Energy efficiency calculation method and device of cloud computing facility, equipment and medium

Abstract

The invention discloses an energy efficiency calculation method and device for a cloud computing facility, equipment and a medium. Servers of the cloud computing facility are divided into different grades according to obtained performance parameters; establishing a service time and server resource integrated energy consumption model according to the divided grades so as to determine the resource efficiency of each server; and converting server deployment with maximum energy efficiency into an integer linear programming problem according to the resource efficiency, determining an objective function and constraint conditions, and solving to obtain the energy efficiency degree of the cloud computing facility. According to the scheme, the energy efficiency calculation precision of the cloud computing facility can be improved, uneven energy resource distribution is avoided, and therefore the energy utilization efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud computing, and in particular, to an energy efficiency calculation method, device, equipment and medium for cloud computing facilities. Background Art

[0002] In the era of rapid development of information technology today, cloud computing has become an important part of power facilities. A cloud computing center usually consists of thousands of servers. With the popularization of big data, artificial intelligence, etc. in power applications, the cloud computing center is facing unprecedented data processing pressure. Servers consume a large amount of electric energy during operation, and with the change of load, the energy consumption level fluctuates greatly. Evaluating the energy efficiency of different cloud computing facilities is beneficial to energy scheduling support to ensure the efficient operation of each cloud computing facility.

[0003] Traditional energy efficiency evaluation schemes often ignore the performance differences and different energy consumption characteristics among servers, resulting in uneven distribution of energy resources and low energy utilization efficiency. Summary of the Invention

[0004] In view of the above defects, the present invention provides an energy efficiency calculation method, device, equipment and medium for cloud computing facilities, which can improve the accuracy of energy efficiency calculation of cloud computing facilities, avoid uneven distribution of energy resources, and thus improve energy utilization efficiency.

[0005] An embodiment of the present invention provides an energy efficiency calculation method for cloud computing facilities, and the method includes: Dividing the servers of the cloud computing facility into different levels according to the obtained performance parameters; Establishing an energy consumption model integrating service time and server resources according to the divided levels to determine the resource efficiency of each server; Converting the deployment of the server with the maximum energy efficiency into an integer linear programming problem according to the resource efficiency, determining the objective function and the constraint conditions, and solving to obtain the energy efficiency degree of the cloud computing facility.

[0006] Preferably, dividing the servers of the cloud computing facility into different levels according to the obtained performance parameters includes: Obtaining the memory bandwidth utilization rate and the number of cycle instructions executed by the servers of the cloud computing as the performance parameters; Determining the servers with a memory bandwidth utilization rate lower than a preset first threshold or a number of cycle instructions executed lower than a preset second threshold as second-level servers; Determining the servers with a memory bandwidth utilization rate not lower than the first threshold and a number of cycle instructions executed not lower than the second threshold as first-level servers.

[0007] As a preferred solution, dividing the servers of the cloud computing facility into different levels according to the obtained performance parameters includes: Obtain the number of computing units, single instruction multiple data width, loop unrolling degree, and workgroup size of the server as the performance parameters; According to the performance parameters, use a preset performance model to calculate the performance metrics of different servers; Determine the servers with performance metrics less than a preset third threshold as first-level servers; Determine the servers with performance metrics not less than the third threshold as second-level servers.

[0008] Preferably, the method further includes; Perform offline critical path analysis using a vertex-down sorting algorithm to determine the critical path; Identify the usage scenarios of the cloud computing facility; When a first-level server or a second-level server completes the previous operator calculation and enters the idle state, analyze according to the critical path. When an operator in the to-be-started state with a clear dependency relationship is identified, schedule the identified operator to run on the idle server in the order of levels; or, When a first-level server or a second-level server completes the operator calculation and enters the idle state, check the power sensor and the task queue, obtain the current power consumption data from the power sensor, and calculate the peak power consumption of all operators in the system after scheduling an operator to execute on a certain idle device; when the peak power consumption of all operators after scheduling does not exceed the system power consumption upper limit value, schedule the operator to run on the idle device.

[0009] Preferably, establish an energy consumption model integrating service time and server resources according to the divided levels to determine the resource efficiency of each server, including: Establish a user service request set including a number of user request vectors according to the obtained instance set; Establish an energy consumption model of the server according to the working parameters of the server; Calculate the energy consumed by each server according to the energy consumption model; Calculate the energy consumed by each server according to the consumed energy; Determine the resource efficiency of each server according to the energy consumed by each server.

