Knowledge-transfer-driven computational power effeiciency modeling method and apparatus for data center

Abstract

Disclosed in the present invention are a knowledge-transfer-driven computational power efficiency modeling method and apparatus for a data center. The method comprises: acquiring labeled efficiency data of a source computational power device and a target computational power device; pre-processing the efficiency data of the source computational power device and the target computational power device, so as to obtain respective key efficiency characteristic spaces of the source computational power device and the target computational power device; constructing a cross-computational-power-device efficiency model on the basis of imbalanced optimal transport; training the efficiency model, so as to implement the learning of differences between efficiency characteristic spaces of heterogeneous computational power devices; and using the trained efficiency model to perform efficiency modeling of a target computational power device to be evaluated, so as to evaluate the efficiency performance of said target computational power device. The present invention ameliorates the problems of reusing efficiency data of a computational power device being difficult, the modeling costs being high, and the precision of few-sample modeling being poor which are caused by there being a domain drift in efficiency characteristic spaces of heterogeneous computational power devices in a scenario where data acquisition is limited.

Description

Knowledge transfer driven data center computing power energy efficiency modeling method and device TECHNICAL FIELD

[0001] The present application belongs to the technical field of data centers, and specifically relates to a knowledge transfer driven data center computing power energy efficiency modeling method and device. BACKGROUND

[0002] In recent years, with the continuous expansion of global cloud data centers, their energy consumption has also increased dramatically. It is estimated that the global cloud data center electricity consumption will reach 240-340TWh in 2022, accounting for about 1%-1.3% of the global final electricity consumption. In addition, with the deployment of complex artificial intelligence applications by super-large-scale public cloud providers, especially large language models, the energy consumption problem of data centers has been further exacerbated. As a key infrastructure of data centers, the average resource utilization rate of the massive heterogeneous computing power equipment, i.e. various servers, is 12%-18%, but the energy consumption accounts for 42% of the overall energy consumption of data centers, so how to realize the efficient use of computing power resources has become the key to data center energy saving. At present, most energy efficiency optimization researches mainly aim to improve the resource utilization efficiency of computing power equipment, i.e. computing power equipment parameter optimization and computing power resource dynamic scheduling, etc. The common point of these researches is to find the correlation between the resource characteristics of computing power equipment and energy consumption, and this is also the significance of establishing the computing power equipment energy efficiency model.

[0003] With the rapid development of machine learning (ML), its powerful nonlinear fitting ability has been used to implement the energy efficiency modeling of computing devices. For example, "Lin W, Wu G, Wang X, et al. An artificial neural network approach to power consumption model construction for servers in cloud data centers [J]. IEEE Transactions on Sustainable Computing, 2020, 5(3): 329-340." uses the energy efficiency data of servers running different types of workloads to train an artificial neural network model for energy efficiency modeling. Similarly, "Wu W, Lin W, He L, et al. A power consumption model for cloud servers based on elman neural network [J]. IEEE Transactions on Cloud Computing, 2021, 9(4): 1268-1277." proposes a server time series energy efficiency model based on Elman neural network, which realizes the time series prediction of the energy efficiency of computing devices. Although the current large number of computing device energy efficiency modeling methods based on ML show superior performance, the research on energy efficiency modeling of computing devices in real data centers still faces the following two challenges:

[0004] Energy efficiency modeling of computing devices in data-restricted scenarios: On the one hand, the energy efficiency model of computing devices based on traditional ML methods often needs to collect a large amount of labeled data for model training in order to obtain an effective energy efficiency model. However, due to the short running time of newly deployed computing devices in data centers, the amount of energy efficiency data that can be obtained is also difficult to meet the training needs of the model. At the same time, limited by the commercial privacy of data centers, a large amount of energy efficiency data of computing devices, which directly reflects the running state of data centers, is often prohibited from being opened to the outside, which also leads to the challenge of difficulty and high cost of data collection for energy efficiency modeling of computing devices in real data centers. Therefore, how to realize the energy efficiency modeling of computing devices in real data centers in a data-restricted scenario is of great significance.

[0005] Energy efficiency data reuse across computing power devices: On the other hand, the energy efficiency model of computing power devices based on traditional ML methods can only perform better when the training data and test data are independent and identically distributed. However, in actual data centers, in order to meet the computing resource needs of different users, data centers often need to deploy a large number of heterogeneous computing power devices, such as general computing power devices with CPU chips as computing cores, and intelligent computing power devices with GPU, FPGA, TPU, NPU and other AI chips as computing cores. Due to the differences in hardware or running load, the energy efficiency performance of different computing power devices is different during running, which further causes the energy efficiency feature distribution of different computing power devices to drift, which makes it difficult for the energy efficiency model based on traditional ML to perform better generalization performance. Therefore, to get a high-precision energy efficiency model, a large amount of energy efficiency data of the specified target server needs to be collected for model training, thereby bringing high data collection cost. In fact, although different computing power devices have differences in running load or hardware, the data collected from different servers contains shared energy efficiency features, and traditional ML methods are difficult to effectively mine knowledge in a non-independent and identically distributed scenario. Therefore, how to further mine valuable knowledge contained in the collected energy efficiency data, realize energy efficiency data reuse across computing power devices, and improve the energy efficiency modeling performance of the target computing power device with only a small amount of labeled data, is of great significance for data center computing power device energy efficiency modeling.

