Data center task scheduling modeling method based on multi-dimensional identification and mean field theory

Abstract

The invention belongs to the technical field of data center resource management under an identification network architecture, and discloses a data center task scheduling modeling method based on a multi-dimensional identification and mean field theory. The method comprises the steps of setting a load index of a server, then confirming a queuing strategy according to the set load index and returning a multi-dimensional identifier of the server, and finally carrying out modeling on the system by utilizing an average field queuing theory and predicting performance indexes under different resource configurations so as to maximize the overall performance and stability of the system. Optimization of comprehensive performance is achieved, real-time scheduling and configuration optimization are combined to form a closed loop, the data center can automatically adapt to load changes, and elastic optimization configuration of resources is achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data center resource management under the identification network architecture, and in particular relates to a data center task scheduling modeling method based on multi-dimensional identification and mean field theory. BACKGROUND

[0002] The traditional TCP / IP system is based on an end-to-end communication model, which simplifies network interaction as a dialogue between two hosts, leading to traffic transmission along a pre-set static routing path, which is prone to cause single-point congestion and resource competition. With the rapid development of mobile communication, industrial internet, Internet of Vehicles, Internet of Things, and integration of space and earth, the core needs of users have changed from simple host access to convenient and reliable access to required content and services at any time and place. However, the traditional IP addressing method exposes three outstanding problems due to its static nature, singleness, and lack of semantic information: it is difficult to support efficient mobility management, service and security policies are difficult to bind flexibly, and the expansion and innovation of network architecture are severely limited.

[0003] With the rapid development of cloud computing technology, data centers, as the physical basis for supporting cloud services, are facing unprecedented challenges. The widespread application of cloud computing has brought an explosive growth in data volume, which requires data centers to have higher processing power, stronger storage performance, and better energy efficiency. In this context, hardware manufacturers such as NVIDIA are constantly pushing technological breakthroughs by increasing the number of cores in servers to improve the processing power of individual servers to meet the growing demand for computing. However, as the number of server cores increases, the resource optimization problem in the data center server cluster becomes more complex. Traditional resource management methods often struggle to adapt to dynamic changes and high concurrency requests in a multi-core environment, leading to uneven resource utilization, prolonged response times, and even performance bottlenecks.

[0004] In traditional single-queue system analysis, the entire data center or a group of servers is often simplified as a model of a centralized queue and multiple service stations. However, this method has significant limitations when analyzing large-scale data centers composed of a large number of homogeneous servers: first, the overall state space of the system grows exponentially with the number of servers, making the model too complex to solve; second, this model cannot accurately depict the dynamic coupling and mutual influence between servers due to load balancing strategies such as shortest queue priority. These limitations make the traditional single-queue model inaccurate and computationally expensive when evaluating system macro performance and optimizing resource allocation. SUMMARY

[0005] To address the aforementioned technical issues, this application provides a data center task scheduling modeling method based on multidimensional identification and mean field theory. Combining multidimensional identification system and mean field queuing theory, the proposed scheduling method can achieve more comprehensive dynamic resource management in multi-core server environments, providing strong support for the efficient operation of data centers.

[0006] To achieve the above objectives, this application employs the following technical solution:

[0007] This application discloses a data center task scheduling modeling method based on multidimensional identification and mean field theory. The method comprises three stages: stage one is real-time task scheduling; stage two is establishing a mean field model; and stage three is configuration optimization. The real-time operational data generated in stage one drives the update of the mean field model in stage two. The macroscopic performance indicators of stage two drive the optimization in stage three and change the operational data of stage one based on the optimization results. Specifically, the method includes the following steps:

[0008] Phase 1 real-time task scheduling specifically involves: deploying an SDN controller in the data center; the SDN controller maintaining a global resource state table; and encapsulating server state information into a server state tuple. The SDN controller calculates load metrics for each server. For each server The SDN controller is based on the current server Status information, estimating the expected dwell time for a new task if assigned to this server. Based on load level indicators and expected length of stay The SDN controller is for each server Construct a weighted cost function When a new task arrives, the SDN controller traverses the global resource state table and selects the option that maximizes the cost function. Minimum server As the allocation target, the SDN controller will assign servers. The identifier is returned to the scheduler, which then forwards the task to the server, entering phase two;

