, some real-time online services are extremely sensitive to latency and must be processed immediately, while other offline data processing tasks have higher time flexibility. Furthermore, some computing tasks allow for a certain degree of reduction in computational precision or service quality to reduce resource consumption. Therefore, accurately characterizing the operational characteristics of different types of computing loads and constructing reasonable optimization scheduling strategies based on this is a crucial technical foundation for achieving efficient data center operation.

[0006] Meanwhile, with the development of cloud computing service models, the relationship between data centers and users has gradually shifted from a single service provider relationship to a multi-party collaborative relationship. Users are not only consumers of computing resources but also significant providers of computing load. When data centers adjust load through methods such as task latency, task migration, or service degradation, it inevitably impacts the user's service experience. Therefore, achieving a balance between reducing operating costs and ensuring a good user service experience during data center optimization has become a crucial technical challenge that urgently needs to be addressed in current data center operation and management.

[0007] Although a great deal of research has been conducted in recent years on the energy efficiency optimization and load regulation of data centers, from the perspective of overall technological development, existing methods still have many shortcomings in areas such as refined modeling of data center loads, quantification of user service experience, and multi-entity collaborative optimization. These shortcomings are mainly reflected in the following aspects.

[0008] First, existing research generally lacks fine-grained classification in computational load modeling. Many studies simply categorize data center computing tasks into "tunable loads" and "non-tunable loads," or classify tasks into interactive loads and batch processing loads based solely on latency sensitivity. This coarse-grained classification fails to accurately reflect the differences in time flexibility, spatial mobility, and service quality adjustability among different types of computing tasks. In reality, the adjustment methods available for different computing tasks during scheduling vary significantly. For example, some tasks only have latency capabilities, while others are better suited for migration between different data centers. Without fine-grained load modeling, it is difficult to fully explore the potential flexible adjustment capabilities within a data center, thus limiting the effectiveness of optimized scheduling strategies.

[0009] Secondly, existing optimization methods often prioritize minimizing data center operating costs, neglecting the significant impact of computing service quality on user experience. In actual operation, when data centers adjust load by delaying task execution or reducing service quality, it can lead to increased task response time or decreased service quality, adversely affecting user services. Without reasonable constraints on user experience, excessive load adjustment may increase the risk of service defaults and even cause user churn. Therefore, quantifying the impact of different adjustment behaviors on user service experience during the optimization scheduling process and incorporating it into the optimization model is a significant technical challenge in current research.

[0010] Furthermore, existing research still has shortcomings in multi-agent collaborative optimization. Traditional data center scheduling methods mostly make optimization decisions from the perspective of a single operator, meaning the data center uniformly decides the scheduling method for computing tasks. However, in real-world operating environments, users, as providers of computing tasks, have a certain degree of decision-making power and willingness to participate in task scheduling. When the data center adjusts task execution methods, users need to weigh economic incentives against service experience. Therefore, data center operation optimization is essentially a multi-agent game problem. If user autonomy is ignored, single-agent optimization methods cannot accurately reflect the behavioral characteristics of the actual operating environment and are unlikely to achieve stable scheduling results. Summary of the Invention

[0011] The purpose of this invention is to provide a data center optimization operation method that considers computing load type and user satisfaction. By constructing a refined load model and a two-layer collaborative optimization model, it achieves coordinated optimization between data center operation economy and user service experience, effectively reducing data center operating costs while ensuring user satisfaction and improving the utilization level of renewable energy.

[0012] The objective of this invention can be achieved through the following technical solutions: A data center optimization operation method that considers computing load type and user satisfaction includes the following steps: Build flexible workload constraints for data centers; Construct rigid workload constraints for the data center; Conduct data center user satisfaction modeling; Based on the constraints of flexible workload in data centers, rigid workload in data centers, and satisfaction, a two-layer collaborative optimization model between data centers and users is constructed. Solving the model yields the optimal time-series scheduling sequence for flexible workloads.

[0013] Furthermore, the data center and user two-layer collaborative optimization model includes an upper-layer model and a lower-layer model.

[0014] Furthermore, the objective function of the upper-level model is: Cgrid represents the cost of purchasing electricity from the grid for the data center; Cnet represents the network bandwidth cost incurred by the data center in migrating workloads; Cess represents the cost of using the data center's energy storage system; and Ccomp represents the incentive cost paid by the data center to lower-level users.

[0015] Furthermore, the cost in the objective function of the upper-level model is: Where fgrid,t is the electricity purchase price at time t; Pgrid,t is the power purchased by the data center from the grid at time t; fnet is the unit bandwidth equivalent load leasing price; Bnet,t is the network bandwidth equivalent load available for task migration between data centers at time t; fch and fdis are the unit depreciation and maintenance cost coefficients for charging and discharging the data center energy storage system, respectively; Pch,t and Pdis,t are the charging and discharging power of the energy storage system at time t, respectively; fadj,t and fcut,t are the unit load adjustment incentive price and unit load reduction compensation price at time t, respectively; Ladj,t is the workload adjustment amount for FW-I, FW-II, and FW-III users at time t; and Lcut,t is the workload reduction amount for FW-IV users at time t.

[0016] Furthermore, the constraints of the upper-level model include: data center power balance constraints, data center energy storage system operation constraints, and data center physical operation constraints.

[0017] Furthermore, the template function for the lower-level model is: Csta represents the equivalent economic benefit derived from user service satisfaction; Ccomp represents the incentive benefit gained from user participation in the response.

[0018] Furthermore, the constraints of the lower-level model include data center flexible workload constraints, data center rigid workload constraints, and satisfaction constraints.