[0006] SUMMARY

[0007] The main purpose of the present application is to overcome the shortcomings and deficiencies of the prior art, and to provide a knowledge transfer driven data center computing power energy efficiency modeling method and device, which learns the energy efficiency feature distribution difference between the source computing power device and the target computing power device, and then uses the valuable knowledge contained in the large amount of labeled energy efficiency historical data of the source computing power device to improve the energy efficiency modeling accuracy of the target computing power device with only a small amount of labeled energy efficiency data.

[0008] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0009] In a first aspect, the present application provides a knowledge transfer driven data center computing power energy efficiency modeling method, comprising the following steps:

[0010] Obtain the energy efficiency data of the source computing power device and the target computing power device; the source computing power device includes a large amount of labeled energy efficiency data, and the target computing power device includes a small amount of labeled energy efficiency data;

[0011] Preprocess the energy efficiency data of the source computing power device and the target computing power device respectively to obtain their respective key energy efficiency feature spaces; the key energy efficiency feature space refers to the feature most related to the energy efficiency of the computing power device;

[0012] An energy efficiency model of cross-computing devices is constructed based on unbalanced optimal transmission, a target function of the energy efficiency model is to obtain an optimal unbalanced optimal transmission plan and an energy efficiency modeling function, key energy efficiency feature space edge distribution and conditional distribution difference of a source computing device and a target computing device are simultaneously learned by optimizing the target function, and then valuable energy efficiency knowledge contained in energy efficiency data of the source computing device is used to improve energy efficiency modeling precision of the target computing device with only a small amount of labeled energy efficiency data;

[0013] The energy efficiency model is trained to learn the key energy efficiency feature space difference between the heterogeneous computing devices, and then the energy efficiency knowledge transfer of the effective heterogeneous computing devices is realized.

[0014] The trained energy efficiency model is used for energy efficiency modeling of the target computing device to be evaluated, and the energy efficiency performance is evaluated.

[0015] As a preferred technical solution, the energy efficiency data of the source computing device and the target computing device is obtained, specifically:

[0016] The computing device with a large amount of labeled energy efficiency historical data is set as the source computing device X S , and the energy efficiency data set of the source computing device is represented as D S ={x S ;y S}, the number of labeled samples is N s ;

[0017] The computing device with a small amount of labeled energy efficiency data is set as the target computing device X T , and the energy efficiency data set corresponding to the target computing device is represented as D T ={x T ;y T}, the number of samples is N t , wherein the labeled energy efficiency data is represented as , the number of energy efficiency data is n t ; wherein, n t <<N t <<N s ;

[0018] Let the energy efficiency feature space joint probability distribution of X S be represented as P S , and the energy efficiency feature space joint probability distribution of X T be represented as P T .

[0019] As a preferred technical solution, the energy efficiency data of the source computing device and the target computing device is preprocessed respectively to obtain the respective key energy efficiency feature space, specifically:

[0020] The key energy efficiency feature spaces of the source computing power equipment and the target computing power equipment are respectively reduced in dimension by using a principal component analysis method, and then the key energy efficiency feature spaces of the source computing power equipment and the target computing power equipment are respectively obtained.

[0021] In order to eliminate the influence of the differences in the dimensions and numerical ranges of the energy efficiency values of the labeled energy efficiency data between different computing power equipment on the energy efficiency model training, the energy efficiency values of the source computing power equipment and the target computing power equipment are respectively normalized by using Min-Max, and are expressed as:

[0022] Wherein, y represents the energy efficiency value of the computing power equipment, max(y) and min(y) respectively represent the maximum energy efficiency value and the minimum energy efficiency value in each energy efficiency data set, and y' represents the normalized energy efficiency value, and the value range is located in [0, 1].

[0023] As a preferred technical solution, the objective function of the energy efficiency model is as follows:

[0024] Wherein, f(·) represents the energy efficiency modeling function, which belongs to the reproducing Hilbert space represents the joint probability distribution of the unlabeled target computing power equipment energy efficiency data, represents the projection space of the two joint probability distributions, represents the i-th energy efficiency feature data of the source computing power equipment, represents the corresponding energy efficiency value; represents the j-th energy efficiency feature data of the target computing power equipment, represents the corresponding energy efficiency value; 1 represents the unit matrix; Γ i,j represents the unbalanced optimal transmission plan, C(·) represents a joint cost function, which contains the energy efficiency sample distance of the source computing power equipment and the target computing power equipment; Ω(·) represents a regularization term, D KL (·) represents the KL divergence, D KL (z)=zlog(z)-z; δ≥0 represents a parameter for balancing the complexity of the cost function C(·) and the energy efficiency modeling function f; λ1 and λ2 respectively represent penalty hyperparameters; the target computing power equipment training data contains n t labeled energy efficiency data is used to fit the energy efficiency modeling function f(·) in a semi-supervised setting, 1≤j≤n t as a constraint condition for the energy efficiency model training.