[0009] Phase Two specifically involves: constructing a mean-field model based on all homogeneous servers in the data center to obtain the steady-state probability distribution. Based on the obtained steady-state probability distribution Calculate system-level macro-performance metrics, including average task dwell time. Average server power and average task energy consumption ;

[0010] Phase three specifically involves defining the data center resource allocation problem as a constrained optimization problem, where the decision variable is the number of servers. and the number of cores per server The macro performance indicators of Phase 2 drive the optimization of Phase 3 and change the running data of Phase 1 based on the optimization results, thus realizing a closed loop of real-time scheduling, performance modeling, configuration optimization and scheduling adjustment.

[0011] A further improvement of this application is that step 1 specifically includes the following steps:

[0012] Step 1.1, SDN Controller Initialization and Server Agent Deployment: Deploy the SDN controller in the data center and deploy a lightweight monitoring agent on each server to collect server status information, including CPU utilization. Total CPU Memory usage Total memory Current number of tasks Core number and single core service rate ;

[0013] Step 1.2: Through the deployed lightweight monitoring agent, the collected server status information is reported to the SDN controller. The SDN controller maintains a global resource status database, forming a global, real-time, and consistent view of the resources throughout the entire data center. For each server... Encapsulate server status information into a server status tuple. :

[0014]

[0015] in: For server The number of cores For server Single-core service speed, For server Total memory, For server Total CPU computing resources For server Currently used memory, For server Currently used CPU resources For server Number of current tasks pending;

[0016] Step 1.3, for each server , the SDN controller computes a comprehensive, normalized load degree indicator : :

[0017]

[0018] wherein, is a configurable weight parameter, used to adjust the relative importance of memory and CPU resources in load evaluation;

[0019] Step 1.4, for each server , the SDN controller estimates an expected residence time required for a new task to be assigned to the server based on current server state information, the residence time including queuing time and service time, each server is modeled as an M / M / c queue, the expected residence time of M / M / c queue :

[0020]

[0021] wherein, is the instantaneous server load factor, is the recent server instantaneous arrival rate, calculated by the SDN controller, is formula, representing the probability of a new task needing to wait in queue;

[0022] Step 1.5, based on the load degree indicator of Step 1.3 and the expected residence time of Step 1.4, the SDN controller constructs a weighted cost function for each server :

[0023]

[0024] wherein, is a scheduling weight parameter, used to balance between immediate energy consumption tendency and task delay tendency, when a new task arrives, the SDN controller traverses the global resource state table established in Step 1, selects the server that makes the cost function minimum as the optimal target, the SDN controller returns the identifier of the server to the scheduler, the scheduler then forwards the task to the server.

[0025] Further improvement of the present application is that in phase two, all homogeneous servers of the data center are modeled as an average interaction system, and the probability distribution of all server states is constructed to form an average field interaction system, specifically, an average field model is constructed to obtain a steady-state probability distribution comprises the following steps:

[0026] Step 2.1.1, since the scheduling strategy is equivalent to the global shortest queue (JSQ), in the average field limit, the instantaneous arrival rate of a new task received by the server with the current number of tasks :

[0027]

[0028] wherein: is the global task arrival rate, is the cumulative probability distribution function, is the number of servers, is the probability distribution, is the summation index vector;

[0029] Step 2.1.2, construct the average field steady-state equation set: establish the master equation describing the evolution of the probability distribution

[0030]

[0031] Let the differential term be 0, that is, the nonlinear average field steady-state equation set when the system reaches steady state is obtained:

[0032]

[0033] wherein, is the probability that the number of tasks of any server at time is equal to is the single-core service rate of the current number of tasks is the instantaneous arrival rate, is the single-core service rate, , is the instantaneous arrival rate, is the single-core service rate, is the instantaneous arrival rate of the current number of tasks is the single-core service rate of the current number of tasks is the probability distribution of the current number of tasks is the probability distribution of the current number of tasks

[0034] ​​​​​​Step 2.1.3, solving average field steady-state equation and calculating performance index: the constructed average field steady-state equation group is solved by using fixed-point iteration algorithm, and the steady-state probability distribution is obtained .