[0019] Furthermore, the characteristic is that the satisfaction constraint is: Where n is the segment number; an and bn are the slope and intercept of the linear function in the nth segment. The user satisfaction level at time t represents the FW-class workload; LFW-Ⅰ,t represents the backlog of FW-Ⅰ class workloads in the data center at time t.

[0020] Furthermore, the constraints for flexible workloads in data centers are: Wherein, λt is the total load arrival rate of the data center at time t; λFW-Ⅰ,t, λFW-Ⅱ,t, λFW-Ⅲ,t, and λFW-Ⅳ,t are the arrival rates of the four types of workloads FW-Ⅰ to IV in the data center at time t; λRW-0,t is the arrival rate of the workload of this invention in the data center at time t; LFW-Ⅰ,t is the backlog of the FW-Ⅰ type of workload in the data center at time t; LFW-Ⅰ,act,t is the actual processing volume of the FW-Ⅰ type of workload at time t; μFW-Ⅰ,t is the processing capacity allocated to the FW-Ⅰ type of workload in the data center at time t; and μmax FW-Ⅰ,t is the maximum processing volume of this type of workload at time t. DFW-Ⅰ represents the maximum allowable latency for FW-Ⅰ type workloads; LFW-Ⅱ,t represents the processing volume of FW-Ⅱ type workloads in the data center at time t; LFW-Ⅱ,in,t represents FW-Ⅱ type workloads migrated into this data center from other data centers at time t; LFW-Ⅱ,out,t represents FW-Ⅱ type workloads migrated out of this data center at time t; μmax FW-Ⅱ,t represents the maximum processing volume of FW-Ⅱ type workloads at time t; Bmax net,t represents the maximum network bandwidth equivalent load available for task migration between data centers at time t; LFW-Ⅲ,t represents the backlog of FW-Ⅲ type workloads in the data center at time t; LFW-Ⅲ,in,t represents the FW-Ⅲ type workloads migrated into this data center from other data centers at time t; LFW-Ⅲ,out,t represents the FW-Ⅲ type workloads migrated out of this data center at time t; LFW-Ⅲ,act,t represents the actual processing volume of FW-Ⅲ type workloads at time t; μFW-Ⅲ,t represents the processing capacity allocated to FW-Ⅲ type workloads in the data center at time t; μmax FW-Ⅲ,t represents the maximum processing volume of this type of workload at time t; DFW-Ⅲ represents the maximum allowable latency time for FW-Ⅲ type workloads; LFW-Ⅳ,t represents the actual processing volume of FW-Ⅳ type workloads in the data center at time t; LFW-Ⅳ,cut,t represents the workload reduction due to service degradation at time t; μmax FW-Ⅳ,t represents the maximum processing volume of FW-Ⅳ type workloads at time t; βmax FW-Ⅳ is the maximum allowable reduction ratio for FW-Ⅳ type loads.

[0021] Furthermore, the rigid workload constraint for data centers is: Wherein, LRW-0,t is the actual processing volume of RW-0 type workloads that the data center must handle at time t; μmaxRW-0,t is the maximum processing volume of RW-0 type workloads at time t.

[0022] Compared with the prior art, the present invention has the following beneficial effects: (1) A multi-dimensional feature-driven method for fine-grained classification of data center computing loads is proposed. This invention comprehensively considers the differences in computing tasks across three dimensions: time flexibility, spatial mobility, and service quality adjustability, and then finely classifies data center computing loads. Flexible computing tasks are further divided into time-shiftable loads, spatially migrated loads, spatiotemporally adjustable loads, and quality-reducible loads, which, together with rigid loads, constitute a complete load classification system. By establishing differentiated operational constraint models for different types of loads, the operational characteristics of data center computing tasks can be described more accurately, providing a more precise model basis for subsequent optimization scheduling.

[0023] (2) Constructing a quantitative evaluation model for user satisfaction regarding computing service experience. To address the problem that traditional optimization methods neglect user service experience, this invention introduces a user satisfaction index to quantitatively characterize user experience from multiple dimensions, such as task latency, cross-data center migration, and service quality degradation. Differentiated satisfaction functions are constructed for different types of computing loads, and a comprehensive user satisfaction evaluation index is established to achieve a quantitative assessment of data center service quality. This method can effectively constrain over-adjustment behavior during the optimization scheduling process, thereby ensuring user service experience while tapping the potential for data center load adjustment.

[0024] (3) A two-layer collaborative optimization operation model between data centers and users is proposed. Addressing the multi-agent decision-making problem between data centers and users, this invention constructs a two-layer optimization model. The upper-layer model takes the data center operator as the decision-making entity, aiming to minimize operating costs, and guides users to participate in load adjustment by designing load adjustment incentive strategies; the lower-layer model takes the user as the decision-making entity, aiming to maximize overall utility, and weighs economic benefits against service experience. Through the coupled solution of the two-layer model, collaborative optimization decision-making between the data center and users is achieved, thereby effectively reducing data center operating costs and improving the utilization level of renewable energy while ensuring user satisfaction. Attached Figure Description