[0025] The computing power equipment includes a general computing power equipment with a CPU chip as a computing core and an intelligent computing power equipment with an AI chip as a computing core.

[0026] The energy efficiency model is trained to realize energy efficiency feature space difference learning between heterogeneous computing power equipment, and specifically,

[0027] A large amount of labeled energy efficiency data of the source computing power equipment and a small amount of labeled energy efficiency data of the target computing power equipment are used as training data.

[0028] f(·) is fixed, and Γ is solved first, and the optimization objective function is iteratively updated based on the Majorization-Minimization algorithm to obtain the transmission plan Γ.

[0029] The transmission plan Γ is fixed, and the target function is further represented as follows:

[0030] wherein, The predicted energy efficiency value is represented by n t j≤N t ;

[0031] Finally, the training data is used to fit the above optimization objective until the model converges or the maximum number of iterations is reached, and thus the optimal energy efficiency modeling function f(·) is obtained.

[0032] The unlabeled energy efficiency data of the target computing power equipment is input into the trained energy efficiency model to realize energy efficiency evaluation of the target computing power equipment, and the mean square error and the mean absolute error are used to quantitatively analyze the performance of the energy efficiency prediction model.

[0033] In a second aspect, the present application provides a knowledge transfer driven data center computing power energy efficiency modeling system, which is applied to the knowledge transfer driven data center computing power energy efficiency modeling method and includes a data acquisition module, a preprocessing module, an energy efficiency model construction module, an energy efficiency model training module and an energy efficiency evaluation module.

[0034] The data acquisition module is used to acquire energy efficiency data of source computing power equipment and target computing power equipment; the source computing power equipment includes a large amount of labeled energy efficiency data, and the target computing power equipment includes a small amount of labeled energy efficiency data.

[0035] The preprocessing module is used to preprocess the energy efficiency data of the source computing power equipment and the target computing power equipment respectively to obtain respective key energy efficiency feature spaces; the key energy efficiency feature space refers to the most relevant features to the energy efficiency of the computing power equipment.

[0036] The energy efficiency model construction module is configured to construct an energy efficiency model of the cross-computing power device based on the unbalanced optimal transmission, wherein a target function of the energy efficiency model is to obtain an optimal unbalanced optimal transmission plan and an energy efficiency modeling function, and by optimizing the target function, energy efficiency edge distribution and conditional distribution difference of the source computing power device and the target computing power device are simultaneously learned, and then valuable energy efficiency knowledge contained in the energy efficiency data of the source computing power device is utilized to improve energy efficiency modeling precision of the target computing power device with only a small amount of labeled energy efficiency data.

[0037] The energy efficiency model training module is configured to train the energy efficiency model, learn key energy efficiency feature space difference between the heterogeneous computing power devices, and then realize effective energy efficiency knowledge transfer of the heterogeneous computing power devices.

[0038] The energy efficiency evaluation module is configured to use the trained energy efficiency model for energy efficiency modeling of a target computing power device to be evaluated, and evaluate energy efficiency performance.

[0039] In a third aspect, the present application provides an electronic device, which comprises:

[0040] at least one processor; and

[0041] a memory in communication with the at least one processor; wherein

[0042] the memory stores computer program instructions executable by the at least one processor, and the computer program instructions are executed by the at least one processor to enable the at least one processor to execute the knowledge transfer driven data center computing power energy efficiency modeling method.

[0043] In a fourth aspect, the present application provides a computer readable storage medium storing a program, and the program is executed by a processor to implement the knowledge transfer driven data center computing power energy efficiency modeling method.

[0044] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0045] The present application is based on the actual situation of data center computing power device energy efficiency modeling, and proposes a knowledge transfer driven data center computing power energy efficiency modeling method to solve the problem of energy efficiency modeling of computing power devices lacking labeled energy efficiency data. By learning the key energy efficiency feature space distribution difference between the source computing power device and the target computing power device, the energy efficiency modeling precision of the target computing power device with only a small amount of labeled energy efficiency data is improved by utilizing valuable knowledge contained in a large amount of labeled energy efficiency historical data of the source computing power device. The present application remedies the defects of the prior art energy efficiency modeling method, improves the generalization performance of the energy efficiency model for the target computing power device, reduces the data collection cost of real data center energy efficiency modeling, and improves the usability of the energy efficiency modeling method in actual data centers. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description only represent some of the embodiments of the present application, and all other drawings obtained by those skilled in the art without creative effort based on these drawings also fall within the scope of protection of the present application.