[0035] Further improvement of the present application is that in step 2.1.3, solving average field steady-state equation, specifically comprising the following steps:

[0036] Step 2.1.3.1, initialization: setting initial steady-state probability distribution ;

[0037] Step 2.1.3.2, setting iteration times , according to the current distribution Calculate the cumulative probability distribution , the cumulative probability distribution is brought into the calculation formula of instantaneous arrival rate to calculate the arrival rate function , the arrival rate function and are substituted into the average field steady-state equation group, and the new distribution is obtained by solving.

[0038] Step 2.1.3.3, calculating error and termination judgment: if ||, is the preset, exit the loop and output the steady-state probability distribution , otherwise return to step 2.1.3.1.

[0039] Further improvement of the present application is that in stage two, the macro performance index is respectively:

[0040] Average task residence time:

[0041]

[0042] Average server power:

[0043]

[0044] Average task energy consumption:

[0045]

[0046] Wherein, is the server idle power, is the server full load power.

[0047] Further improvement of the present application is that stage three comprises the following steps:

[0048] Step 3.1, constructing a two-dimensional grid, and for each candidate configuration in the two-dimensional grid , for a plurality of candidate server numbers in a preset range in combination with a single server core number ;

[0049] Step 3.2, sequentially perform the check constraint condition, if the constraint condition is satisfied, take the combination as the system input parameter, perform the load degree index and the expected residence time , obtain the corresponding average task residence time and average task energy consumption ;

[0050] Step 3.3, calculate the objective function value of the combination , minimize the objective function value:

[0051]

[0052] select the combination with the minimum objective function value as the optimal system configuration , distribute the optimal system configuration to the infrastructure management platform of the data center, and re-adjust the candidate configuration .

[0053] Further improvement of the present application is that in step 3.2, the constraint condition of the constraint check is:

[0054] system stability constraint: server load factor , to ensure stable operation of the system;

[0055] server number constraint: ; is a set of positive integers, is the maximum server number;

[0056] core number constraint: , is the maximum core number;

[0057] cost constraint: , wherein is the cost per core;

[0058] power constraint: , wherein, is the maximum total power consumption;

[0059] resource constraint: and , wherein is the maximum task number estimate.

[0060] The beneficial effects of the present application are:​​

[0061] The application considers the influence of time delay and energy consumption through a multi-dimensional cost function, and realizes the optimization of comprehensive performance.

[0062] The application realizes more accurate load balancing than traditional sampling strategies by using the global view of SDN, and avoids sampling overhead and local optimization.

[0063] The mean field theory model of the application greatly reduces the complexity of modeling a super large scale system, provides accurate and fast insight into the macro behavior of the scheduling system, and overcomes the problem of rapid increase of memory and computing resource consumption due to the increase of the number of system states in the state space composed of a large number of server nodes and their states.

[0064] The application combines real-time scheduling with offline or online (or online) configuration optimization to form a closed loop, so that the data center can automatically adapt to load changes and realize elastic optimization configuration of resources. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 The figure is a schematic diagram of the scheduling system of the application.

[0066] Figure 2 The figure is a general flow chart of the task scheduling modeling method of the scheduling system architecture of the application.

[0067] Figure 3 The figure is a flow chart of the fixed-point iteration algorithm of the mean field model of the application. DETAILED DESCRIPTION

[0068] In the following, the embodiments of the application will be disclosed with drawings, and many practical details will be described in the following description for the purpose of clear illustration. However, it should be understood that these practical details should not be used to limit the application. That is, these practical details are not necessary in some embodiments of the application.

[0069] The application provides a data center task scheduling system based on multi-dimensional identification and mean field theory, as shown in Figure 1 The scheduling system includes an application layer, a control layer and a data layer, the control layer is composed of an SDN controller and various controller applications, mainly realizes the control and management of the data layer, including network topology discovery, routing calculation, traffic control, etc., the traffic classification module and the QoS guarantee module are also deployed in the SDN controller, which are abstracted as two modules of resource optimization and task scheduling in the application; the data layer is composed of various network devices, such as switches, routers, host clusters, etc., mainly realizes the forwarding and processing of network data, including the forwarding of data packets, traffic control, load feedback mechanism, etc.; the application layer provides various application services of network management and control, such as balancing, security management, quality of service guarantee, etc.