[0010] Preferably, the method further includes: Determine the description vector of the service quality of the independent application task under the cloud computing facility according to different service quality parameters; Construct a service matrix for the cloud computing facility to provide services to users according to the description vector; Perform normalization calculation on the service matrix to determine the normalized matrix; Introduce the processed values in the normalization matrix into a one-dimensional weight matrix to establish a weight factor set for the indicators, and determine the quality of service indicators for the target service. Determine the weight factor according to the quality of service indicator as the weight value of the quality of service parameter value. Construct a user service evaluation model based on the obtained weight value. Calculate according to the user service evaluation model to determine the quality of service evaluation value vector and obtain the quality of service matrix.

[0011] Furthermore, the energy efficiency of the cloud computing facility ; where is the number of users on the j-th physical node, is the energy consumption of the i-th user on the j-th physical node, represents the quality of service evaluation value of the service obtained by the i-th user on the j-th physical node.

[0012] An embodiment of the present invention further provides an energy efficiency calculation device for a cloud computing facility. The device includes: A partitioning module for partitioning the servers of the cloud computing facility into different levels according to the obtained performance parameters; An efficiency module for establishing an energy consumption model integrating service time and server resources according to the partitioned levels to determine the resource efficiency of each server; A calculation module for converting the deployment of the server with the maximum energy efficiency into an integer linear programming problem according to the resource efficiency, determining the objective function and the constraint conditions, and solving to obtain the energy efficiency of the cloud computing facility.

[0013] An embodiment of the present invention further provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the energy efficiency calculation method of the cloud computing facility as described in any one of the above embodiments.

[0014] An embodiment of the present invention further provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program. Wherein, when the computer program runs, it controls the device where the computer-readable storage medium is located to execute the energy efficiency calculation method of the cloud computing facility as described in any one of the above embodiments.

[0015] The energy efficiency calculation method, device, equipment and medium of the cloud computing facility provided by the present invention classify the servers of the cloud computing facility into different levels according to the obtained performance parameters; establish an energy consumption model integrating service time and server resources according to the classified levels to determine the resource efficiency of each server; transform the server deployment with maximized energy efficiency into an integer linear programming problem according to the resource efficiency, determine the objective function and the constraint conditions, and solve to obtain the energy efficiency degree of the cloud computing facility. This solution can improve the energy efficiency calculation accuracy of the cloud computing facility, avoid uneven distribution of energy resources, and thus improve the energy utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic flowchart of a method for calculating the energy efficiency of a cloud computing facility provided by an embodiment of the present invention; Figure 2 is a schematic structural diagram of an energy efficiency calculation device of a cloud computing facility provided by an embodiment of the present invention; Figure 3 is a schematic structural diagram of a terminal device provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0018] See Figure 1 , which is a schematic flowchart of a method for calculating the energy efficiency of a cloud computing facility provided by an embodiment of the present invention. The method includes steps S1 to S3: Step S1: Classify the servers of the cloud computing facility into different levels according to the obtained performance parameters; Step S2: Establish an energy consumption model integrating service time and server resources according to the classified levels to determine the resource efficiency of each server; Step S3: Transform the server deployment with maximized energy efficiency into an integer linear programming problem according to the resource efficiency, determine the objective function and the constraint conditions, and solve to obtain the energy efficiency degree of the cloud computing facility.

[0019] In the specific implementation of this embodiment, the servers of the cloud computing facility are classified into different levels according to the performance parameters of the servers of the cloud computing facility.

[0020] As a preferred embodiment, the servers can be divided into first-level servers and second-level servers. The first-level servers are used to implement high-performance operator execution; the second-level servers are used to provide performance lower than that of the first-level servers; to help the first-level servers reduce power consumption.

[0021] It should be noted that in this embodiment, the servers are divided into two levels. In other embodiments, the servers can be divided into other different levels.

[0022] Establish a resource provisioning model that integrates service time and server resources, and based on the server complementary deployment mechanism, to determine the resource efficiency of each server; Establish an energy efficiency evaluation model, transform the server deployment that maximizes energy efficiency into an integer linear programming problem according to the resource efficiency, determine the objective function and the constraints, and solve to obtain the energy efficiency of the cloud computing facility.