[0047] FIG. 1 is a flow chart of a knowledge transfer driven data center computing power energy efficiency modeling method according to an embodiment of the present application;

[0048] FIG. 2 is a structural schematic diagram of a knowledge transfer driven data center computing power energy efficiency modeling system according to an embodiment of the present application;

[0049] FIG. 3 is a structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort fall within the scope of protection of the present application.

[0051] In the present application, the phrase "embodiment" means that the specific features, structures or characteristics described in conjunction with the embodiment can be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily mean the same embodiment, nor is it an independent or alternative embodiment to other embodiments. Those skilled in the art understand explicitly and implicitly that the embodiments described in the present application can be combined with other embodiments.

[0052] Referring to FIG. 1, a knowledge transfer driven data center computing power energy efficiency modeling method according to an embodiment of the present application includes the following steps:

[0053] Step 1: Set the computing power device with a large amount of labeled energy efficiency historical data as the source computing power device X S , and the corresponding energy efficiency dataset is represented as D S ={x S ;y S}, and the number of labeled samples is N s . At the same time, set the computing power device with a small amount of labeled data as the target computing power device X T , and the corresponding energy efficiency dataset is represented as D T ={x T ;y T}, the number of samples is N t , where the labeled data is represented as The number of samples is n t , and it is assumed that the energy efficiency feature space joint probability distribution of X S is represented as P S , and the energy efficiency feature space joint probability distribution of X T is represented as P T . It should be noted that n t <<N t <<N s .

[0054] Step 2: Preprocess the energy efficiency data of the source computing power device and the target computing power device respectively to obtain their respective key energy efficiency features.

[0055] Further, in order to obtain the key energy efficiency features of the source computing power device and the target computing power device while ensuring the key energy efficiency feature space of the source computing power device and the target computing power device, first, the principal component analysis (PCA) is used to reduce the dimension of the energy efficiency feature space of the source computing power device and the target computing power device. Then, the differences in energy efficiency feature space between different types of computing power devices deployed in the data center are analyzed, including but not limited to general computing power devices with CPU chips as computing cores, intelligent computing power devices with GPU, FPGA, TPU, NPU, etc. AI chips as computing cores, and the feasibility of energy efficiency knowledge transfer between different types of computing power devices is explored.

[0056] At the same time, in order to eliminate the influence of the differences in dimension and numerical range of the energy efficiency values of the labeled energy efficiency data between different computing power devices on the model training, the Min-Max normalization is used to normalize the energy efficiency values of the source computing power device and the target computing power device, represented as:

[0057] Where y represents the energy efficiency value of the computing power device, max(y) and min(y) represent the maximum and minimum energy efficiency values in each data set respectively, and y' represents the normalized label, with a value range of [0, 1].

[0058] Step 3: Construct a semi-supervised domain adaptive regression model based on unbalanced optimal transport, and its objective function can be represented as:

[0059] Where f(·) represents the energy efficiency modeling function, which belongs to the reproducing Hilbert space Γ i,j ∈Γ represents the unbalanced optimal transport plan. represents the joint probability distribution of the target computing power device in the semi-supervised setting. denotes the projection space of two joint probability distributions. denotes the i-th energy efficiency feature data of the source computing device, denotes the corresponding energy efficiency value. denotes the j-th energy efficiency feature data of the target computing device, denotes the corresponding energy efficiency value. Ω(·) denotes a regularization term, D KL (·) denotes the KL divergence, D KL (z) = z log(z) - z. δ ≥ 0 denotes a parameter used to balance the complexity of the cost function C(·) and the energy efficiency modeling function f(·). In addition, n t pieces of labeled energy efficiency data in the target server training data are used to fit the energy efficiency modeling function f(·) in a semi-supervised setting, as a constraint condition of the model. C(·) is represented as a joint cost function, which includes the energy efficiency sample distance of the source computing device and the target computing device, and a loss function L(·) measuring the difference between the corresponding energy efficiency values, which is represented as:

[0060] where q(·) denotes a function measuring the difference between the energy efficiency data of the source computing device and the energy efficiency data of the target computing device σ is a hyperparameter used to balance the difference in the energy efficiency feature space and the energy efficiency value loss L(·).

[0061] In summary, the final goal of the above objective function is to find an optimal unbalanced optimal transport plan Γ and an energy efficiency modeling function f(·). By optimizing the above objective function, the energy efficiency marginal distribution and conditional distribution difference of the source computing device and the target computing device are simultaneously learned, and the valuable energy efficiency knowledge contained in the large amount of energy efficiency data of the source computing device is used to improve the energy efficiency modeling accuracy of the target computing device with only a small amount of sample data.