[0070] The scheduling system based on multidimensional identifiers comprises a central SDN controller and multiple distributed schedulers. The SDN controller maintains a global resource view database and collects tuple data from all servers in real time via a southbound interface. Scheduler: Located in the data plane, it receives arriving tasks and queries the SDN controller for the optimal server address.

[0071] like Figure 2 As shown, this application provides a data center task scheduling modeling method based on multidimensional identification and mean-field theory. The data center task scheduling modeling method includes three stages: stage one is real-time task scheduling, stage two is establishing a mean-field model, and stage three is configuration optimization. The real-time operating data generated in stage one drives the update of the mean-field model in stage two. The macroscopic performance indicators of stage two drive the optimization in stage three and change the operating data of stage one based on the optimization results. Stage two is the expected dwell time in stage one. This provides the foundation for computation. The runtime data generated in Phase One... Instantaneous arrival rate, steady-state probability distribution, etc., are used to calibrate the second-constructed mean-field model, and the third-stage search outputs the optimal configuration. The system is reconfigured, and feedback is sent to Phase 1, thereby changing the operating parameters of Phase 1. Phase 2 runs periodically or triggered by a specific event, and the output... and Used for optimization of driving phase three.

[0072] The data center task scheduling modeling method includes the following steps:

[0073] Phase 1 real-time task scheduling specifically involves: deploying an SDN controller in the data center; the SDN controller maintaining a global resource state table; and encapsulating server state information into a server state tuple. The SDN controller calculates load metrics for each server. For each server The SDN controller is based on the current server Status information, estimating the expected dwell time for a new task if assigned to this server. Based on load level indicators and expected length of stay The SDN controller is for each server Construct a weighted cost function When a new task arrives, the scheduler queries the SDN controller for the optimal target. The SDN controller then traverses the global resource state table to select the target that best suits the cost function. Minimum server As the allocation target, the SDN controller will assign servers. the identifier is returned to the dispatcher, which then forwards the task to the server, entering phase two;

[0074] Phase two specifically comprises: based on all the homogeneous servers of the data center, an average field model is constructed to obtain a steady-state probability distribution , based on the obtained steady-state probability distribution , a macro performance index of the system level is calculated, and the macro performance index includes average task residence time , average server power , and average task energy consumption ;

[0075] Phase three specifically comprises: the data center resource configuration problem is defined as a constrained optimization problem, and the decision variables of the optimization problem are the number of servers and the number of single-server cores , the macro performance index of phase two drives phase three optimization and changes the running data of phase one according to the optimization result, to realize the closed loop of real-time scheduling-performance modeling-configuration optimization-scheduling adjustment.

[0076] In phase one, the system parameter configuration is first completed, and the system parameter configuration comprises: initial configuration of system basic parameters, including global task arrival rate , single-core service rate , server idle power (W), server full load power , scheduling weight parameter , load weight parameter , energy consumption weight , unit conversion factor for adjusting the importance of energy consumption, number of servers and number of cores per server ;

[0077] Phase one specifically comprises the following steps:

[0078] Step 1.1, SDN controller initialization and server agent deployment: deploy an SDN controller in the data center, and deploy a lightweight monitoring agent on each server, which is responsible for collecting server state information, and the server state information includes CPU usage , total CPU , memory usage , total memory , current number of tasks , number of cores and single-core service rate ;

[0079] Step 1.2, the collected server status information is reported to the SDN controller through the deployed lightweight monitoring agent, the SDN controller maintains a global resource status database, forming a global, real-time, consistent view of the entire data center resources, for each server , the server status information is encapsulated into a server status tuple :

[0080]

[0081] wherein: is the number of cores of the server , is the single-core service rate of the server , is the total memory of the server , is the total CPU computing resource of the server , is the currently used memory of the server , is the currently used CPU resource of the server , is the number of current tasks to be processed of the server ;

[0082] Step 1.3, for each server , the SDN controller calculates a comprehensive, normalized load degree index according to the server status tuple :

[0083]

[0084] wherein, is a configurable weight parameter, used to adjust the relative importance of memory and CPU resources in load evaluation;

[0085] Step 1.4, for each server , the SDN controller estimates an expected residence time required for a new task to be assigned to the server according to the current server status information, the residence time includes queuing time and service time, each server is modeled as an M / M / c queue, the expected residence time of the M / M / c queue :