[0025] Figure 1 Classify data center workloads; Figure 2 Optimized for collaboration between the data center and users at both levels; Figure 3 Electricity price curves for each data center; Figure 4 For flexible workload tasks in data centers; Figure 5 Incentive costs paid for data centers in different scenarios; Figure 6 To improve user satisfaction in different scenarios; Figure 7 Contribute to renewable energy in data center areas; Figure 8 This refers to the utilization rate of renewable energy. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0027] This invention proposes a data center optimization operation method that considers both computational load type and user satisfaction. Addressing existing problems in data center operation optimization, such as insufficiently refined computational load modeling, lack of quantitative description of user service experience, and the difficulty of balancing operational economy and service quality with single-entity optimization, this invention first refines the classification of data center computational load based on differences in time flexibility, spatial migration capability, and service quality adjustability. This classification categorizes loads into time-shiftable, spatially migrated, spatiotemporally adjustable, quality-reducible, and rigid loads, and establishes corresponding operational constraint models. Second, it constructs a user satisfaction evaluation model oriented towards computational service experience, quantitatively characterizing user experience from multiple dimensions such as task latency, cross-data center migration, and service quality reduction. Furthermore, it establishes a two-layer collaborative optimization model between the data center and users. The upper layer aims to minimize data center operating costs by incentivizing user participation in load adjustment; the lower layer aims to maximize overall user utility by balancing economic incentives and service experience. By solving the two-layer optimization model, collaborative optimization of computational load scheduling and energy utilization is achieved. This invention can effectively reduce data center operating costs and improve renewable energy utilization while ensuring user service quality. The output of the upper-level model includes one or more of the following: grid power purchase, wind and solar local consumption power, energy storage charging and discharging power, energy storage time-series state of charge, energy storage charging and discharging 0-1 state variables, unit load adjustment incentive price for each time period, and bandwidth occupied by task migration.

[0028] In the upper-level model of this invention, the data center acts as the leader, aiming to minimize total operating costs. By providing appropriate incentives to its users, the data center guides them to proactively respond and adjust a certain proportion of their workload. This flexible workload adjustment provides the data center with greater resource scheduling flexibility, reducing its own operating costs.

[0029] (1) Objective function Data center operators aim to minimize total operating costs, and their objective function is expressed as: (twenty four) In the formula: Cgrid is the cost of purchasing electricity from the grid for the data center; Cnet is the network bandwidth cost incurred by the data center in migrating workloads; Cess is the cost of using the data center's energy storage system; Ccomp is the incentive cost paid by the data center to lower-level users, including load adjustment incentive cost and load reduction incentive cost.

[0030] The specific costs are as follows: (25) (26) (27) (28) In the formula: fgrid,t is the electricity purchase price at time t; Pgrid,t is the power purchased by the data center from the grid at time t; fnet is the unit bandwidth equivalent load leasing price; Bnet,t is the network bandwidth equivalent load available for task migration between data centers at time t; fch and fdis are the unit depreciation and maintenance cost coefficients for charging and discharging the data center energy storage system, respectively; Pch,t and Pdis,t are the charging and discharging power of the energy storage system at time t, respectively; fadj,t and fcut,t are the unit load adjustment incentive price and unit load reduction compensation price at time t, respectively; Ladj,t is the workload adjustment amount for FW-I, FW-II, and FW-III users at time t; Lcut,t is the workload reduction amount for FW-IV users at time t.

[0031] (2) Constraints Data center power balance constraints: Data center power consumption is comprised of both IT load and infrastructure power consumption (such as cooling and power distribution systems). This paper introduces power usage effectiveness (PUE) to provide a unified characterization of the relationship between data center IT load and infrastructure load. Therefore, the power of the data center at time t can be expressed as: (29) (30) In the formula: Pidc,t is the total power of the data center at time t; Pit,t is the power of IT equipment at time t; PUE reflects the overall energy efficiency level of the data center; Pwt,t and Ppv,t are the power of wind turbines and photovoltaic power generation at time t, respectively.

[0032] (31) In the formula: Pmin grid,t and Pmax grid,t are the minimum and maximum power flowing from the grid into the data center, respectively; Pmax wt,t is the upper limit of wind turbine power generation; and Pmax pv,t is the upper limit of photovoltaic power generation.

[0033] Data center energy storage system operating constraints: Constraints on energy storage systems mainly include power balance constraints and upper and lower limits constraints on energy storage capacity and power.

[0034] (32) In the formula: Eess,t and Eess,t-1 are the states of charge of the energy storage system at time t and time t-1, respectively; ηch and ηdis are the charging efficiency and discharging efficiency of the energy storage system, respectively; Emax ess,t and Emix ess,t are the upper and lower limits of the energy storage system capacity, respectively; Pmax ch and Pmax dis are the maximum charging and discharging power of the system, respectively; Nch,t and Ndis,t are binary state variables (0 or 1) used to limit the energy storage from charging and discharging at the same time.

[0035] Data center physical operational constraints: The total computing service capacity of a data center depends on the number of servers currently in operation. To ensure service stability and hardware security, and to prevent server overload leading to downtime or high latency, the number of servers in operation and resource utilization must meet the following requirements: (33) In the formula: Mon,t is the number of servers that are on at time t; Mser is the total number of servers configured in the data center; Msta,t and Mstop,t are the number of servers that are started and stopped at time t, respectively. To avoid the loss of lifespan caused by frequent hardware start-ups and shutdowns, |Mon,t-Mon,t-1| needs to be limited; δt is the server resource occupancy rate at time t; δmin and δmax are the minimum resource occupancy rate and the maximum safe resource occupancy rate, respectively.

[0036] In the lower-level model, users act as followers, aiming to maximize overall utility, including service experience and economic benefits. Guided by incentives set by the data center, users autonomously decide on the proportion of workload they participate in, based on their workload flexibility and service needs, and a comprehensive balance between the economic compensation and service reliability brought by the workload. Through reasonable adjustment, users can maintain the stability of their service experience while receiving appropriate incentive compensation.

[0037] (1) Objective function With the goal of maximizing the overall utility of data center users, the objective function is expressed as: (34) In the formula: Csta is the equivalent economic benefit calculated based on user service satisfaction; Ccomp is the incentive benefit obtained by users participating in the response.

[0038] To standardize and quantify service satisfaction and incentive benefits, a utility conversion coefficient is introduced to transform the dimensionless satisfaction index into economic value. Therefore, Csta can be expressed as: (35) In the formula: χ is the satisfaction-utility conversion coefficient (yuan / unit satisfaction).

