A cluster computing power equipment management method and device based on deep learning
By using a parallel echo state network model of two reservoirs and a mutual information sparse screening mechanism, combined with recursive least squares online updates, the shortcomings of state modeling and parameter updates in the management of cluster computing power devices are solved, and a clear representation of device status and stability and adaptability of management decisions are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 北京深启科技有限公司
- Filing Date
- 2026-02-26
- Publication Date
- 2026-06-09
AI Technical Summary
Existing cluster computing power equipment management methods are difficult to effectively characterize the equipment operating status in large-scale, dynamically changing environments, and existing models are difficult to continuously adjust according to real-time status changes, resulting in insufficient stability and adaptability of management decisions.
An echo state network model with two parallel reservoirs and a mutual information sparse screening mechanism are adopted to separate and model the computing power status and energy consumption and heat status of the cluster computing power equipment. The output layer parameters are adjusted hourly through a recursive least squares online update mechanism to build an intelligent management process.
This improves the clarity of state representation for cluster computing devices, reduces dynamic state interference, enhances the adaptability of the management process, and reduces computational complexity.
Smart Images
Figure CN122174633A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning-based cluster computing power equipment operation status modeling and intelligent management technology, and in particular to a deep learning-based cluster computing power equipment management method and apparatus. Background Technology
[0002] With the continuous development of cloud computing, artificial intelligence, and big data technologies, large-scale cluster computing power devices are widely deployed in scenarios such as data processing, model training, inference services, and complex business scheduling. Cluster computing power devices typically consist of multiple heterogeneous computing nodes connected via a network to form a collaborative operating environment. Their operational status is influenced by various factors, including task load, resource allocation, energy consumption, and thermal status. Therefore, effectively managing the operational status of cluster computing power devices has become a crucial issue for ensuring stable system operation and rational resource allocation.
[0003] Existing cluster computing power device management methods mostly rely on rule-driven or statistical feature-based monitoring and analysis methods, which judge and control the device's operating status through preset thresholds or empirical rules. These methods are applicable to certain scenarios with small system scale or relatively stable operating modes. However, in large-scale, dynamically changing cluster environments, the device's operating status exhibits obvious temporal correlation and nonlinear characteristics, and it is difficult to fully characterize the state evolution process by simply relying on static rules or linear models.
[0004] Some technical solutions attempt to introduce machine learning or deep learning models to model cluster states. However, existing methods often mix and match multiple state indicators such as computing power utilization, energy consumption, and temperature, failing to structurally distinguish the dynamic characteristics of states from different sources. This can easily lead to aliasing of state representations, affecting the stability of subsequent management decisions. Furthermore, existing models mostly employ offline training or batch updates, making it difficult to continuously adjust model parameters based on real-time state changes during device operation. This limits the model's adaptability to changes in the operating environment.
[0005] Therefore, existing cluster computing power device management technologies still have shortcomings in terms of state modeling structure design, state filtering mechanism, and model parameter update method. There is an urgent need for a management method and device that can separate modeling for different state sources and support continuous updates during operation. Summary of the Invention
[0006] One objective of this invention is to propose a cluster computing power device management method and device based on deep learning. This invention introduces an echo state network model of two parallel reservoirs and a mutual information sparse screening mechanism to perform separate modeling and state reorganization on the computing power state sequence and energy consumption heat state sequence of the cluster computing power device. It also combines a recursive least squares online update mechanism to adjust the output layer parameters hourly, constructing an intelligent management process that continuously evolves with the operating state, and realizing the continuous generation and dynamic updating of cluster computing power device management decisions.
[0007] A method and apparatus for managing cluster computing power devices based on deep learning according to an embodiment of the present invention includes the following steps:
[0008] S1. Collect operational status data from the cluster's computing power devices. This data includes a computing power utilization rate data sequence and energy consumption and temperature data sequences. S2. Perform time alignment processing on the computing power utilization rate data sequence and the energy consumption and temperature data sequences to form a computing power status input sequence and an energy consumption and thermal status input sequence. S3. Construct an echo state network model with two parallel reservoirs. The first reservoir receives the computing power status input sequence, and the second reservoir receives the energy consumption and thermal status input sequence. S4. Perform recursive mapping in the first and second reservoirs respectively to generate a computing power reservoir state matrix and an energy consumption and thermal reservoir state matrix. S5. Analyze the computing power reservoir state matrix and the energy consumption and thermal status input sequence. The state matrix of the hot water reservoir is expanded to construct a joint state component set; S6, the mutual information values between each state component in the joint state component set and the management output variables are calculated; S7, the state components whose mutual information values satisfy the preset sparse constraints are retained to form a sparse fused state vector; S8, an output layer linear mapping structure is constructed based on the sparse fused state vector, which contains parameter subsets corresponding to the first and second reservoirs; S9, the parameter subsets are updated online using a recursive least squares algorithm, and the cluster computing power equipment management decision is generated and output based on the updated output layer linear mapping structure.