[0086]

[0087] wherein, is the instantaneous server Load factor For recent servers The instantaneous arrival rate is calculated by the SDN controller. yes The formula represents the probability that a new task needs to wait in the queue;

[0088] Step 1.5: Based on the load level index in Step 1.3 And the expected length of stay in step 1.4 The SDN controller is for each server Construct a weighted cost function :

[0089]

[0090] in, For scheduling weight parameters, This is used to balance immediate energy consumption tendency and task latency tendency. When a new task arrives, the SDN controller traverses the global resource state table established in step 1 and selects the resource that best suits the cost function. Minimum server As the optimal target, the SDN controller will... The identifier is returned to the scheduler, which then forwards the task to the server.

[0091] In Phase Two, all homogeneous servers in the data center are modeled as an average interactive system, and the probability distribution of all server states is... To construct a mean-field interaction system, specifically, a mean-field model is built to obtain the steady-state probability distribution. Includes the following steps:

[0092] Step 2.1: Since the scheduling strategy is equivalent to the Global Shortest Queue (JSQ), under the average field limit, the current number of tasks... The instantaneous arrival rate of a new task received by the server :

[0093]

[0094] in: For the global task arrival rate, Let be the cumulative probability distribution function. For the number of servers, For probability distribution, To sum the index vector;

[0095] Step 2.2: Construct the mean-field steady-state equations: Establish a system describing the probability distribution. The master equation of evolution:

[0096]

[0097] Let the differential term be 0, i.e. get the nonlinear mean field steady state equations of the system when it reaches steady state:

[0098]

[0099] where, is the probability that the number of tasks in the system at time is equal to , is the current number of tasks , , is the instantaneous arrival rate, is the single-core service rate, is the instantaneous arrival rate of the current number of tasks , is the single-core service rate of the current number of tasks , is the probability distribution of the current number of tasks , is the probability distribution of the current number of tasks ;

[0100] Step 2.3, solving the mean field steady state equations and calculating performance indicators: using fixed point iteration algorithm to solve the mean field steady state equations constructed, solving the steady state probability distribution .

[0101] As shown in Figure 3 , solving the mean field steady state equations, specifically including the following steps:

[0102] Step 2.3.1, initialization: set the initial steady state probability distribution ;

[0103] Step 2.3.2, set the number of iterations , calculate the cumulative probability distribution according to the current distribution , bring the cumulative probability distribution into the calculation formula of the instantaneous arrival rate to calculate the arrival rate function , substitute the arrival rate function and into the mean field steady state equations, and solve to get the new distribution ;

[0104] Step 2.3.3, calculate the error and terminate the judgment: if ||, As a preset, exit the loop and output the steady-state probability distribution , otherwise return to step 2.3.1.

[0105] In phase two, the macro performance indicators are respectively:

[0106] Average task residence time:

[0107]

[0108] Average server power:

[0109]

[0110] Average task energy consumption:

[0111]

[0112] Wherein, is the server idle power, is the server full load power.

[0113] Phase three includes the following steps:

[0114] Step 3.1, construct a two-dimensional grid, one dimension is the number of servers , the other dimension is the number of cores of the server , the grid density initial search uses step size 10, step size 5, fine search uses step size 2, step size 1, for each candidate configuration in the two-dimensional grid , for a plurality of candidate server numbers and single server core numbers in the preset range; since the number of servers and the number of cores per server are integer variables and have limited value ranges, discrete search is suitable; secondly, the objective function may have multiple local optimal solutions, and grid search can fully explore the solution space; finally, this method is simple to implement, the results are reliable, and is suitable for engineering applications.

[0115] Step 3.2, check the constraint conditions in turn, if the constraint conditions are met, take the combination as the system input parameter, execute the load level indicator and the expected residence time , get the corresponding average task residence time and average task energy consumption ; the constraint conditions for constraint checking are:

[0116] System stability constraint: server load factor to ensure the system runs stably;

[0117] Server number constraint: ; is a set of positive integers, is the maximum number of servers;

[0118] Core number constraint: , is the maximum number of cores;

[0119] Cost constraint: where is the cost per core;

[0120] Power constraint: where, is the maximum total power consumption;

[0121] Resource constraint: and where is the maximum number of tasks estimate.