[0039] (2) Constraints: The lower-level constraints are mainly composed of the operating characteristics of various data center loads. As an emerging high-energy-consuming and strongly coupled load, the power consumption behavior of data centers exhibits a certain degree of flexibility in both time and space dimensions. The energy consumption level and adjustability potential of data centers depend not only on the physical constraints of the infrastructure, but also to a large extent on the operating characteristics of the internal workloads themselves. Different types of workloads have significant differences in latency sensitivity, portability, and scalability. Accurately characterizing the behavioral characteristics of different workloads has become a key foundation for achieving efficient resource scheduling and ensuring service quality. In existing research, workloads are usually divided into interactive workloads and batch workloads based on latency sensitivity. Interactive workloads emphasize low latency and high service quality requirements, while batch workloads have strong spatiotemporal transfer characteristics and can be executed within a specified time, depending on the actual situation. Based on this, this paper divides data center workloads into flexible workloads (FW) and rigid workloads (RW) of this invention. Further considering time tolerance, spatial elasticity, and quality reduction characteristics, flexible workloads are divided into time-shiftable workloads, spatially portable workloads, spatiotemporally adjustable workloads, and quality-reducible workloads. For a detailed classification and characteristics of various workloads, please refer to [link / reference]. Figure 1 See Table 1.

[0040] Table 1 Data Center Workload Characteristics As shown in Table 1, Time-Movable Load (FW-I) refers to computing tasks that are limited by requirements such as local data storage and cannot be migrated across regions, but have flexibility in completion time. They allow for delayed execution within a given time window, but must be completed before the deadline. Equation (1) represents the arrival rate of various types of loads in the data center.

[0041] (1) In the formula: λt is the total load arrival rate of the data center at time t; λFW-Ⅰ,t, λFW-Ⅱ,t, λFW-Ⅲ,t, and λFW-Ⅳ,t are the arrival rates of the four types of workloads FW-Ⅰ to Ⅳ in the data center at time t; λRW-0,t is the arrival rate of the workload of this invention in the data center at time t.

[0042] (2) In the formula: LFW-Ⅰ,t represents the backlog of FW-Ⅰ type workloads in the data center at time t; LFW-Ⅰ,act,t represents the actual processing volume of FW-Ⅰ type workloads at time t.

[0043] (3) (4) (5) In the formula: μFW-Ⅰ,t represents the processing capacity allocated to the data center for FW-Ⅰ type load at time t; μmax FW-Ⅰ,t represents the maximum processing capacity of this type of load at time t; and DFW-Ⅰ represents the maximum allowable latency time for FW-Ⅰ type load.

[0044] Spatially portable workloads (FW-II) refer to computing tasks that have high processing time requirements, typically lack the ability to be moved in the time dimension, but are not sensitive to geographical location. Relying on high-speed fiber optic networks between data centers, this type of workload can be migrated and executed between different data centers.

[0045] (6) In the formula: LFW-Ⅱ,t represents the processing volume of FW-Ⅱ type workloads in the data center at time t; LFW-Ⅱ,in,t represents the FW-Ⅱ type workloads that migrated from other data centers to this data center at time t; LFW-Ⅱ,out,t represents the FW-Ⅱ type workloads that migrated out of this data center at time t.

[0046] (7) (8) In the formula: μmax FW-Ⅱ,t is the maximum processing capacity of FW-Ⅱ type load at time t; Bmax net,t is the maximum network bandwidth equivalent load available for task migration between data centers at time t.

[0047] Spatiotemporally adjustable workload (FW-Ⅲ) refers to a computing task that has both time-shifting and spatial migration capabilities, and its resource consumption can be dynamically adjusted in both spatiotemporal dimensions according to changes in workload.

[0048] (9) In the formula: LFW-Ⅲ,t represents the backlog of FW-Ⅲ type workloads in the data center at time t; LFW-Ⅲ,in,t represents the FW-Ⅲ type workloads migrated into this data center from other data centers at time t; LFW-Ⅲ,out,t represents the FW-Ⅲ type workloads migrated out of this data center at time t; and LFW-Ⅲ,act,t represents the actual processing volume of FW-Ⅲ type workloads at time t.

[0049] (10) (11) (12) In the formula: μFW-Ⅲ,t represents the processing capacity of the data center at time t for FW-Ⅲ type load; μmax FW-Ⅲ,t represents the maximum processing capacity of this type of load at time t; and DFW-Ⅲ represents the maximum allowable latency time for FW-Ⅲ type load.

[0050] Quality-Reducible Load (FW-Ⅳ) refers to computing tasks that data center users are willing to sacrifice some service quality or computing accuracy under certain incentive mechanisms. These tasks can reduce the current resource usage by degrading the service.

[0051] (13) In the formula: LFW-Ⅳ,t is the actual processing volume of the data center's FW-Ⅳ type workload at time t; LFW-Ⅳ,cut,t is the workload reduction due to service degradation at time t.

[0052] For quality-reducible loads, the actual load processing volume must not exceed the maximum processing capacity allocated to that type of load, and the reduced load must be controlled within the minimum service quality allowed by the user.

[0053] (14) (15) In the formula: μmax FW-Ⅳ,t is the maximum processing capacity of FW-Ⅳ type load at time t; βmax FW-Ⅳ is the maximum allowable reduction ratio coefficient of FW-Ⅳ type load, which reflects the user's maximum tolerance for service quality degradation.

[0054] Data center rigid workloads: The workload of this invention refers to (RW-0) computing tasks that have high real-time requirements, are constrained by geographical location or data privacy, cannot be postponed, cannot be migrated across data centers, and must be executed in full with no reduction in service quality. Since the workload of this invention has no adjustment capabilities, its actual processing volume at time t is equal to the load arrival volume at that time.