[0023] Assume that there are servers in the current system and user server requests. Then, the V server deployment based on maximizing energy efficiency can be formalized as the following integer linear programming ILP problem, and the objective function can be obtained as: ; The constraints are expressed as: ; ; ; Among them, represents the resource demand of the i-th user at time t, represents the resource amount that the j-th server can provide at time t, represents the resource demand of the i-th user at the time variable , represents the resource amount that the j-th server can provide at the time variable . Among them, the time variable belongs to the elements in the time set R and is used to limit different time situations in the constraints. is an integer programming variable of the ILP, and its value is 0 or 1. If takes the value of 1, it means that within the time segment , the i-th user server request is loaded (mapped) on the j-th server; if it is 0, there is no corresponding mapping relationship. The output matrix Shows the deployment results of the server. The first two constraint conditions respectively indicate that the resources provided by the server should be greater than the total demand of the user resources, and the total resource demand of the server on each service cannot exceed the resources provided by the server. The last constraint condition is that a server can only be loaded onto one server.

[0024] In the solution of this application, by introducing a second-level server, while ensuring the overall service performance, it can share part of the load for the first-level server, thereby reducing its operating power consumption and achieving energy conservation and emission reduction. A resource provision model integrating service time and server resources is established. This model is based on the complementary server deployment mechanism, comprehensively considering multiple dimensions such as the performance, power consumption, and service time of the server. Through algorithm optimization, the task allocation of each level of server is dynamically adjusted to achieve the maximization of energy efficiency. It can improve the energy efficiency calculation accuracy of cloud computing facilities, avoid uneven distribution of energy resources, and thus improve the energy utilization efficiency.

[0025] In another embodiment provided by the present invention, when performing level division in step S1, it includes the following steps: Obtain the memory bandwidth utilization rate and the number of cycles per instruction of the server of cloud computing as the performance parameters; This application sets two thresholds, namely the memory bandwidth utilization and the number of cycles per instruction The operator with a memory bandwidth utilization less than or an IPC less than is a candidate operator that can be offloaded to the second-level server for execution. Then, the performance of the corresponding operator of the candidate operator will be optimized on the second-level server and may be scheduled by the runtime. By using the above threshold-based method, operators that are not suitable for offloading to the second-level server can be screened out as early as possible to save the workload of operator optimization on the second-level server.

[0026] Determine the server with a memory bandwidth utilization rate lower than a preset first threshold or a number of cycles per instruction lower than a preset second threshold as the second-level server; Determine the server with a memory bandwidth utilization rate not lower than the first threshold and a number of cycles per instruction not lower than the second threshold as the first-level server.

[0027] In this application, by introducing a second-level server, while ensuring the overall service performance, part of the load can be shared for the first-level server, thereby reducing its operating power consumption and achieving energy conservation and emission reduction. A resource provisioning model integrating service time and server resources is established. Based on the server complementary deployment mechanism, this model comprehensively considers multiple dimensions such as server performance, power consumption, and service time, and through algorithm optimization, dynamically adjusts the task allocation of each level of servers to maximize energy efficiency. An energy efficiency evaluation model is established. This model converts the service quality level value that can be provided by the unit energy consumption value under cloud computing facilities into a scalar value within a certain range; realizing the qualitative evaluation of the energy efficiency of cloud computing facilities.

[0028] In this application, the servers are divided into first-level servers and second-level servers. The first-level servers are mainly used to implement high-performance operator execution; the second-level servers are not used to provide performance superior to that of the first-level servers; the second-level servers are used to run some operators that are traditionally executed on the first-level servers with lower power consumption, so as to help the first-level servers reduce power consumption and achieve the goal of saving the total system power consumption, thereby achieving the goal of energy conservation.

[0029] This application characterizes the workload of cloud computing facilities at the fine-grained operator level. By using the memory bandwidth utilization rate of operators on the first-level servers and the number of instructions executed per clock cycle (Instructions Per Cycle, IPC) as indicators, those operators that can achieve comparable or even better execution performance on the second-level servers than on the first-level servers are first determined. Then, by further analyzing the task data of the cloud computing facility workload, smaller operators that are not on the critical path of execution are identified. These operators will only be offloaded to the second-level server when it is ensured that offloading to the second-level server for execution plus data movement will not extend the execution time of the critical path of the task data. To address the challenges in runtime scheduling, this application designs a runtime system that schedules operators based on the previous offline analysis and operator characterization research. The runtime system will check the runtime power consumption to determine which operators should be executed on the first-level servers and the second-level servers and in what order, so as to minimize data movement and energy consumption without exceeding the user-specified power limit.

[0030] The first-level servers do not participate in operator calculations and are mainly responsible for operator allocation and scheduling operators to the relevant first-level servers and second-level servers at runtime.