[0009] Optionally, S3 specifically includes:
[0010] Within the same echo state network framework, a first reservoir and a second reservoir are configured in parallel. They are topologically independent, with no direct connection path at the reservoir level between them. The first reservoir establishes an input connection only with the computing power state input sequence, while the second reservoir establishes an input connection only with the energy consumption and heat state input sequence. The computing power state input sequence is not transmitted to the second reservoir, and the energy consumption and heat state input sequence is not transmitted to the first reservoir, thus achieving structural isolation of state sources at the input level. Each of the first and second reservoirs contains several reservoir nodes, which form an internal recursive connection structure through fixed random connections. This internal recursive connection structure remains unchanged during model operation, and the internal connection weights do not participate in parameter updates during the training phase. The first and second reservoirs exhibit structural differences in at least one of the following parameters: reservoir node configuration, internal connection sparsity parameter, or spectral radius parameter. The system exhibits different time response characteristics for dynamic computing power status and dynamic energy consumption / heat status. During the operation of the echo state network model with two parallel reservoirs, the first reservoir independently performs state recursion operations based on the computing power status input sequence to generate a computing power reservoir state matrix, while the second reservoir independently performs state recursion operations based on the energy consumption / heat status input sequence to generate an energy consumption / heat reservoir state matrix. The computing power reservoir state matrix and the energy consumption / heat reservoir state matrix do not undergo merging, weighting, or cross-mapping operations at the reservoir level. In the parallel dual-reservoir structure, a cross-reservoir state fusion path is introduced only before the output layer is constructed. This path consists of state components selected by the mutual information filtering step; reservoir state components that do not pass the mutual information filtering step do not enter the cross-reservoir fusion path. Through structural configuration, the first and second reservoirs independently model different state sources at the reservoir level and form a state input structure oriented towards management output through a restricted cross-reservoir fusion path before the output layer.
[0011] Optionally, the construction of the joint state component set in step S5 specifically includes:
[0012] The state matrices of the computing power reservoir and the energy consumption hot water reservoir are read according to the time index order, and the corresponding reservoir state vector is extracted at each time index position. The state vectors from the computing power reservoir state matrix and the energy consumption hot water reservoir state matrix at the same time index position are recorded as independent state components. In the time dimension, the state components corresponding to each time index position are arranged and aggregated to form a joint state component set containing the computing power reservoir state components and the energy consumption hot water reservoir state components. Each state component in the joint state component set retains its source identifier and time index information.
[0013] Optionally, S7 specifically includes:
[0014] S71. After obtaining the mutual information value of each state component in the joint state component set, the mutual information value of each state component is compared with the preset sparse constraint condition one by one. The preset sparse constraint condition limits the number or proportion of state components entering the state fusion process. Only when the mutual information value meets the preset sparse constraint condition is the corresponding state component marked as a valid state component.
[0015] S72. Perform a removal operation on state components that do not meet the preset sparse constraints. Class state components do not participate in the state vector construction process, thereby forming a sparse state component subset that contains only valid state components in the joint state component set.
[0016] S73. In the sparse state component subset, retain the source reservoir identifier and time index identifier for each effective state component.
[0017] S74. At the same time index position, read the valid state component of the corresponding time index identifier and perform an ordered arrangement operation on the state component from the computing power reservoir and the state component from the energy consumption hot water reservoir according to the preset reservoir order rules. The ordered arrangement operation only changes the organization order of the state components and does not change the numerical content of the state components.
[0018] S75. Combine the effective state components arranged in an ordered manner at the same time index position to form a sparse state vector corresponding to the time index position. The dimension of the sparse state vector is determined by the number of effective state components. Unselected state components do not occupy a position in the sparse state vector.
[0019] S76. In the time dimension, according to the time index from small to large, perform sequential splicing and recording operations on the sparse state vectors corresponding to each time index position to form a sparse fused state vector covering several time index positions.
[0020] The sparse fusion state vector formed by steps S71-S76 contains only state components that satisfy the mutual information sparse constraint condition, and preserves the source reservoir information and time sequence information of the state components.
[0021] Optionally, the state fusion process in step S7 specifically includes:
[0022] After comparing each state component in the joint state component set with the preset sparse constraints, the state components whose mutual information values satisfy the preset sparse constraints are selected as the state fusion input set. At each time index position, the state component at the corresponding time index position is read from the state fusion input set, and the state components are grouped and recorded according to the source reservoir type.
[0023] Within the same time index position, the state components from the computing power reservoir and the state components from the energy consumption hot water reservoir are arranged sequentially according to the preset reservoir order rules to form a time slice state sequence; a vectorization combination operation is performed on the time slice state sequence to generate a fused state vector corresponding to the time index position;
[0024] In the time dimension, sequential aggregation operations are performed on the fused state vectors corresponding to each time index position according to the time index order to form a state fusion result covering several time index positions; in the state fusion process, no numerical weighting, summation or mapping operations are performed between state components.