[0122] Step 3.3, calculate the objective function value of the combination minimize the objective function value:

[0123]

[0124] Select the combination with the minimum objective function value as the optimal system configuration , and issue the optimal system configuration to the infrastructure management platform of the data center, trigger resource adjustment actions such as initializing new servers, shutting down old servers into energy-saving mode, or adjusting the core allocation of virtual machines, and re-adjust the candidate configuration through automated scripts or APIs .

[0125] This application combines multi-dimensional identification system and mean field queuing theory to propose a resource optimization modeling that can achieve more comprehensive dynamic resource management in a multi-core server environment. First, the SDN technology architecture is used to obtain the global perspective of the entire data center, then the queuing strategy is confirmed according to the set load index and the server multi-dimensional identifier is returned, finally the mean field queuing theory is used to model the system, predict the performance indicators under different resource configurations, and maximize the overall performance and stability of the system. With the continuous development of cloud computing and the continuous expansion of data center scale, this advanced resource optimization modeling method will play an increasingly important role and provide strong support for the efficient operation of data centers.

[0126] ​The above merely illustrates the embodiments of the present application but should not be taken as limitations. For those skilled in the art, the present application can have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. within the spirit and principles of the present application shall be included in the scope of the claims of the present application.

Claims

1. A data center task scheduling modeling method based on multidimensional identification and mean field theory, characterized in that: The data center task scheduling modeling method comprises three stages: stage one is real-time task scheduling, stage two is establishment of an average field model, and stage three is configuration optimization; real-time operation data generated in stage one drives average field model updating in stage two, macro performance indexes in stage two drive optimization in stage three and change operation data in stage one according to the optimization result; and the method comprises the following steps: Phase 1 real-time task scheduling specifically involves: deploying an SDN controller in the data center; the SDN controller maintaining a global resource state table; and encapsulating server state information into a server state tuple. The SDN controller calculates load metrics for each server. For each server The SDN controller is based on the current server Status information, estimating the expected dwell time for a new task if assigned to this server. Based on load level indicators and expected length of stay The SDN controller is for each server Construct a weighted cost function When a new task arrives, the SDN controller traverses the global resource state table and selects the option that maximizes the cost function. Minimum server As the allocation target, the SDN controller will assign servers. The identifier is returned to the scheduler, which then forwards the task to the server, entering phase two; The second stage specifically comprises: constructing an average field model based on all the homogeneous servers of the data center to obtain a steady state probability distribution , calculating system-level macro performance indexes based on the obtained steady state probability distribution , wherein the macro performance indexes comprise average task residence time , average server power and average task energy consumption ; The third stage is specifically: defining the data center resource configuration problem as a constrained optimization problem, the decision variable of the optimization problem being the number of servers and the number of cores of a single server The macro performance indicators of the second stage drive the third stage optimization and change the running data of the first stage according to the optimization results, realizing a closed loop of real-time scheduling-performance modeling-configuration optimization-scheduling adjustment. 2.The method of claim 1, wherein: Stage one specifically comprises the following steps Step 1.1, SDN controller initialization and server agent deployment: deploy SDN controller in data center and deploy lightweight monitoring agent in each server, responsible for collecting server status information, including CPU usage , total CPU , memory usage , total memory , current task number , core number and single-core service rate ; Step 1.2, the collected server status information is reported to the SDN controller through the deployed lightweight monitoring agent, the SDN controller maintains a global resource status database, forming a global, real-time and consistent view of the entire data center resources, and for each server , the server status information is encapsulated into a server status tuple : ， in: For server The number of cores For server Single-core service speed, For server Total memory, For server Total CPU computing resources For server Currently used memory, For server Currently used CPU resources For server Number of current tasks pending processing; Step 1.3, for each server , the SDN controller computes a comprehensive, normalized load degree indicator , based on the server state tuples : ， wherein, is a configurable weight parameter, for adjusting the relative importance of memory and CPU resources in the load evaluation; Step 1.4, for each server , the SDN controller estimates the expected residence time required for a new task to be assigned to the server based on the current server state information , the residence time includes the queuing time and the service time, models each server as an M / M / c queue, the expected residence time of an M / M / c queue : ， wherein, is the instantaneous server load factor, is the recent server instantaneous arrival rate, computed by the SDN controller, is the formula, which represents the probability that a new task needs to be queued. Step 1.5: Based on the load level index in Step 1.3 And the expected length of stay in step 1.4 The SDN controller is for each server Construct a weighted cost function : ， wherein, is a scheduling weight parameter, , for balancing between instant energy consumption tendency and task delay tendency, when a new task arrives, the SDN controller traverses the global resource state table established in step 1, and selects the server that minimizes the cost function as the optimal target, the SDN controller returns the identifier of the server to the scheduler, which then forwards the task to the server. 3.The method of claim 1, wherein: In phase two, all the homogeneous servers of the data center are modeled as an average interaction system, and the probability distribution of all the server states An average field interaction system is constituted, specifically, an average field model is built, and a steady state probability distribution is obtained The method comprises the following steps: Step 2.