[0055] (16) (17) Where: LRW-0,t is the actual processing volume of RW-0 type workloads that the data center must handle at time t; μmaxRW-0,t is the maximum processing volume of RW-0 type workloads at time t; Based on the refined definition of workload characteristics described above, different types of workloads exhibit varying degrees of flexibility in terms of timing, spatial distribution, and quality reduction, providing data centers with diverse control methods for demand response. In response to the diversity of user computing service demands and to fully tap the response potential of data center users, data centers implement corresponding incentive subsidies to guide users to adjust different types of computing workloads according to their own business characteristics and service quality requirements. While users receive economic incentives, data centers can reduce operating costs and improve energy efficiency, achieving synergistic optimization between data centers and users.

[0056] Data center user satisfaction modeling: While adjusting data center workloads to reduce operating costs, it inevitably impacts user experience. With the evolution of data center service models, continuously meeting diverse and multi-dimensional user experience needs has become a key issue in optimizing data center operations. To accurately reflect the impact of adjusting different types of workloads on user experience, this paper introduces user satisfaction (US) as a core dimension for evaluating user service quality. By quantitatively characterizing the experience loss during the adjustment process for different types of workloads, it effectively avoids service quality degradation caused by over-adjustment. Uk,t is defined as the user satisfaction function for the k-th type of workload at time t, with a value range of [0,1], where 1 represents complete satisfaction (no adjustment impact) and 0 represents complete dissatisfaction (exceeding the tolerance limit).

[0057] For FW-I type workloads, as the backlog of LFW-I,t increases, task queuing time increases, and user satisfaction exhibits a non-linear decreasing trend. An exponential function is used to characterize this time-latency characteristic: (18) In the formula: UFW-Ⅰ,t is the user satisfaction at time t for workload type FW-Ⅰ; Lth FW-Ⅰ,t is the user's psychological threshold for the backlog of this type of workload; αFW-Ⅰ,t is the time sensitivity factor; θ1 is the satisfaction decay index (θ1>1), which is used to reflect the user's strong sensitivity to long latency.

[0058] For FW-II workloads, the decline in user experience mainly stems from network latency and data transmission risks caused by cross-data center transfers. When migrating large-scale computing tasks, the risk of network congestion increases, which can affect service stability.

[0059] (19) In the formula: UFW-Ⅱ,t represents the user satisfaction at time t for FW-Ⅱ type workload; αFW-Ⅱ,t represents the cross-data center migration impact coefficient; the ratio of LFW-Ⅱ,out,t to Bmax net,t reflects the network bandwidth occupancy rate. The higher the occupancy rate, the greater the risk of network congestion and the worse the user experience.

[0060] For FW-III type workloads, which possess both time-shifting and spatial migration capabilities, user experience is jointly affected by task latency and migration behavior. To reflect the spatiotemporal coupling characteristics, a weighted geometric average is used to construct a satisfaction model. This model shows that prolonged workload backlog or large-scale migration leads to a decrease in overall user satisfaction.

[0061] (20) In the formula: UFW-Ⅲ,t represents the user satisfaction with FW-Ⅲ type workload at time t; Lth FW-Ⅲ represents the user's psychological threshold for the backlog of FW-Ⅲ type workload; αtime FW-Ⅲ and αspace FW-Ⅲ are the sensitivity weight coefficients of this type of workload in the time dimension and space dimension, respectively.

[0062] For FW-IV type workloads, user satisfaction depends on the proportion of workload reduction due to decreased service quality. The greater the reduction, the greater the corresponding decline in service quality. Originally, a satisfaction model based on the reduction ratio was used. (twenty one) In the formula: UFW-Ⅳ,t is the user satisfaction at time t for FW-Ⅳ type workload; αFW-Ⅳ is the quality sensitivity coefficient; θ2 is the quality attenuation coefficient.

[0063] For RW-0 workloads, due to their extremely high requirements for real-time performance and reliability, unlike other adjustable flexible workloads, there is no room for "degradation" in the quality of service (QoS) of RW-0 workloads. Their QoS must always be guaranteed, therefore, their user satisfaction URW-0,t is set to a constant value: (twenty two) To evaluate the service quality of data centers for various types of workloads, and taking into account the importance of different types of workloads in their business, a unified weighted overall user satisfaction index (USt) is constructed to quantitatively characterize the overall service level of the data center.

[0064] (twenty three) In the formula: κ={FW-Ⅰ,FW-Ⅱ,FW-Ⅲ,FW-Ⅳ,RW-0}; ωk is the priority weight coefficient of the k-th type of workload; λk,t is the arrival rate of the k-th type of workload at time t; Uk,t is the user satisfaction of the k-th type of workload.

[0065] This invention constructs a two-layer collaborative optimization model between a data center and its users, which is a typical master-slave game problem. The upper-layer model, with the data center as the leader, guides user responses through incentives; the lower-layer model, with end users as followers, autonomously decides on adjustments and reductions of various types of flexible workloads under incentive guidance to maximize its overall utility. Since the user satisfaction function introduced in the lower-layer model exhibits nonlinear characteristics, and the two-layer solution model itself is nonconvex, to ensure solution efficiency and global optimality, this paper employs piecewise linearization combined with Karush-Kuhn-Tucker (KKT) conditions to equivalently transform the two-layer model into a single-layer mixed-integer linear programming problem for solution.

[0066] In the lower-level model, the satisfaction function UFW-Ⅰ,t for FW-Ⅰ type workloads exhibits exponential decay. To satisfy the convexity requirement of the KKT conditions, a piecewise linearization method is used to transform it into a linear constraint: (36) In the formula: n is the segment number; an and bn are the slope and intercept of the nth linear function. Similarly, other nonlinear satisfaction functions can be processed.