[0025] Optionally, S8 specifically includes:
[0026] After obtaining the sparse fused state vector, a linear mapping structure for the output layer is constructed based on the sparse fused state vector. The state components in the sparse fused state vector are distinguished according to the source reservoir type, and divided into a set of state components corresponding to the first reservoir and a set of state components corresponding to the second reservoir. Parameter subsets are configured for different sets of state components. A mapping relationship is established between the first parameter subset and the state component set of the first reservoir, and a mapping relationship is established between the second parameter subset and the state component set of the second reservoir. By performing linear combination operations on each parameter subset and the corresponding set of state components, a structured parameter configuration for the output layer is formed.
[0027] Optionally, S9 specifically includes:
[0028] After completing the construction of the linear mapping structure of the output layer, a recursive least squares online update structure is established for the first parameter subset of the first reservoir state component set and the second parameter subset of the second reservoir state component set in the output layer, respectively. The recursive least squares online update structure is kept in a continuously enabled state during the model operation.
[0029] When each time index position is reached, the sparse fusion state vector corresponding to the time index position is read, and the management output variable corresponding to the time index position is read synchronously. The sparse fusion state vector is used as the update input of the current time index position, and the management output variable is used as the update reference of the current time index position, so that the update process and the state generation process maintain a one-to-one correspondence on the time index.
[0030] Using the data at the current time index position as a single update sample, recursive least squares update calculations are performed on the first parameter subset and the second parameter subset respectively. The update calculation is based only on the data at the current time index position and the parameter state and covariance information retained at the previous time index position, without introducing the data sequence of historical time index positions for batch backtracking calculations.
[0031] During the recursive least squares update process, the corresponding gain vectors are calculated for the first and second parameter subsets respectively, and the parameters are corrected based on the gain vectors so that the parameter correction is completed at the current time index position. After the parameter correction is completed, the updated parameter subset is retained as the initial parameter state for the recursive update at the next time index position.
[0032] After the parameter update is completed at any time index position, the output layer linear mapping structure immediately performs a linear mapping operation on the sparse fused state vector at the time index position using the updated first parameter subset and second parameter subset, and manages the output to be continuously generated as the time index advances;
[0033] During the online update process, the update processes of the first parameter subset and the second parameter subset are executed independently. No coupling relationship is established across parameter subsets during the parameter update phase. Moreover, the parameter update operation is performed only once at any time index position. The entire update process proceeds continuously along the time index direction, thus forming a time-index-based online parameter update mechanism based on the recursive least squares algorithm.
[0034] The beneficial effects of this invention are:
[0035] (1) By constructing an echo state network model of two reservoirs in parallel, the computing power state and energy consumption and heat state are separated and modeled at the structural level to avoid dynamic interference between different states and improve the clarity of state representation.
[0036] (2) Introduce a sparse constraint screening mechanism based on mutual information to select state components from the joint state components that satisfy the constraint conditions with the correlation degree of the management output variable, thereby reducing redundant states participating in subsequent calculations;
[0037] (3) By constructing sparse fusion state vectors, the orderly organization of states across reservoirs can be achieved while maintaining the source and temporal order information of states;
[0038] (4) The output layer adopts a parameter subset partitioning structure and combines the recursive least squares algorithm to perform online updates according to the time index, so that the model parameters are continuously adjusted with the running process, enhancing the management process's adaptability to changes in the running state;
[0039] (5) The overall method completes state modeling and parameter updates without changing the internal structure of the reservoir, reducing the computational complexity during model operation. Attached Figure Description
[0040] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0041] Figure 1 This is a flowchart of a cluster computing power device management method and apparatus based on deep learning proposed in this invention;
[0042] Figure 2 This is a schematic diagram of a dual-reservoir parallel echo state network and mutual information sparse fusion structure of a cluster computing power device management method and device based on deep learning proposed in this invention.
[0043] Figure 3 This is a schematic diagram of the online update structure of the output layer based on the recursive least squares algorithm of the cluster computing power device management method and device based on deep learning proposed in this invention. Detailed Implementation
[0044] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0045] refer to Figure 1-3 A method and apparatus for managing cluster computing power devices based on deep learning, comprising the following steps:
[0046] S1. Collect operational status data from the cluster's computing power devices. This data includes a computing power utilization rate data sequence and energy consumption and temperature data sequences. S2. Perform time alignment processing on the computing power utilization rate data sequence and the energy consumption and temperature data sequences to form a computing power status input sequence and an energy consumption and thermal status input sequence. S3. Construct an echo state network model with two parallel reservoirs. The first reservoir receives the computing power status input sequence, and the second reservoir receives the energy consumption and thermal status input sequence. S4. Perform recursive mapping in the first and second reservoirs respectively to generate a computing power reservoir state matrix and an energy consumption and thermal reservoir state matrix. S5. Analyze the computing power reservoir state matrix and the energy consumption and thermal status input sequence. The state matrix of the hot water reservoir is expanded to construct a joint state component set; S6, the mutual information values between each state component in the joint state component set and the management output variables are calculated; S7, the state components whose mutual information values satisfy the preset sparse constraints are retained to form a sparse fused state vector; S8, an output layer linear mapping structure is constructed based on the sparse fused state vector, which contains parameter subsets corresponding to the first and second reservoirs; S9, the parameter subsets are updated online using a recursive least squares algorithm, and the cluster computing power equipment management decision is generated and output based on the updated output layer linear mapping structure.