[0062] Step 4: Use the large amount of labeled data of the source computing device and the small amount of labeled data of the target computing device as training data to train the semi-supervised domain adaptation regression model constructed in step 3, so as to obtain the optimal transport plan Γ and the energy efficiency modeling function f(·). Under the given training data set, since the objective function defined in step 3 has smoothness, in the process of model training, the method of fixing one and solving the other can be adopted. First, f(·) can be fixed, and Γ can be solved first. Based on the Majorization-Minimization algorithm, the optimization objective function in step 3 is iteratively updated, which is represented as follows:

[0063] wherein diag(·) represents a matrix diagonal element extraction operation, represents a matrix multiplication operation, exp(·) represents an exponential function form. m represents an m-dimensional unit matrix. n represents an n-dimensional unit matrix. Let λ = λ1 = λ2. k represents the number of iterations.

[0064] When the transmission plan Γ is fixed, the objective function constructed in step 3 can be further represented as follows:

[0065] wherein, represents a predicted energy efficiency value.

[0066] Further, the above objective function is rewritten in a kernel form, represented as:

[0067] wherein k(·) represents a kernel function. j ω represents a weight. 1≤j≤n t As a constraint model training constraint condition.

[0068] In addition, the above objective function is rewritten in a Lagrange form as:

[0069] wherein, In addition, ρ represents a parameter, and K is represented as:

[0070] By setting the first derivative of the Lagrange form objective function to 0, the optimal parameter can be obtained, represented as:

[0071] Step 5: The unmarked energy efficiency data of the target computing power equipment is input into the trained model, so as to realize the energy efficiency evaluation of the target computing power equipment, and the performance of the energy efficiency prediction model is quantitatively analyzed by using MSE and MAE, and the calculation formula is as follows:

[0072] wherein M represents the number of test samples; p i and respectively represent the true value and the prediction of the target computing power equipment.

[0073] In another embodiment of the present application, the feasibility of the present application is verified, mainly including the following two steps:

[0074] Step 1: Experimental setting, specifically as follows:

[0075] Dataset selection: The embodiment is carried out on the historical dataset of a real data center SURFsara. The embodiment selects the energy efficiency data of six CPU architecture computing power devices and three GPU architecture computing power devices for three consecutive days for experiment. During the embodiment, it is assumed that the labeled energy efficiency data of the source computing power device has a scale of 15,000, and the energy efficiency data of the target computing power device has a scale of 500, of which only 10 pieces of labeled energy efficiency data are available. The specific device information is shown in Table 1.

[0076] Table 1. Information of computing power devices for experiment

[0077] Comparison method: In the embodiment, the method proposed in the application is compared with four popular traditional ML energy efficiency modeling methods (namely Linear Regression, SVM, AdaBoost and Random Forest) and two energy efficiency modeling methods based on transfer learning (namely JDOT and DARE-GRAM).

[0078] Simulation setting: The embodiment sets up cross-computing power device energy efficiency modeling performance analysis in two scenarios (namely cross-isomorphic computing power device energy efficiency modeling and cross-heterogeneous computing power device energy efficiency modeling). For cross-isomorphic computing power device energy efficiency modeling, the embodiment verifies the consideration between CPU architecture computing power devices. For cross-heterogeneous computing power device energy efficiency modeling, the embodiment considers the energy efficiency between CPU architecture computing power devices and GPU architecture computing power devices.

[0079] Step 2: Performance comparison, the specific content is as follows:

[0080] The embodiment shows the performance comparison of the method proposed in the application and the cross-computing power device energy efficiency modeling based on traditional ML methods. It can be seen that the method proposed in the application shows excellent performance under different task setting types. The traditional ML methods show poor results due to the difficulty in dealing with the differences in energy efficiency feature spaces between different computing power devices. The comparison results are shown in Table 2.

[0081] Table 2. Performance comparison of cross-computing power device energy efficiency modeling based on traditional ML methods in the embodiment

[0082] The embodiment shows the performance comparison of the method proposed in the application and the cross-isomorphic computing power device energy efficiency modeling based on transfer learning methods. It can be seen that the method proposed in the application is obviously superior to other comparison methods in the cross-isomorphic computing power device energy efficiency modeling task. The comparison results are shown in Table 3.

[0083] Table 3 Performance comparison of the embodiment and the cross-isomorphic computing power device energy efficiency modeling method based on transfer learning

[0084] The embodiment shows the performance comparison of the method and the cross-heterogeneous computing power device energy efficiency modeling method based on transfer learning. It can be seen that the method is obviously superior to other comparison methods in the cross-heterogeneous computing power device energy efficiency modeling task, which further illustrates the effectiveness of the method. The comparison results are shown in Table 4.

[0085] Table 4 Performance comparison of the embodiment and the cross-heterogeneous computing power device energy efficiency modeling method based on transfer learning

[0086] It should be noted that for the foregoing method embodiments, in order to facilitate description, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the present application, certain steps can be performed in other order or simultaneously.