[0067] After the linearization process described above, the utility maximization problem for lower-level users is equivalently transformed into a continuous convex programming problem. By constructing the Lagrangian function of the lower-level model, the original two-level master-slave game model is transformed into a single-level mathematical programming problem with equilibrium constraints. For the complementary relaxation equations with nonlinear characteristics in the KKT conditions, auxiliary Boolean variables are introduced and the Big-M method is used for exact linearization. Finally, the original two-level problem is reconstructed into a mixed-integer linear programming model, which can be directly solved efficiently and centrally using commercial solvers.

[0068] To verify the effectiveness of the proposed model and algorithm, a simulation analysis was conducted using a real-world data center in a certain location. The scheduling cycle was 24 hours, and the time step was 1 hour. Three data centers located in different geographical locations were used in the simulation. The electricity prices in the areas where the three data centers were located were different, as shown in Figure 3. The servers and other equipment in each data center had the same hardware parameters, and the impact of different equipment performance was not considered. The overall energy efficiency (PUE) level of each data center was 1.4. Assuming that the flexible workload and the workload of this invention accounted for 70% and 30% of the total workload in the data center, respectively, the initial arrival time of the flexible workload before optimization is shown in Figure 4. DFW-Ⅰ and DFW-Ⅲ were taken as 6 hours and 8 hours, respectively, and the latency of other types of workload was taken as 0.3 seconds.

[0069] To verify the effectiveness of the proposed two-layer collaborative optimization strategy, the following three comparison scenarios were designed: Scenario 1: The data center does not implement any incentive mechanism, and all workloads arriving at the data center are treated as the workloads of this invention and processed immediately.

[0070] Scenario 2: Considering only the operating costs of the data center, the data center performs time and space transfer or quality reduction on various flexible workloads submitted by users to minimize its own operating costs, ignoring the satisfaction of lower-level users.

[0071] Scenario 3: The model proposed in this invention takes into account both data center operating costs and user service quality.

[0072] Table 2 shows the total operating costs of the data centers under different optimization scenarios. Overall, the total operating costs of the three data centers are highest in Scenario 1 and lowest in Scenario 3. In Scenario 1, because the data center treats all computing tasks as the load of this invention, it cannot adjust the computing tasks according to changes in electricity prices or the characteristics of renewable energy output, making it difficult to utilize off-peak electricity periods or peak periods for renewable energy generation. Therefore, the total operating cost of the data center is highest in this scenario. In Scenario 2, the data center aims to minimize operating costs and adjusts the flexible load to the maximum extent, ignoring the end-user experience boundaries. In this case, the operating cost of the data center decreases significantly. The results show that compared to Scenario 1, the total operating costs of the three data centers decrease by RMB 231,000, RMB 245,000, and RMB 104,000, respectively. In Scenario 3, as presented in this invention, compared to Scenario 2, the total operating cost of the data center increases slightly due to the introduction of user satisfaction constraints and the consideration of a reasonable incentive compensation mechanism. However, compared to Scenario 1, this scenario still demonstrates a certain degree of economic efficiency, with the total operating costs of the three data centers decreasing by approximately 29.3%, 22.7%, and 11.8%, respectively, compared to Scenario 1. It is evident that the two-tier collaborative optimization model constructed in this invention, which considers both data centers and users, effectively reduces data center operating costs by rationally guiding the adjustment of various flexible workloads, while avoiding excessive adjustments that could negatively impact user service experience.

[0073] Table 2 Total operating costs of data centers under different scenarios Figure 5 illustrates the differences in incentive costs paid by three data centers to downstream users under different scenarios. Incentive costs reflect the economic costs incurred by data centers to guide flexible workloads to participate in responses. In Scenario 1, the data center implements no incentive mechanism; all tasks are processed on demand and immediately, resulting in no incentive costs. In Scenario 2, the data center's incentive costs are higher because, in this scenario, the data center, in pursuit of minimizing operating costs, uses incentives to guide a large number of flexible workloads to participate in adjustments, achieving time-space migration or service quality reduction. In Scenario 3, the data center, based on users' sensitivity to task migration and quality reduction, sets reasonable unit load adjustment incentive prices and unit load reduction compensation prices. A game equilibrium is achieved between the data center and users, stimulating users' willingness to respond while avoiding excessive incentives, thus keeping incentive costs within an economically reasonable range.

[0074] User satisfaction is a crucial indicator of data center service quality; a lower value indicates a more severe impairment in user experience. Figure 6 compares user satisfaction across three data centers in different scenarios. In Scenario 1, the data center does not intervene with time-shifting or quality reduction for flexible workloads. After a user submits a computing task, the data center processes it immediately. Therefore, in this scenario, user satisfaction remains high across all three data centers, generally stable above 0.90, indicating a relatively high level of service experience. However, as discussed in the previous analysis of data center operating costs, this near-optimal service quality comes at the cost of high data center operating costs, making the overall operating model uneconomical. In Scenario 2, user satisfaction drops significantly, with all three data centers falling to a low level of 0.55-0.65. This is primarily due to the data center's unilateral pursuit of minimizing operating costs and over-utilization of the adjustment potential of flexible workloads. For example, a large amount of time-shiftable workloads (FW-Ⅰ) are accumulated for extended periods, exceeding users' psychological thresholds; simultaneously, quality-reducible workloads (FW-Ⅳ) are consistently degraded, operating at a lower service quality level. While this scenario achieves lower-cost data center operation, it comes at the cost of significantly reduced user experience, potentially leading to increased service default risk and user churn. In contrast, Scenario 3 proposed in this invention achieves a more reasonable balance between data center operational economics and user experience. Under a two-tier collaborative optimization framework involving both data centers and users, user satisfaction in all three data centers remains within a reasonable range of 0.75 to 0.85. Compared to Scenario 2, Scenario 3 demonstrates a significant improvement in user experience. The fundamental reason lies in the fact that, in this model, lower-level users, as rational followers, can comprehensively weigh the incentive benefits provided by the data center against the loss of service experience based on their own sensitivity to task latency and quality degradation. When the load backlog or quality reduction approaches the user's tolerance limit, the user no longer allows flexible workloads to participate in adjustment.