[0047] In this embodiment, S3 specifically includes:
[0048] Within the same echo state network framework, a first reservoir and a second reservoir are configured in parallel. They are topologically independent, with no direct connection path at the reservoir level between them. The first reservoir establishes an input connection only with the computing power state input sequence, while the second reservoir establishes an input connection only with the energy consumption and heat state input sequence. The computing power state input sequence is not transmitted to the second reservoir, and the energy consumption and heat state input sequence is not transmitted to the first reservoir, thus achieving structural isolation of state sources at the input level. Each of the first and second reservoirs contains several reservoir nodes, which form an internal recursive connection structure through fixed random connections. This internal recursive connection structure remains unchanged during model operation, and the internal connection weights do not participate in parameter updates during the training phase. The first and second reservoirs exhibit structural differences in at least one of the following parameters: reservoir node configuration, internal connection sparsity parameter, or spectral radius parameter. The system exhibits different time response characteristics for dynamic computing power status and dynamic energy consumption / heat status. During the operation of the echo state network model with two parallel reservoirs, the first reservoir independently performs state recursion operations based on the computing power status input sequence to generate a computing power reservoir state matrix, while the second reservoir independently performs state recursion operations based on the energy consumption / heat status input sequence to generate an energy consumption / heat reservoir state matrix. The computing power reservoir state matrix and the energy consumption / heat reservoir state matrix do not undergo merging, weighting, or cross-mapping operations at the reservoir level. In the parallel dual-reservoir structure, a cross-reservoir state fusion path is introduced only before the output layer is constructed. This path consists of state components selected by the mutual information filtering step; reservoir state components that do not pass the mutual information filtering step do not enter the cross-reservoir fusion path. Through structural configuration, the first and second reservoirs independently model different state sources at the reservoir level and form a state input structure oriented towards management output through a restricted cross-reservoir fusion path before the output layer.
[0049] In this embodiment, the recursive mapping in step S4 specifically includes:
[0050] Reservoir state update equations are established in the first and second reservoirs respectively. At each time index position, the corresponding input sequence and the reservoir state vector at the previous time index position are used together as the recursive calculation input.
[0051] A linear mapping is performed on the input sequence by the input weight matrix, and a linear mapping is performed on the reservoir state vector at the previous time index by the internal connection weight matrix. The results of the two linear mappings are then summed.
[0052] A non-linear activation function is applied to the summation result to generate the reservoir state vector at the current time index position. The recursive mapping result is continuously updated in the time dimension as the input sequence changes.
[0053] During the recursive mapping process, the first reservoir and the second reservoir independently perform state update calculations without establishing a state transfer path or weight coupling relationship across reservoirs. The input weight matrix and internal connection weight matrix used in the recursive mapping remain fixed during the operation of the echo state network model of the two reservoirs in parallel, thus forming the state matrix of the computing power reservoir and the state matrix of the energy consumption hot water reservoir.
[0054] In this embodiment, the construction of the joint state component set in step S5 specifically includes:
[0055] The state matrices of the computing power reservoir and the energy consumption hot water reservoir are read according to the time index order, and the corresponding reservoir state vector is extracted at each time index position. The state vectors from the computing power reservoir state matrix and the energy consumption hot water reservoir state matrix at the same time index position are recorded as independent state components. In the time dimension, the state components corresponding to each time index position are arranged and aggregated to form a joint state component set containing the computing power reservoir state components and the energy consumption hot water reservoir state components. Each state component in the joint state component set retains its source identifier and time index information.
[0056] In this embodiment, S6 specifically includes:
[0057] For each state component in the joint state component set, extract the numerical sequence that is aligned with the management output variable at the same time index position to form the state component sequence and the management output sequence;
[0058] The state component sequence and the management output sequence are subjected to interval quantization to obtain the discrete value set of the state component and the discrete value set of the management output. The joint occurrence frequency of any discrete value of the state component and any discrete value of the management output at each time index position is counted to form a joint frequency matrix. The joint frequency matrix is then normalized to obtain the joint distribution.
[0059] The marginal distributions of the state components and the marginal distributions of the management output are obtained from the joint distribution. The logarithmic ratio term is calculated for each discrete pair of values in the joint distribution and a weighted summation operation is performed to obtain the mutual information value and write it into the mutual information list.
[0060] In this embodiment, S7 specifically includes:
[0061] S71. After obtaining the mutual information value of each state component in the joint state component set, the mutual information value of each state component is compared with the preset sparse constraint condition one by one. The preset sparse constraint condition limits the number or proportion of state components entering the state fusion process. Only when the mutual information value meets the preset sparse constraint condition is the corresponding state component marked as a valid state component.
[0062] S72. Perform a removal operation on state components that do not meet the preset sparse constraints. Class state components do not participate in the state vector construction process, thereby forming a sparse state component subset that contains only valid state components in the joint state component set.