1. Since the scheduling strategy is equivalent to the Global Shortest Queue (JSQ), under the average field limit, the current number of tasks... The instantaneous arrival rate of a new task received by the server : ， wherein: is the global task arrival rate, is the cumulative probability distribution function, is the number of servers, is the probability distribution, is the summation index vector; Step 2. 2, Constructing the mean-field steady-state equations: Establishing the description of the probability distribution Evolution of the master equation: ， Let the differential term be zero, i.e. obtain the nonlinear mean-field steady-state equations when the system has reached a steady state: ， wherein, is the number of tasks at time is the number of tasks at time is the probability that the number of tasks at time is the single-core service rate for the current number of tasks is the single-core service rate for the current number of tasks , is the instantaneous arrival rate for the current number of tasks is the single-core service rate for the current number of tasks is the instantaneous arrival rate for the current number of tasks is the single-core service rate for the current number of tasks is the probability distribution for the current number of tasks is the probability distribution for the current number of tasks is the probability distribution for the current number of tasks is the probability distribution for the current number of tasks is the probability distribution for the current number of tasks is the probability distribution for the current number of tasks Step 2. Solving the mean-field steady-state equations and calculating performance metrics: The constructed mean-field steady-state equations are solved using a fixed-point iteration algorithm, and the steady-state probability distribution is obtained . 4.The method of claim 3, wherein: In step 2.3, the average field steady-state equation is solved, specifically comprising the following steps: Step 2. 3.1, Initialization: Set initial steady-state probability distribution ; Step 2.3.2, Set the number of iterations , according to the current distribution Calculate the cumulative probability distribution , the cumulative probability distribution into the calculation formula of the instantaneous arrival rate Calculate the arrival rate function , the arrival rate function and Substitute into the mean field steady-state equation, and solve to get the new distribution ; Step 2.3.3, compute error, termination test: if ||, is preset, exit loop and output steady-state probability distribution , otherwise go to step 2.3.

1. 5.The method of claim 4, wherein: In stage two, the macro performance indexes are as follows: Average task residence time: ， Average server power: ， Average task energy consumption: ， wherein, Pserver idle is the server idle power, Pserver full is the server full load power. 6.The method of claim 1, wherein: Stage three comprises the following steps: Step 3.1, constructing a two-dimensional grid, for each candidate configuration in the two-dimensional grid for a plurality of candidate server numbers within a preset range in combination with a single-server core number ​ Step 3.2, check the constraint condition one by one, if the constraint condition is satisfied, take the combination as the system input parameter, execute the load degree index and the expected residence time , get the corresponding average task residence time and average task energy consumption ; ; Step 3.3, compute the objective function value of the combination minimize the objective function value: ， selecting a candidate configuration with the minimum target function value combining as the optimal system configuration issuing the optimal system configuration to an infrastructure management platform of the data center to trigger a resource adjustment action through an automated script or an API to readjust the candidate configuration . 7.The method of claim 6, wherein: In step 3.2, the constraint condition of the constraint check is as follows: System stability constraints: server load factor to ensure system stability; Server number constraint: ; is a set of positive integers, is the maximum server number; Core count constraint: , is the maximum core count; Cost constraints: where is the cost per core; Power constraints: wherein, is the maximum total power consumption; Resource constraints: and where is the maximum task number estimate.

Images