[0031] The cloud computing facility includes two offline phases and a runtime scheduler. These two offline phases are used to determine candidate operators that can be offloaded to the second-level server and optimize the performance of these operators on the second-level server. The first offline phase performs power consumption and performance characterization of the operators and makes second-level server offloading decisions for the operators based on the computational intensity and memory bandwidth utilization of the operators. The second offline phase selects the optimal parameter configuration combination according to the performance model to optimize the performance of those candidate operators; the second offline phase also selects small operators for offloading to the second-level server according to the critical path analysis. The runtime scheduler performs operator scheduling according to the offloading decisions and user requirements (minimize energy consumption without degrading performance or maximize performance while strictly ensuring the power limit).

[0032] In yet another embodiment provided by the present invention, when performing the level division in step S1, the following steps may further be included: After determining which operators can be offloaded to the second-level server through the characterization study of the operators, the present application implements the corresponding second-level server operators of these operators using OpenCL, and optimizes their performance through rapid design space exploration to select the optimal parameter configuration combination and critical path analysis.

[0033] To optimize the performance of the second-level server operators, the number of computing units, single instruction multiple data width, loop unrolling degree, and workgroup size are obtained as the performance parameters.

[0034] The number of computing units CU is crucial for the construction of the second-level server operator pipeline and hardware utilization. The operator pipeline can be replicated multiple times to generate multiple CUs to achieve higher throughput. However, more CUs consume more hardware resources, which often reduces the operating frequency of the second-level server, thereby affecting performance.

[0035] Single instruction multiple data width Determines how many work items are executed in one instruction. Thanks to the potential memory coalescing access opportunities, a larger SIMD width can improve data parallel processing efficiency. However, a wider SIMD also puts more pressure on the memory bandwidth, and at the same time, due to the increased possibility of control path divergence, a wider SIMD may reduce the operator performance.

[0036] Determining the loop unrolling degree is also very challenging because a greater degree of loop unrolling provides more computational concurrency but also consumes more second-level server resources shared by other components.

[0037] The workgroup size determines the number of work items in each workgroup. A larger workgroup size allows the compiler to apply more optimizations to fully utilize the hardware resources without using additional second-level server logic resources. However, the workgroup size is also limited by the available shared second-level server resources and the maximum workgroup size supported by the hardware.

[0038] Determining the optimal combination of the above four parameter configurations is very time-consuming. To avoid the time-consuming process of evaluating the operators optimized for each parameter configuration combination, this application introduces a performance model to indirectly evaluate and compare the performance of the operators for each parameter configuration combination. In particular, given the operator input and parameter configuration, the performance is evaluated by a proxy metric called the Latency Indicator (LI), and the performance metric LI of different servers is calculated using a preset performance model.

[0039] ; where and are the dimension sizes of the two-dimensional input , is the dimension size of the two-dimensional input , so the numerator in the formula quantifies the amount of computation of the entire operator. is the base time quantity, represents the workgroup size, represents the number of CUs, represents the width of the SIMD, and ul represents the loop unrolling degree. is the second-level server working frequency pre-allocated to the operator by the offline compiler according to the second-level server resource usage, and it is also a function determined by influencing factors including the number of CUs, the width of the SIMD, the degree of loop unrolling, etc. A smaller LI value indicates better operator performance.

[0040] The performance of the operator is related to the working frequency and the input size. Smaller input and higher working frequency can bring better performance, and both can be captured in the latency indicator. As a bridge, it connects the relationship between the three configuration parameters and the performance. reflects the trade-off between the hardware resource constraints and the goal of using more CUs, wider SIMD, and higher degree of loop unrolling to obtain better performance. The formula also considers the impact of the workgroup size on the performance: a larger workgroup helps to obtain higher performance without violating the hardware constraints.

[0041] Given an operator and its input size, determine the optimal parameter configuration combination. For each combination, this application uses the offline compiler provided by the vendor to obtain The compiler analyzes to obtain the corresponding parameter configuration combination operator The process only takes a few seconds. An operator may have hundreds of optimized parameter configuration combinations, and obtaining all combinations only takes hundreds of seconds in total, which is far less than the compilation and execution time (up to hundreds of hours) involved in the second-level server operators that directly evaluate all possible different combinations of the parameter optimization design space on the second-level server. The input size and can be obtained through offline analysis.

[0042] In another embodiment provided by the present invention, the method further includes the following steps: Second-level server operator offloading decision based on critical path analysis: In addition to using operator characterization research to determine second-level server candidate operators, this application also uses critical path analysis to select those small operators that are not on the critical path for offloading to the second-level server. Running these small operators on the second-level server may require longer execution time than running on the first-level server.