[0063] S73. In the sparse state component subset, retain the source reservoir identifier and time index identifier for each effective state component.
[0064] S74. At the same time index position, read the valid state component of the corresponding time index identifier and perform an ordered arrangement operation on the state component from the computing power reservoir and the state component from the energy consumption hot water reservoir according to the preset reservoir order rules. The ordered arrangement operation only changes the organization order of the state components and does not change the numerical content of the state components.
[0065] S75. Combine the effective state components arranged in an ordered manner at the same time index position to form a sparse state vector corresponding to the time index position. The dimension of the sparse state vector is determined by the number of effective state components. Unselected state components do not occupy a position in the sparse state vector.
[0066] S76. In the time dimension, according to the time index from small to large, perform sequential splicing and recording operations on the sparse state vectors corresponding to each time index position to form a sparse fused state vector covering several time index positions.
[0067] The sparse fusion state vector formed by steps S71-S76 contains only state components that satisfy the mutual information sparse constraint condition, and preserves the source reservoir information and time sequence information of the state components.
[0068] In this embodiment, the state fusion process in step S7 specifically includes:
[0069] After comparing each state component in the joint state component set with the preset sparse constraints, the state components whose mutual information values satisfy the preset sparse constraints are selected as the state fusion input set. At each time index position, the state component at the corresponding time index position is read from the state fusion input set, and the state components are grouped and recorded according to the source reservoir type.
[0070] Within the same time index position, the state components from the computing power reservoir and the state components from the energy consumption hot water reservoir are arranged sequentially according to the preset reservoir order rules to form a time slice state sequence; a vectorization combination operation is performed on the time slice state sequence to generate a fused state vector corresponding to the time index position;
[0071] In the time dimension, sequential aggregation operations are performed on the fused state vectors corresponding to each time index position according to the time index order to form a state fusion result covering several time index positions; in the state fusion process, no numerical weighting, summation or mapping operations are performed between state components.
[0072] In this embodiment, S8 specifically includes:
[0073] After obtaining the sparse fused state vector, a linear mapping structure for the output layer is constructed based on the sparse fused state vector. The state components in the sparse fused state vector are distinguished according to the source reservoir type, and divided into a set of state components corresponding to the first reservoir and a set of state components corresponding to the second reservoir. Parameter subsets are configured for different sets of state components. A mapping relationship is established between the first parameter subset and the state component set of the first reservoir, and a mapping relationship is established between the second parameter subset and the state component set of the second reservoir. By performing linear combination operations on each parameter subset and the corresponding set of state components, a structured parameter configuration for the output layer is formed.
[0074] In this embodiment, S9 specifically includes:
[0075] After completing the construction of the linear mapping structure of the output layer, a recursive least squares online update structure is established for the first parameter subset of the first reservoir state component set and the second parameter subset of the second reservoir state component set in the output layer, respectively. The recursive least squares online update structure is kept in a continuously enabled state during the model operation.
[0076] When each time index position is reached, the sparse fusion state vector corresponding to the time index position is read, and the management output variable corresponding to the time index position is read synchronously. The sparse fusion state vector is used as the update input of the current time index position, and the management output variable is used as the update reference of the current time index position, so that the update process and the state generation process maintain a one-to-one correspondence on the time index.
[0077] Using the data at the current time index position as a single update sample, recursive least squares update calculations are performed on the first parameter subset and the second parameter subset respectively. The update calculation is based only on the data at the current time index position and the parameter state and covariance information retained at the previous time index position, without introducing the data sequence of historical time index positions for batch backtracking calculations.
[0078] During the recursive least squares update process, the corresponding gain vectors are calculated for the first and second parameter subsets respectively, and the parameters are corrected based on the gain vectors so that the parameter correction is completed at the current time index position. After the parameter correction is completed, the updated parameter subset is retained as the initial parameter state for the recursive update at the next time index position.
[0079] After the parameter update is completed at any time index position, the output layer linear mapping structure immediately performs a linear mapping operation on the sparse fused state vector at the time index position using the updated first parameter subset and second parameter subset, and manages the output to be continuously generated as the time index advances;
[0080] During the online update process, the update processes of the first parameter subset and the second parameter subset are executed independently. No coupling relationship is established across parameter subsets during the parameter update phase. Moreover, the parameter update operation is performed only once at any time index position. The entire update process proceeds continuously along the time index direction, thus forming a time-index-based online parameter update mechanism based on the recursive least squares algorithm.
[0081] In this embodiment, the gain vector in step S9 specifically includes:
[0082] In recursive least squares online updates, the gain vector is used to convert the input information at the current time index into parameter correction values. To maintain consistency with the output layer parameter subset structure, this invention calculates the gain vectors for the first and second parameter subsets respectively, with the two calculation processes being independent of each other in terms of input, buffer state, and update results.
[0083] Upon arrival at any time index position, the state component vectors corresponding to the first reservoir and the state component vectors corresponding to the second reservoir are extracted from the sparse fused state vector and used as the first update input and the second update input, respectively. Simultaneously, the updated state information retained at the previous time index position is read from the cache. This updated state information includes at least the covariance matrix cache and forgetting factor configuration of the previous time index position. The first parameter subset and the second parameter subset maintain their respective covariance matrix caches, which are not shared.