[0075] Based on the above analysis of data center operation economy and user satisfaction, it can be seen that the data center and user two-layer collaborative optimization model constructed in this invention effectively alleviates the contradiction between data center operation cost and user experience. At a reasonable cost slightly higher than that of scenario 2, it effectively taps into the flexibility of data center operation while avoiding service experience deterioration due to over-adjustment, thus achieving a win-win situation for data center operation economy and user service experience.

[0076] Users, as lower-level followers, gain incentive benefits by adjusting their flexible workloads. Similarly, due to the significant randomness and volatility of renewable energy sources such as wind and solar power, their output often fails to naturally match traditional load demands over time. Data centers, as typical high-energy-consuming loads, possess a degree of spatiotemporal flexibility and adjustable service quality in their flexible workloads. On the user side, adjustments to some delayable or elastically processable computing tasks allow for task execution to be shiftable and spatially mobile within a certain time window, thus forming an adjustable computing load. Data centers, in turn, optimize the execution sequence of workloads based on changes in renewable energy output, achieving a coordinated match between computing load and renewable energy generation curves over time by appropriately delaying or advancing the processing time of some tasks. When renewable energy output is high, data centers can increase the intensity of computing tasks, improve server resource utilization, and thus enhance the local consumption level of renewable energy. When renewable energy output is insufficient, the processing priority of some non-critical tasks is reduced or their execution is delayed to improve the data center's adaptability to renewable energy fluctuations. The renewable energy output situation in the data center area is shown in Figure 7.

[0077] Figure 8 illustrates the renewable energy utilization of data centers under different scenarios, showing variations in renewable energy absorption levels. Taking Data Center 1 as an example, in Scenario 1, the overall renewable energy absorption level of the data center is low and exhibits significant periodic fluctuations. Since all workloads lack adjustable capacity, the data center load cannot be dynamically adjusted based on renewable energy output, resulting in ineffective utilization of renewable energy and demonstrating that under the traditional load operation mode of this invention, the data center's capacity for green energy absorption is limited. In Scenario 2, the data center significantly improves renewable energy absorption capacity by centrally scheduling flexible loads to ensure that computing tasks are highly matched to the characteristics of renewable energy output in terms of time. However, optimization in this scenario primarily relies on the data center's forced scheduling of user-submitted computing tasks, and the improvement in absorption level comes at the cost of sacrificing user service experience and satisfaction. In Scenario 3, the data center's renewable energy absorption level is high. Through a two-layer game mechanism between the data center and users, the data center guides rational users to autonomously adjust their load behavior by issuing incentives. Specifically, time-shiftable loads are transferred to periods with higher renewable energy output, while spatially movable loads are transferred and scheduled between different data centers. The results show that, compared with scenario 1, the data center absorption capacity in scenario 3 is significantly improved, basically reaching the level of scenario 2.

[0078] A comprehensive comparison of the three scenarios reveals that the proposed two-layer collaborative optimization strategy has significant advantages, achieving renewable energy consumption effects close to those in scenario 2, even without considering user experience. This strategy guides users to participate in load adjustments through a price incentive mechanism, further matching data center load and renewable energy output while ensuring user satisfaction. By fully leveraging the spatiotemporal flexibility of computing load, it effectively mitigates the randomness and volatility of renewable energy output.

[0079] To address the issues of flexible adjustment potential of data center computing load and ensuring user service experience, this paper proposes a data center optimization operation method that considers computing load type and user satisfaction. The effectiveness of the model is verified through numerical examples. The main work and conclusions are as follows: (1) The characteristics of data center workloads are described in detail. Based on the adjustability of tasks in the time dimension, space dimension and service quality dimension, data center workloads are divided into five categories: time-shiftable workloads, space-migratable workloads, time-space adjustable workloads, quality-reducible workloads and the workload of this invention.

[0080] (2) A differentiated user satisfaction evaluation model was constructed. To quantitatively characterize the impact of different load adjustment behaviors on user service experience, a user satisfaction index was introduced, and corresponding satisfaction function models were constructed for different types of workloads. By considering factors such as task backlog, cross-data center migration scale, and service quality reduction ratio, a comprehensive user satisfaction evaluation index was established, realizing the quantitative assessment of data center service quality.

[0081] (3) A two-layer collaborative optimization model between data centers and users is proposed. The upper-layer model takes data center operators as the decision-making body, aims to minimize operating costs, and guides users to participate in load regulation by setting reasonable incentive mechanisms. The lower-layer model takes data center users as the decision-making body, aims to maximize comprehensive utility, and balances economic incentives and service experience. Finally, the numerical analysis results show that the proposed two-layer collaborative optimization strategy can effectively reduce data center operating costs and significantly improve renewable energy utilization while ensuring that user satisfaction is at a reasonable level. This strategy guides users to actively participate in load regulation through incentive mechanisms, fully explores the spatiotemporal flexibility of computing tasks, and achieves collaborative optimization between data center operating economy, user service experience, and renewable energy absorption capacity.