[0043] This application uses the following method for offline critical path analysis. The algorithm proposed by this application is based on the calculation of vertex downward sorting. The vertex downward sorting is calculated by adding the computational complexity of the previous vertex on the longest path from any source vertex to . For a given data, the algorithm workflow of critical path analysis is as follows: Starting from the source vertex, calculate the downward sorting of each vertex in the complete graph.

[0044] Determine that the sink point with the largest lower rank is on the critical path.

[0045] According to the previous relationship, from the selected sink vertex to the source vertex, add the visited vertices to the critical path.

[0046] The search process of the critical path stops once it reaches the source point. Among the selected sink vertex and the predecessor relationship of the sink vertex, the longest path between any connected sink vertices is the critical path to be found.

[0047] According to the obtained critical path information, the following method can be used to select which small operators to offload to the second-level server. For any small operator not on the critical path, this application implements and optimizes it with OpenCL and measures its second-level server execution performance. Only when the execution time of the operator on the second-level server plus the data movement time between the CPU and the second-level server is less than the execution time of the critical path, will the operator be offloaded to the second-level server. This ensures that the critical path will not be extended.

[0048] The runtime system schedules operators based on the operator characterization results and offline critical path analysis. The runtime system adopts different scheduling strategies for two usage scenarios in the data center respectively: Usage scenario Minimize energy consumption without sacrificing training throughput; Usage scenario 2: Maximize performance (throughput) without exceeding a given power cap.

[0049] The runtime system assumes the following information is available when scheduling operators: the peak power consumption and average power consumption of the operators running on the first-level servers, and the power consumption of the operators running on the second-level servers. The above information can be collected through offline characterization studies.

[0050] Next, this application will discuss how the runtime system schedules operators for these two usage scenarios. The working processes of the two scheduling strategies are as follows: Scheduling scheme for usage scenario 1: When the first-level or second-level server finishes the previous operator calculation and enters the idle state, the runtime system examines the task queue. In the queue, some operators have clearly defined their dependencies and are in the pending start state. If, according to the results of the offline analysis, there are operators to be run scheduled to execute on these idle servers (first-level or second-level), the system will schedule these operators to the corresponding idle servers to run. In particular, if the first-level server is idle and only the operators prepared for the second-level server are ready in the queue, the system will preferentially schedule these operators to the first-level server to run without waiting. In addition, if there are multiple second-level server operators applicable to different input scales, the system will select the optimal operator to execute on the second-level server according to the type of operator input.

[0051] Scheduling scheme for usage scenario 2: When the first-level or second-level server finishes the operator calculation and is idle, the runtime system checks both the power sensor and the task queue simultaneously. The system obtains the current system power consumption data from the power sensor. If scheduling a certain operator to execute on an idle device (first-level or second-level server) will not cause the total system power consumption to exceed the power cap, the system will immediately schedule this operator to run on this idle device. If there is no such operator, the system will remain in a waiting state until a new operator is ready. This strictly ensures that the power consumption during system operation does not exceed the power cap. When both the first-level server and the second-level server are idle and do not exceed the power cap set by the system, the runtime system preferentially schedules operators to run on the first-level server to achieve the best performance.

[0052] Establish a resource provision model that integrates time (M service time) and space (server resources): Based on a complementary server deployment mechanism in terms of time and space (server service time and server resource requirements), the resource requirements of each complementary server can be maximally close to the host resource capacity they are loaded with throughout the service period. Then, the problem is transformed into an optimization problem of performance and energy consumption with multiple constraints, which can improve the energy efficiency of each resource and enable various resources of the system to work in a state of load balancing as much as possible within a relatively long time interval.

[0053] In another embodiment provided by the present invention, step S2 specifically includes the following steps: User service request set Responded by the M instance set. Each user resource request is a multi-dimensional resource vector. If there are requests, then is a -dimensional user request vector, denoted as , where , represents an M request, represents the resource vector of the i-th user's request for M, is for user is the set of the server service time, request service start time, and service end time that the user

[0054] The energy consumption model is expressed as: ; Where: is the activity factor, C represents the load capacitance, V represents the working voltage, f is the clock frequency corresponding to the working voltage, and p represents the energy consumption.

[0055] The energy consumed by the active server is ; Where: M is the total number of active servers, j ∈ M, the subscript j represents the parameters of the corresponding server, and a is a constant. A constant greater than 0.