[0084] Subsequently, within the update structure of each parameter subset, a "normalized denominator" is first calculated. This normalized denominator is obtained by combining three parts: a baseline term obtained by taking the reciprocal of the forgetting factor, a product term of the current update input and the covariance matrix cache, and an inner product term of the product term and the current update input. The normalized denominator characterizes the scale of the current update input in the sense of covariance, and it is a scalar value. After calculating the normalized denominator, the "numerator vector" is calculated. The numerator vector is obtained by performing a matrix-vector multiplication between the covariance matrix cache and the current update input. Finally, the numerator vector is divided by the normalized denominator to obtain the gain vector of the corresponding parameter subset at the current time index.
[0085] After obtaining the gain vector, it is combined with the prediction error at the current time index to form a parameter correction. The prediction error is determined by the difference between the management output variable and the predicted output of the output layer linear mapping structure at the current time index. After calculating the parameter correction, parameter write updates are performed on the corresponding parameter subset at the current time index, and the covariance matrix cache is updated synchronously. The update of the covariance matrix cache is scaled with a forgetting factor as a weight, and the correction term consisting of the gain vector and the current update input is subtracted, thus retaining the update state information for the next time index.
[0086] Example 1:
[0087] To verify the feasibility and effectiveness of this invention in practical applications, it was applied to a typical cluster computing power device operation and management scenario. In this scenario, the cluster consists of multiple computing power nodes, each of which simultaneously undertakes computing tasks and generates energy consumption and heat changes during operation. With the dynamic changes in task load, different devices within the cluster exhibit significant differences in computing power utilization, energy consumption levels, and temperature conditions. Traditional management methods based on fixed thresholds or simple statistical rules are unable to reflect the state evolution characteristics in a timely manner, easily leading to problems such as delayed management decisions and unstable state judgments. In this scenario, the method of this invention continuously collects the operating status data of the cluster computing power devices and processes the computing power utilization-related data and energy consumption and temperature-related data separately. Through time alignment operations, state data from different sources are mapped to a unified time index system, avoiding information offset caused by inconsistent sampling rhythms. Subsequently, a dual-reservoir parallel echo state network structure is constructed, enabling the computing power state sequence and energy consumption and heat state sequence to be modeled separately at the structural level, thereby avoiding dynamic interference caused by the mixing of multiple source states in the same reservoir. During operation, the computing power status input sequence is sent to the first reservoir, while the energy consumption and temperature status input sequences are sent to the second reservoir. The two types of reservoirs maintain differentiated settings in their internal structure and parameter configurations to reflect the characteristics of different state change rhythms. The reservoir states are continuously updated through a recursive mapping method, forming a reservoir state matrix that evolves over time. Based on this, state components from different reservoirs but at the same time index position are organized to construct a joint state component set, providing a unified data foundation for subsequent state selection.
[0088] For the joint state component set, this invention further calculates the mutual information value between each state component and the management output variable to measure the degree of correlation between different state components and management decisions. By setting sparse constraints, only state components whose mutual information values fall within a preset range are retained, while other state components do not participate in the subsequent fusion process, thereby reducing the impact of state redundancy on the management model. Based on the state components that meet the sparse constraints, a sparse fusion state vector is constructed according to the time index order. This state vector simultaneously retains the structural information of computing power state and energy consumption and heat state. Based on the sparse fusion state vector, this invention constructs an output layer linear mapping structure and divides the output layer parameters into multiple parameter subsets, each corresponding to reservoir state components from different sources. During cluster operation, the output layer parameters are updated step by step according to the time index using a recursive least squares algorithm. Each time, only the data at the current time index position is used to complete a parameter correction, enabling the management model to continuously adjust with changes in the cluster's operating state without interrupting operation or retraining the model. Through the above application process, it can be observed that the method of this invention shows significant advantages in terms of the stability and continuity of management decision generation when the cluster load changes frequently. In the comparative analysis, the traditional rule-based management method resulted in a wide range of fluctuations in equipment status judgment, while the method of this invention resulted in a more concentrated trend in status changes, and the management output remained consistent with the status evolution. Simultaneously, because the state components underwent mutual information filtering and sparse constraint processing, the number of state dimensions involved in the calculation was effectively controlled, and the operational burden of the management model remained within a stable range. The table below presents the comparative results of different management methods on several key indicators under the conditions of this embodiment, visually demonstrating the performance of the method of this invention during implementation through structured data.