[0082] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A data center optimization operation method considering computing load type and user satisfaction, characterized in that, The method includes the following steps: Build flexible workload constraints for data centers; Construct rigid workload constraints for the data center; Conduct data center user satisfaction modeling; Based on the constraints of flexible workload in data centers, rigid workload in data centers, and satisfaction, a two-layer collaborative optimization model between data centers and users is constructed. Solving the model yields the optimal time-series scheduling sequence for flexible workloads.

2. The data center optimization operation method considering computing load type and user satisfaction as described in claim 1, characterized in that, The data center and user two-layer collaborative optimization model includes an upper-layer model and a lower-layer model.

3. A data center optimization operation method considering computing load type and user satisfaction as described in claim 2, characterized in that, The objective function of the upper-level model is: Cgrid represents the cost of purchasing electricity from the grid for the data center; Cnet represents the network bandwidth cost incurred by the data center in migrating workloads; Cess represents the cost of using the data center's energy storage system; and Ccomp represents the incentive cost paid by the data center to lower-level users.

4. A data center optimization operation method considering computing load type and user satisfaction as described in claim 3, characterized in that, The cost in the objective function of the upper-level model is: Where fgrid,t is the electricity purchase price at time t; Pgrid,t is the power purchased by the data center from the grid at time t; fnet is the unit bandwidth equivalent load leasing price; Bnet,t is the network bandwidth equivalent load available for task migration between data centers at time t; fch and fdis are the unit depreciation and maintenance cost coefficients for charging and discharging the data center energy storage system, respectively; Pch,t and Pdis,t are the charging and discharging power of the energy storage system at time t, respectively; fadj,t and fcut,t are the unit load adjustment incentive price and unit load reduction compensation price at time t, respectively; Ladj,t is the workload adjustment amount for FW-I, FW-II, and FW-III users at time t; and Lcut,t is the workload reduction amount for FW-IV users at time t.

5. A data center optimization operation method considering computing load type and user satisfaction as described in claim 4, characterized in that, The constraints of the upper-level model include: data center power balance constraints, data center energy storage system operation constraints, and data center physical operation constraints.

6. A data center optimization operation method considering computing load type and user satisfaction as described in claim 5, characterized in that, The template function for the lower-level model is: Csta represents the equivalent economic benefit derived from user service satisfaction; Ccomp represents the incentive benefit gained from user participation in the response.

7. A data center optimization operation method considering computing load type and user satisfaction as described in claim 6, characterized in that, The constraints of the lower-level model include data center flexible workload constraints, data center rigid workload constraints, and satisfaction constraints.

8. A data center optimization operation method considering computing load type and user satisfaction according to claim 7, characterized in that, The satisfaction constraint is: Where n is the segment number; an and bn are the slope and intercept of the linear function in the nth segment. The user satisfaction level at time t represents the FW-class workload; LFW-Ⅰ,t represents the backlog of FW-Ⅰ class workloads in the data center at time t.

9. A data center optimization operation method considering computing load type and user satisfaction as described in claim 8, characterized in that, The constraints for flexible workloads in data centers are: Wherein, λt is the total load arrival rate of the data center at time t; λFW-Ⅰ,t, λFW-Ⅱ,t, λFW-Ⅲ,t, and λFW-Ⅳ,t are the arrival rates of the four types of workloads FW-Ⅰ to IV in the data center at time t; λRW-0,t is the arrival rate of the workload of this invention in the data center at time t; LFW-Ⅰ,t is the backlog of the FW-Ⅰ type of workload in the data center at time t; LFW-Ⅰ,act,t is the actual processing volume of the FW-Ⅰ type of workload at time t; μFW-Ⅰ,t is the processing capacity allocated to the FW-Ⅰ type of workload in the data center at time t; and μmax FW-Ⅰ,t is the maximum processing volume of this type of workload at time t. DFW-Ⅰ represents the maximum allowable latency for FW-Ⅰ type workloads; LFW-Ⅱ,t represents the processing volume of FW-Ⅱ type workloads in the data center at time t; LFW-Ⅱ,in,t represents FW-Ⅱ type workloads migrated into this data center from other data centers at time t; LFW-Ⅱ,out,t represents FW-Ⅱ type workloads migrated out of this data center at time t; μmax FW-Ⅱ,t represents the maximum processing volume of FW-Ⅱ type workloads at time t; Bmax net,t represents the maximum network bandwidth equivalent load available for task migration between data centers at time t; LFW-Ⅲ,t represents the backlog of FW-Ⅲ type workloads in the data center at time t; LFW-Ⅲ,in,t represents the FW-Ⅲ type workloads migrated into this data center from other data centers at time t; LFW-Ⅲ,out,t represents the FW-Ⅲ type workloads migrated out of this data center at time t; LFW-Ⅲ,act,t represents the actual processing volume of FW-Ⅲ type workloads at time t; μFW-Ⅲ,t represents the processing capacity allocated to FW-Ⅲ type workloads in the data center at time t; μmax FW-Ⅲ,t represents the maximum processing volume of this type of workload at time t; DFW-Ⅲ represents the maximum allowable latency time for FW-Ⅲ type workloads; LFW-Ⅳ,t represents the actual processing volume of FW-Ⅳ type workloads in the data center at time t; LFW-Ⅳ,cut,t represents the workload reduction due to service degradation at time t; μmax FW-Ⅳ,t represents the maximum processing volume of FW-Ⅳ type workloads at time t; βmax FW-Ⅳ is the maximum allowable reduction ratio for FW-Ⅳ type loads.

10. A data center optimization operation method considering computing load type and user satisfaction as described in claim 9, characterized in that, The rigid workload constraints of the data center are: Wherein, LRW-0,t is the actual processing volume of RW-0 type workloads that the data center must handle at time t; μmax RW-0,t is the maximum processing volume of RW-0 type workloads at time t.

Images