[0056] Energy efficiency is an important evaluation index for data centers. When deploying servers, every factor including various resources and service time, etc. should be comprehensively considered. The new energy efficiency definition of the j-th server within a period of time is obtained: ; Where, represents the total energy consumed by the server within a period of time, is the resource efficiency of all resources within this time segment. When allocating resources, improving the resource efficiency of a single server and reducing energy consumption can both achieve the goal of enhancing the energy efficiency of the data center. Here and can be calculated respectively according to the following formulas: ; ; wherein, represents the coefficient related to resources, represents the th user on the th server, K represents the number of servers, T represents the time range of calculation, S represents the time interval where the time period s is located, the usage time of the th kind of resource, is the total energy-related quantity consumed by the jth server within a period of time, is a divisor term related to calculating energy, is the energy consumed by the jth server within the time period , is the effective running time of the jth server within the time period s.

[0057] In the same way, the resource efficiency of the server can be defined, which is the ratio of the sum of the utilization rates of all resources within the entire running time segment of M to the energy consumed by M.

[0058] Energy Efficiency Rate (EER) is an important concept of the energy efficiency ratio of cloud computing data centers, which is usually defined as the unit energy consumption for executing the load. The load is usually the processor frequency allocated to the server to determine the resource efficiency of each server.

[0059] In another embodiment provided by the present invention, the method further includes the following steps: Evaluate the energy efficiency. During the energy efficiency evaluation process, its purpose is to evaluate the performance quality, user-friendliness, and environmental sustainability of the cloud computing environment. The judgment criteria may be quantitative, qualitative, or dynamic. In many cases, it is not necessary to obtain the exact value of the performance energy consumption ratio. Based on the above research background, a reduction method for service quality parameters and a weighted energy efficiency evaluation model are proposed in this application.

[0060] Reduce the measurement values of multiple quality of service parameters for independent application tasks to the same dimension range. Then, obtain the evaluation weight matrix based on the user's multi-dimensional quality of service parameter scores for each service, and calculate the weight values of the user for different quality of service parameters. Obtain the user's final quality of service evaluation matrix from the measurement matrix and weight matrix of the quality of service parameters, and convert the quality of service level value that can be provided by the unit energy consumption value under the cloud computing facility into a scalar value within; Establish an energy efficiency evaluation model for the overall quality of service level value provided by unit energy consumption. Refer to the current energy efficiency labels in the home appliance industry, analyze the obtained energy efficiency data, classify the energy efficiency labels of the cloud data center, and finally convert the energy efficiency value under the cloud computing facility into a qualitative concept, realizing the qualitative evaluation of the energy efficiency under the cloud computing facility.

[0061] The quality of service description vector for a certain independent application task under the cloud computing facility can be defined as: For example are different quality of service parameters, such as service cost, execution time, reliability, available resources, etc. And the performance of the cloud data center can also be regarded as a special quality of service indicator. Therefore, this vector can not only describe the measurement of the actual quality of service indicators of the computing nodes in the data center under the cloud computing facility, but also express the user's requirements for the target quality of service indicators.

[0062] Service can be used to describe the th service provided by the data center for a certain user The service vector can be represented by the quality of service matrix of the user where is the measurement value of the th service of the user for the th dimensional quality of service parameter. Perform a normalization calculation on this matrix, and introduce the array ; Suppose the quality of service parameter of the th service of the th type of resource in a domain is expressed as . Through normalization processing, make ; Due to the inconsistency of the dimensions of different quality of service parameters, the above formula is transformed again to obtain: ; After transformation, the following formula is obtained: ; Each different user will obtain such a quality of service service matrix with a unified dimension. In addition, for each user, only the last 5 services and the 5 commonly used quality of service parameters are retained, so that each user maintains the service matrix.

[0063] Establish a quality of service parameter evaluation weight matrix, and use the weighting method to represent the overall quality of service level. Introduce the processed value after normalizing the quality of service parameters into a one-dimensional weight matrix to establish the weight factor set of the index. Use the method of user consultation to refer to the scale value , and score the relative importance of several indicators. Among them, the reference scale value is the actual demand of the user for the target quality of service indicator. Establish weight factors for different dimensions of quality of service parameters as the weight values of the quality of service parameter values, as shown in the following formula: ; Among them, represents the importance scale value of the user for the th service and the th dimension quality of service parameter index, and ; According to the parameter scale value and weight matrix of each service, obtain the user's evaluation of the data center's last services , as shown in the following formula: ; According to the above formula, ; Calculate the th root of the product of each element in the user scoring matrix : ; Perform normalization processing on to obtain ; Then, it can be obtained that is the quality of service evaluation value vector to be sought. The weighted average of this vector is used to obtain the quality of service level value V_service_quality for the user recently obtained. Then, the overall quality of service level of each physical node within a period of time is the weighted average of the V_service_quality of all users of the physical node within a time interval.