[0089] Table 1: Comparison of Cluster Computing Power Device Management Effectiveness
[0090] Management methods Computing power utilization fluctuation range (%) Energy consumption change rate per unit time Temperature variation dispersion Management decision consistency rate (%) Status response matching rate (%) Traditional rule management 18–25 0.35–0.48 0.42–0.57 72–78 70–76 Statistical model management 12–18 0.22–0.34 0.28–0.41 80–86 79–85 Method of the present invention 6–10 0.12–0.18 0.15–0.22 90–95 91–96
[0091] Table 1 compares the performance of different cluster computing power equipment management methods on several key operational indicators. All listed indicators are derived from continuously collected and statistically analyzed results during cluster operation, reflecting the impact of different management methods on computing power status, energy consumption status, and management output stability under dynamic operating environments. The fluctuation range of computing power utilization shows that under the traditional rule-based management method, this indicator varies within a wide range, indicating significant fluctuations in the utilization of computing resources at different time index positions, and the management strategy's response to load changes is not continuous enough. The statistical model-based management method narrows the fluctuation range of computing power utilization to some extent, but a relatively significant range expansion still exists when load changes are frequent.
[0092] In contrast, the computing power utilization fluctuation range corresponding to the method of this invention is significantly convergent, with a smaller range, indicating that under the same operating conditions, the scheduling results of computing power resources are more concentrated, and the state changes have better continuity. This result is consistent with the way this invention structurally models the computing power state sequence independently, enabling computing power-related state changes to be characterized separately and participate in the generation of management decisions.
[0093] Regarding the energy consumption change rate per unit time indicator, the traditional rule-based management method exhibits a wide range of values, reflecting a lack of constraints on energy consumption changes during the management process and strong instability in energy consumption levels as task status changes. Statistical model management shows some improvement in this indicator, but its value range still exhibits significant fluctuations. When using the method of this invention, the energy consumption change rate per unit time remains within a relatively concentrated range, indicating that during the management decision generation process, the energy consumption change trend can be continuously incorporated into the state modeling and parameter update process, thereby maintaining the continuity of energy consumption changes.
[0094] The temperature variation dispersion index reflects the stability of the thermal state distribution during cluster operation. The data in the table shows that under the traditional rule-based management method, the temperature variation dispersion is within a wide range, indicating significant differences in thermal state between different devices. The statistical model management method shows some convergence on this index, but still exhibits a certain degree of dispersion. After adopting the method of this invention, the temperature variation dispersion remains within a narrower range, indicating that energy consumption and temperature state, under the influence of independent reservoir modeling and subsequent screening mechanisms, can participate more stably in management decisions.
[0095] In terms of both management decision consistency rate and state response matching rate, the numerical ranges corresponding to the method of this invention are generally higher than those of other comparison methods, and the range distribution is more concentrated. This indicates that during continuous operation, the management output generated by this invention maintains a good correspondence with changes in cluster state, and management decisions exhibit consistency characteristics as state evolves. In summary, the above indicators show that the method of this invention establishes a more stable correlation between computing power state, energy consumption state, and management output, verifying its effectiveness in cluster computing power device management scenarios.
[0096] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method and apparatus for managing cluster computing power devices based on deep learning, characterized in that, Includes the following steps: S1. Collect the operating status data of the cluster computing devices. The operating status data includes computing power utilization data sequences and energy consumption and temperature data sequences. S2. Perform time alignment processing on the computing power utilization data sequences and energy consumption and temperature data sequences to form computing power status input sequences and energy consumption and thermal status input sequences. S3. Construct an echo state network model with two reservoirs in parallel. The first reservoir receives the computing power status input sequence, and the second reservoir receives the energy consumption and thermal status input sequence. S4. Perform recursive mapping in the first and second reservoirs respectively to generate the state matrix of the computing power reservoir and the state matrix of the energy consumption hot water reservoir; S5. Perform expansion processing on the state matrix of the computing power reservoir and the state matrix of the energy consumption hot water reservoir to construct a joint state component set; S6. Calculate the mutual information value between each state component in the joint state component set and the management output variable; S7. Retain the state components whose mutual information values satisfy the preset sparse constraints to form a sparse fused state vector; S8. Construct an output layer linear mapping structure based on sparse fusion state vectors. The output layer linear mapping structure contains parameter subsets corresponding to the first and second reservoirs. S9. The recursive least squares algorithm is used to perform online updates on the parameter subsets respectively, and the cluster computing power device management decision is generated and output based on the updated output layer linear mapping structure.
2. The method and apparatus for managing cluster computing power devices based on deep learning according to claim 1, characterized in that, S3 specifically includes: Within the same echo state network framework, a first reservoir and a second reservoir are configured in parallel. They are topologically independent, with no direct connection path at the reservoir level between them. The first reservoir establishes an input connection only with the computing power state input sequence, while the second reservoir establishes an input connection only with the energy consumption and heat state input sequence. The computing power state input sequence is not transmitted to the second reservoir, and the energy consumption and heat state input sequence is not transmitted to the first reservoir, thus achieving structural isolation of state sources at the input level. Each of the first and second reservoirs contains several reservoir nodes, which form an internal recursive connection structure through fixed random connections. This internal recursive connection structure remains unchanged during model operation, and the internal connection weights do not participate in parameter updates during the training phase. The first and second reservoirs exhibit structural differences in at least one of the following parameters: reservoir node configuration, internal connection sparsity parameter, or spectral radius parameter. The system exhibits different time response characteristics for dynamic computing power status and dynamic energy consumption / heat status. During the operation of the echo state network model with two parallel reservoirs, the first reservoir independently performs state recursion operations based on the computing power status input sequence to generate a computing power reservoir state matrix, while the second reservoir independently performs state recursion operations based on the energy consumption / heat status input sequence to generate an energy consumption / heat reservoir state matrix. The computing power reservoir state matrix and the energy consumption / heat reservoir state matrix do not undergo merging, weighting, or cross-mapping operations at the reservoir level. In the parallel dual-reservoir structure, a cross-reservoir state fusion path is introduced only before the output layer is constructed. This path consists of state components selected by the mutual information filtering step; reservoir state components that do not pass the mutual information filtering step do not enter the cross-reservoir fusion path. Through structural configuration, the first and second reservoirs independently model different state sources at the reservoir level and form a state input structure oriented towards management output through a restricted cross-reservoir fusion path before the output layer.