[0087] Based on the same idea as the knowledge transfer driven data center computing power energy efficiency modeling method in the above embodiment, the present application also provides a knowledge transfer driven data center computing power energy efficiency modeling system, which can be used to execute the knowledge transfer driven data center computing power energy efficiency modeling method. For the convenience of description, in the structural schematic diagram of the embodiment of the knowledge transfer driven data center computing power energy efficiency modeling system, only the part related to the embodiment of the present application is shown, and those skilled in the art can understand that the illustrated structure does not constitute a limitation on the device, which can include more or fewer components than the illustrated, or combine certain components, or different component arrangement.

[0088] Please refer to FIG. 2, in another embodiment of the present application, a knowledge transfer driven data center computing power energy efficiency modeling system 100 is provided, which comprises a data acquisition module 101, a preprocessing module 102, an energy efficiency model construction module 103, an energy efficiency model training module 104 and an energy efficiency evaluation module 105;

[0089] The data acquisition module 101 is configured to acquire energy efficiency data of source computing power devices and target computing power devices; the source computing power devices include a large amount of labeled energy efficiency data, and the target computing power devices include a small amount of labeled energy efficiency data;

[0090] The preprocessing module 102 is configured to preprocess the energy efficiency data of the source computing power devices and the target computing power devices respectively to obtain respective key energy efficiency feature spaces;

[0091] The energy efficiency model construction module 103 is configured to construct an energy efficiency model of the cross-computing power device based on the unbalanced optimal transmission, a target function of the energy efficiency model is to obtain an optimal unbalanced optimal transmission plan and an energy efficiency modeling function, and by optimizing the target function, the energy efficiency edge distribution and the conditional distribution difference of the source computing power device and the target computing power device are simultaneously learned, and then the valuable energy efficiency knowledge contained in the energy efficiency data of the source computing power device is used to improve the energy efficiency modeling precision of the target computing power device with only a small amount of sample data.

[0092] The energy efficiency model training module 104 is configured to train the energy efficiency model, learn the energy efficiency feature space difference between the heterogeneous computing power devices, and then realize the energy efficiency knowledge transfer of the effective heterogeneous computing power devices.

[0093] The energy efficiency evaluation module 105 is configured to use the trained energy efficiency model for energy efficiency modeling of the target computing power device to be evaluated, and evaluate the energy efficiency performance.

[0094] It should be noted that the knowledge transfer driven data center computing power energy efficiency modeling system of the present application corresponds to the knowledge transfer driven data center computing power energy efficiency modeling method of the present application, and the technical features and advantages described in the embodiment of the knowledge transfer driven data center computing power energy efficiency modeling method are applicable to the embodiment of the knowledge transfer driven data center computing power energy efficiency modeling system, and the specific content can be referred to the description in the method embodiment, which will not be repeated here, and hereby declared.

[0095] In addition, in the embodiment of the knowledge transfer driven data center computing power energy efficiency modeling system of the above embodiment, the logical division of each program module is only an example, and in actual application, the above function allocation can be completed by different program modules according to the needs, for example, the configuration requirements of the corresponding hardware or the convenience of software implementation, that is, the internal structure of the knowledge transfer driven data center computing power energy efficiency modeling system is divided into different program modules to complete all or part of the functions described above.

[0096] Please refer to FIG. 3, in one embodiment, an electronic device for implementing the knowledge transfer driven data center computing power energy efficiency modeling method is provided, the electronic device 200 can include a first processor 201, a first memory 202 and a bus, and can further include a computer program stored in the first memory 202 and executable on the first processor 201, such as a knowledge transfer driven data center computing power energy efficiency modeling program 203.

[0097] The first memory 202 includes at least one type of readable storage medium, such as flash memory, a mobile hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a magnetic memory, a disk, an optical disk, etc. In some embodiments, the first memory 202 can be an internal storage unit of the electronic device 200, such as a mobile hard disk of the electronic device 200. In other embodiments, the first memory 202 can also be an external storage device of the electronic device 200, such as a plug-in mobile hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the first memory 202 can include both an internal storage unit and an external storage device of the electronic device 200. The first memory 202 can be used to store application software and various data installed in the electronic device 200, such as the code of the knowledge transfer driven data center computing power and energy efficiency modeling program 203, and can also be used to temporarily store data that has been output or will be output.

[0098] The first processor 201 can be composed of an integrated circuit in some embodiments, such as a single packaged integrated circuit, or a plurality of packaged integrated circuits with the same function or different functions, including one or more combinations of a central processing unit (CPU), a microprocessor, a digital processing chip, a graphics processor, and various control chips, etc. The first processor 201 is the control unit of the electronic device, which connects various components of the entire electronic device through various interfaces and lines, and executes various functions and processes data of the electronic device 200 by running or executing programs or modules stored in the first memory 202 and calling data stored in the first memory 202.