[0064] In this application, the energy efficiency is defined as the ratio of the quality of service to the energy consumption to describe the quality of service level that a physical node can provide per unit of energy consumption, and the performance of the system can also be regarded as a special type of quality of service parameter.

[0065] In another embodiment provided by the present invention, when determining the energy efficiency of a cloud computing facility, the following steps are included: Suppose there are physical nodes under a cloud computing facility, and there are users on each node. Then, within a certain time period, the energy consumption of user on a certain physical node is . The energy consumption under the cloud computing facility can be obtained by the following formula, where is the energy consumption coefficient, is the capacitive load, is the voltage of the processor, and is the frequency of the processor.

[0066] ; ; Among them, is the dynamic energy consumption, and is the static energy consumption.

[0067] Therefore, the energy efficiency value of each physical node is . The energy efficiency under the entire cloud computing facility is as shown in the formula. Finally, this value is adjusted to the scalar value within the discourse domain to intuitively quantify the energy efficiency of the cloud computing facility.

[0068] The energy efficiency of the cloud computing facility ; Among them, is the number of users on the jth physical node, is the energy consumption of the ith user on the jth physical node, and represents the quality of service evaluation value of the service obtained by the ith user on the jth physical node.

[0069] When the physical servers in the cloud data center are heterogeneous, the CPU, memory, bandwidth size, and energy consumption of the servers are all different. However, by classifying the cloud computing facilities with the same structure into one category, the heterogeneity can be divided, and parameter values can be determined for each homogeneous sub-environment. For the monitoring of parameters, a power monitor is used to collect the real-time power of the data center nodes, and the overall energy consumption value of the cloud computing system is observed in real time. In addition, monitoring agent software is deployed on each different type of node in the cloud system to sample and measure the multi-dimensional performance indicators under the cloud computing facility in real time.

[0070] An embodiment of the present invention further provides an energy efficiency calculation device for a cloud computing facility. Refer to Figure 2 , which is a schematic structural diagram of the energy efficiency calculation device for the cloud computing facility provided by the embodiment of the present invention. The device includes: A partitioning module, configured to partition the servers of the cloud computing facility into different levels according to the obtained performance parameters; An efficiency module, configured to establish an energy consumption model integrating service time and server resources according to the partitioned levels to determine the resource efficiency of each server; A calculation module, configured to convert the deployment of the server with the maximum energy efficiency into an integer linear programming problem according to the resource efficiency, determine the objective function and the constraint conditions, and solve to obtain the energy efficiency degree of the cloud computing facility.

[0071] It should be noted that the energy efficiency calculation device for the cloud computing facility provided by the embodiment of the present invention can execute the energy efficiency calculation method for the cloud computing facility described in any of the above embodiments, and the specific functions of the energy efficiency calculation device for the cloud computing facility will not be elaborated here.

[0072] Refer to Figure 3 , which is a schematic structural diagram of a terminal device provided by the embodiment of the present invention. The terminal device in this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an energy efficiency calculation program for a cloud computing facility. When the processor executes the computer program, the steps in the above-mentioned embodiments of the energy efficiency calculation method for the cloud computing facility are implemented, such as Figure 1 the steps S1 to S3 shown. Or when the processor executes the computer program, the functions of each module in the above-mentioned device embodiments are implemented.

[0073] Exemplarily, the computer program can be divided into one or more modules / units. The one or more modules / units are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of completing functions, and the instruction segments are used to describe the execution process of the computer program in the terminal device. For example, the computer program can be divided into each module, and the specific functions of each module will not be elaborated again.

[0074] The terminal device can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The terminal device may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the schematic diagram is only an example of the terminal device, and does not constitute a limitation on the terminal device. It may include more or fewer components than shown in the figure, or combine some components, or different components. For example, the terminal device may further include input / output devices, network access devices, a bus, etc.

[0075] The so-called processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the terminal device, and connects various parts of the entire terminal device through various interfaces and lines.

[0076] The memory can be used to store the computer program and / or module. The processor realizes various functions of the terminal device by running or executing the computer program and / or module stored in the memory, and by calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0077] Among them, if the modules / units integrated in the terminal device are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention, it can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0078] The above is the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art of the present technology, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A method for calculating energy efficiency of a cloud computing facility, characterized in that: The method comprises: Classify the servers of the cloud computing facility into different levels according to the obtained performance parameters; According to the classification level, an energy consumption model integrating service time and server resources is established to determine the resource efficiency of each server; According to the resource efficiency, the server deployment for maximizing energy efficiency is converted into an integer linear programming problem, the objective function and constraints are determined, and the problem is solved to obtain the energy efficiency of the cloud computing facility.

2. The method for calculating energy efficiency of a cloud computing facility according to claim 1, wherein: The servers of cloud computing facilities are divided into different levels according to the obtained performance parameters, including: Obtaining the memory bandwidth usage rate and the number of cycle instruction executions of the cloud computing server as the performance parameters; Determine as a second-level server a server whose memory bandwidth usage is lower than a preset first threshold or whose cycle instruction execution number is lower than a preset second threshold; A server whose memory bandwidth usage is not lower than the first threshold and whose cycle instruction execution number is not lower than the second threshold is determined as a first-level server.

3. The method for calculating energy efficiency of cloud computing facilities according to claim 1, wherein: The servers of cloud computing facilities are divided into different levels according to the obtained performance parameters, including: Obtaining the number of computing units, SIMD width, loop unrolling degree, and workgroup size of the server as the performance parameters; According to the performance parameters, the performance indicators of different servers are calculated using a preset performance model; Determine a server whose performance index is less than a preset third threshold as a first-level server; A server whose performance index is not less than the third threshold is determined as a second-level server.

4. The method for calculating energy efficiency of cloud computing facilities according to claim 1, wherein: The method further comprises: Use the vertex downward sorting algorithm to perform offline critical path analysis to determine the critical path; Identify usage scenarios of the cloud computing facility; When the first-level server or the second-level server completes the previous operator calculation and enters the idle state, an analysis is performed according to the critical path. When an operator in a waiting state and with a clear dependency is identified, the identified operator is scheduled to run on the idle server in order of rank; or, When the first-level server or the second-level server completes operator calculation and enters the idle state, it checks the power sensor and task queue, obtains the current power consumption data from the power sensor, and calculates the peak power consumption of all operators in the system after scheduling the operator to a certain idle device for execution; When the peak power consumption of all operators after scheduling does not exceed the upper limit of the system power consumption, the operator is scheduled to run on the idle device.

5. The method for calculating energy efficiency of cloud computing facilities according to claim 1, wherein: According to the classification level, an energy consumption model integrating service time and server resources is established to determine the resource efficiency of each server, including: Establishing a user service request set including a plurality of user request vectors according to the acquired instance set; Establish the energy consumption model of the server according to the working parameters of the server; Calculate the energy consumed by each server according to the energy consumption model; The resource efficiency of each server is determined based on the energy consumed by each server.

6. The method for calculating energy efficiency of cloud computing facilities according to claim 1, wherein: The method further comprises: Determine a description vector of the quality of service of an independent application task under the cloud computing facility according to different quality of service parameters; Constructing a service matrix for the cloud computing facility to provide services to the user according to the description vector; Performing normalization calculation on the service matrix to determine a normalized matrix; The processed values ​​in the normalized matrix are introduced into a one-dimensional weight matrix to establish a weight factor set of an indicator, and a service quality indicator for a target service is determined; Determine a weight factor according to the service quality indicator as a weight value of the service quality parameter value; Construct a user service evaluation model based on the obtained weight values; Calculation is performed according to the user service evaluation model to determine a service quality evaluation value vector and obtain a service quality matrix.

7. The method for calculating energy efficiency of a cloud computing facility according to claim 1, wherein: Energy efficiency of the cloud computing facility ; in, is the number of users on the jth physical node, is the energy consumption of the i-th user on the j-th physical node, It represents the service quality evaluation value of the service obtained by the i-th user on the j-th physical node.

8. An energy efficiency calculation device for cloud computing facilities, characterized in that: The device comprises: A partitioning module, used to partition the servers of the cloud computing facility into different levels according to the obtained performance parameters; The efficiency module is used to establish an energy consumption model integrating service time and server resources according to the divided levels to determine the resource efficiency of each server; The computing module is used to convert the server deployment for maximizing energy efficiency into an integer linear programming problem according to the resource efficiency, determine the objective function and constraints, and solve the problem to obtain the energy efficiency of the cloud computing facility.

9. A terminal device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the method for calculating the energy efficiency of a cloud computing facility as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the energy efficiency calculation method for cloud computing facilities as described in any one of claims 1 to 7.

Images