[0099] FIG. 3 only shows an electronic device with components, and those skilled in the art can understand that the structure shown in FIG. 3 does not constitute a limitation on the electronic device 200, and can include fewer or more components than shown, or combine certain components, or different component arrangements.

[0100] The knowledge transfer driven data center computing power and energy efficiency modeling program 203 stored in the first memory 202 of the electronic device 200 is a combination of a plurality of instructions, which, when running in the first processor 201, can achieve:

[0101] The energy efficiency data of the source computing power device and the target computing power device are obtained; the source computing power device includes a large amount of labeled energy efficiency data, and the target computing power device includes a small amount of labeled energy efficiency data;

[0102] The energy efficiency data of the source computing power device and the target computing power device are respectively preprocessed to obtain respective key energy efficiency feature spaces;

[0103] An energy efficiency model across computing power devices is constructed based on unbalanced optimal transmission, a target function of the energy efficiency model is to obtain an optimal unbalanced optimal transmission plan and an energy efficiency modeling function, energy efficiency edge distribution and conditional distribution difference of the source computing power device and the target computing power device are simultaneously learned by optimizing the target function, and then valuable energy efficiency knowledge contained in the energy efficiency data of the source computing power device is used to improve energy efficiency modeling precision of the target computing power device with only a small amount of sample data;

[0104] The energy efficiency model is trained to learn energy efficiency feature space difference between heterogeneous computing power devices, and then effective energy efficiency knowledge migration of the heterogeneous computing power devices is realized;

[0105] The trained energy efficiency model is used for energy efficiency modeling of a target computing power device to be evaluated, and energy efficiency performance is evaluated.

[0106] Further, the modules / units of the electronic device 200 are implemented in the form of software function units and sold or used as independent products, which can be stored in a nonvolatile computer readable storage medium. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory).

[0107] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer readable storage medium, and when the program is executed, the processes of the above-mentioned embodiment methods can be included. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0108] Any combination of the technical features of the above embodiments can be made. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0109] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations and simplifications of the embodiments of the present application without departing from the spirit and principles of the present application are equivalent replacement methods, and are included in the protection scope of the present application.

Claims

1. A knowledge transfer-driven data center computing power energy efficiency modeling method, characterized in that, Includes the following steps: Acquire energy efficiency data of source computing power devices and target computing power devices; the source computing power devices include a large amount of labeled energy efficiency data, and the target computing power devices include a small amount of labeled energy efficiency data; The energy efficiency data of the source computing power device and the target computing power device are preprocessed to obtain their respective key energy efficiency feature spaces; The key energy efficiency feature space refers to the features most relevant to the energy efficiency of computing equipment. An energy efficiency model across computing devices is constructed based on unbalanced optimal transmission. The objective function of the energy efficiency model is to obtain the optimal unbalanced optimal transmission plan and the energy efficiency modeling function. By optimizing the objective function, the differences in the edge distribution and conditional distribution of key energy efficiency features of the source and target computing devices are learned simultaneously. In turn, the valuable energy efficiency knowledge contained in the energy efficiency data of the source computing device is used to improve the energy efficiency modeling accuracy of the target computing device, which has only a small amount of labeled energy efficiency data. The energy efficiency model is trained to learn the spatial differences in key energy efficiency characteristics among heterogeneous computing devices, thereby achieving effective energy efficiency knowledge transfer among heterogeneous computing devices. The trained energy efficiency model is used to model the energy efficiency of the target computing device to be evaluated, and to assess its energy efficiency performance.

2. The knowledge transfer-driven data center computing power energy efficiency modeling method according to claim 1, characterized in that, The acquisition of energy efficiency data for the source and target computing devices specifically involves: The computing power device with a large amount of labeled historical energy efficiency data is set as the source computing power device X. S The energy efficiency dataset of the source computing power device is represented as D. S ={x S ;y S The number of labeled samples is N. s ; The computing power device with a small amount of labeled energy efficiency data is set as the target computing power device X. T The energy efficiency dataset corresponding to the target computing power device is represented as D. T ={x T ;y T The sample size is N. t Among them, energy efficiency data with labels is represented as The number of energy efficiency data is n t ; where n t <<N t <<N s ; Let X S The joint probability distribution of the energy efficiency characteristic space is represented by P. S X T The joint probability distribution of the energy efficiency characteristic space is represented by P. T .

3. The knowledge transfer-driven data center computing power efficiency modeling method according to claim 1, characterized in that, The energy efficiency data of the source computing power device and the target computing power device are preprocessed to obtain their respective key energy efficiency feature spaces, specifically as follows: Principal component analysis is used to reduce the dimensionality of the key energy efficiency feature spaces of the source computing power device and the target computing power device, respectively, so as to obtain the key energy efficiency feature spaces of the source computing power device and the target computing power device. To eliminate the impact of differences in the units and numerical ranges of energy efficiency values ​​among labeled energy efficiency data from different computing power devices on the training of the energy efficiency model, the energy efficiency values ​​of the source and target computing power devices are normalized using Min-Max, as follows: Where y represents the energy efficiency value of the computing device, max(y) and min(y) represent the maximum and minimum energy efficiency values ​​in each energy efficiency dataset, respectively, and y′ represents the normalized energy efficiency value, which ranges from [0,1].

4. The knowledge transfer-driven data center computing power efficiency modeling method according to claim 1, characterized in that, The objective function of the energy efficiency model is as follows: Here, f(·) represents the energy efficiency modeling function, which belongs to the regenerative Hilbert space. This represents the joint probability distribution of energy efficiency data for unlabeled target computing devices. It represents the projection space of two joint probability distributions. This represents the i-th energy efficiency characteristic data of the source computing power device. This indicates the corresponding energy efficiency value; This represents the j-th energy efficiency characteristic data of the target computing power device. This represents the corresponding energy efficiency value; 1 represents the identity matrix; Γ i,j ∈Γ represents the unbalanced optimal transmission plan, C(·) represents a joint cost function, which includes the energy efficiency sample distance between the source and target computing devices; Ω(·) represents the regularization term, D KL (·) represents the KL divergence, denoted as D. KL (z) = zlog(z) - z; δ≥0 indicates that it is a parameter used to balance the complexity of the cost function C(·) and the energy efficiency modeling function f; λ1 and λ2 represent the penalty hyperparameters, respectively; the target computing power equipment training data contains n t Labeled energy efficiency data were used to fit the energy efficiency modeling function f(·) in a semi-supervised setting. As a constraint for training the energy efficiency model.

5. The knowledge transfer-driven data center computing power efficiency modeling method according to claim 4, characterized in that, The computing power devices include general-purpose computing power devices with CPU chips as the computing core and intelligent computing power devices with AI chips as the computing core.

6. The knowledge transfer-driven data center computing power energy efficiency modeling method according to claim 4, characterized in that, The training of the energy efficiency model to learn the spatial differences in energy efficiency characteristics among heterogeneous computing devices specifically involves: The large amount of labeled energy efficiency data from the source computing power device and the small amount of labeled energy efficiency data from the target computing power device are used as training data. By fixing f(·), we first solve for Γ, and then iteratively update the optimization objective function based on the Majorization-Minimization algorithm to obtain the transmission plan Γ. With the transmission plan Γ fixed, the objective function is further expressed in the following form: in, The predicted energy efficiency value, n t <j≤N t ; Finally, the above optimization objective is fitted using the training data until the model converges or reaches the maximum number of iterations, thus obtaining the optimal energy efficiency modeling function f(·).

7. The knowledge transfer-driven data center computing power efficiency modeling method according to claim 4, characterized in that, Unlabeled energy efficiency data of the target computing power device is input into a trained energy efficiency model to evaluate the energy efficiency of the target computing power device, and the performance of the energy efficiency prediction model is quantitatively analyzed using mean square error and mean absolute error.

8. A knowledge transfer-driven data center computing power and energy efficiency modeling system, characterized in that, The knowledge transfer-driven data center computing power energy efficiency modeling method applied to any one of claims 1-7 includes a data acquisition module, a preprocessing module, an energy efficiency model construction module, an energy efficiency model training module, and an energy efficiency evaluation module; The data acquisition module is used to acquire energy efficiency data of the source computing power device and the target computing power device; the source computing power device includes a large amount of labeled energy efficiency data, and the target computing power device includes a small amount of labeled energy efficiency data; The preprocessing module is used to preprocess the energy efficiency data of the source computing power device and the target computing power device respectively to obtain their respective key energy efficiency feature spaces. The key energy efficiency feature space refers to the features most relevant to the energy efficiency of computing equipment. The energy efficiency model construction module is used to construct an energy efficiency model across computing power devices based on unbalanced optimal transmission. The objective function of the energy efficiency model is to obtain the optimal unbalanced optimal transmission plan and the energy efficiency modeling function. By optimizing the objective function, the differences in the energy efficiency edge distribution and conditional distribution of the source computing power device and the target computing power device are learned simultaneously. In turn, the valuable energy efficiency knowledge contained in the energy efficiency data of the source computing power device is used to improve the energy efficiency modeling accuracy of the target computing power device, which has only a small amount of labeled energy efficiency data. The energy efficiency model training module is used to train the energy efficiency model, learn the key energy efficiency feature spatial differences between heterogeneous computing devices, and thus realize effective energy efficiency knowledge transfer between heterogeneous computing devices. The energy efficiency assessment module is used to apply the trained energy efficiency model to the energy efficiency modeling of the target computing power device to be assessed, and to evaluate the energy efficiency performance.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores computer program instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the knowledge migration-driven data center computing power and energy efficiency modeling method as described in any one of claims 1-7.

10. A computer-readable storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the knowledge transfer-driven data center computing power energy efficiency modeling method according to any one of claims 1-7.

Images