3. The method and apparatus for managing cluster computing power devices based on deep learning according to claim 1, characterized in that, The construction of the joint state component set in step S5 specifically includes: The state matrices of the computing power reservoir and the energy consumption hot water reservoir are read according to the time index order, and the corresponding reservoir state vector is extracted at each time index position. The state vectors from the computing power reservoir state matrix and the energy consumption hot water reservoir state matrix at the same time index position are recorded as independent state components. In the time dimension, the state components corresponding to each time index position are arranged and aggregated to form a joint state component set containing the computing power reservoir state components and the energy consumption hot water reservoir state components. Each state component in the joint state component set retains its source identifier and time index information.
4. The method and apparatus for managing cluster computing power devices based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71. After obtaining the mutual information value of each state component in the joint state component set, the mutual information value of each state component is compared with the preset sparse constraint condition one by one. The preset sparse constraint condition limits the number or proportion of state components entering the state fusion process. Only when the mutual information value meets the preset sparse constraint condition is the corresponding state component marked as a valid state component. S72. Perform a removal operation on state components that do not meet the preset sparse constraints. Class state components do not participate in the state vector construction process, thereby forming a sparse state component subset that contains only valid state components in the joint state component set. S73. In the sparse state component subset, retain the source reservoir identifier and time index identifier for each effective state component. S74. At the same time index position, read the valid state component of the corresponding time index identifier and perform an ordered arrangement operation on the state component from the computing power reservoir and the state component from the energy consumption hot water reservoir according to the preset reservoir order rules. The ordered arrangement operation only changes the organization order of the state components and does not change the numerical content of the state components. S75. Combine the effective state components arranged in an ordered manner at the same time index position to form a sparse state vector corresponding to the time index position. The dimension of the sparse state vector is determined by the number of effective state components. Unselected state components do not occupy a position in the sparse state vector. S76. In the time dimension, according to the time index from small to large, perform sequential splicing and recording operations on the sparse state vectors corresponding to each time index position to form a sparse fused state vector covering several time index positions. The sparse fusion state vector formed by steps S71-S76 contains only state components that satisfy the mutual information sparse constraint condition, and preserves the source reservoir information and time sequence information of the state components.
5. The method and apparatus for managing cluster computing power devices based on deep learning according to claim 1, characterized in that, The state fusion process in step S7 specifically includes: After comparing each state component in the joint state component set with the preset sparse constraints, the state components whose mutual information values satisfy the preset sparse constraints are selected as the state fusion input set. At each time index position, the state component at the corresponding time index position is read from the state fusion input set, and the state components are grouped and recorded according to the source reservoir type. Within the same time index position, the state components from the computing power reservoir and the state components from the energy consumption hot water reservoir are arranged sequentially according to the preset reservoir order rules to form a time slice state sequence; a vectorization combination operation is performed on the time slice state sequence to generate a fused state vector corresponding to the time index position; In the time dimension, sequential aggregation operations are performed on the fused state vectors corresponding to each time index position according to the time index order to form a state fusion result covering several time index positions; in the state fusion process, no numerical weighting, summation or mapping operations are performed between state components.
6. The method and apparatus for managing cluster computing power devices based on deep learning according to claim 1, characterized in that, S8 specifically includes: After obtaining the sparse fused state vector, a linear mapping structure for the output layer is constructed based on the sparse fused state vector. The state components in the sparse fused state vector are distinguished according to the source reservoir type, and divided into a set of state components corresponding to the first reservoir and a set of state components corresponding to the second reservoir. Parameter subsets are configured for different sets of state components. A mapping relationship is established between the first parameter subset and the state component set of the first reservoir, and a mapping relationship is established between the second parameter subset and the state component set of the second reservoir. By performing linear combination operations on each parameter subset and the corresponding set of state components, a structured parameter configuration for the output layer is formed.
7. The method and apparatus for managing cluster computing power devices based on deep learning according to claim 1, characterized in that, S9 specifically includes: After completing the construction of the linear mapping structure of the output layer, a recursive least squares online update structure is established for the first parameter subset of the first reservoir state component set and the second parameter subset of the second reservoir state component set in the output layer, respectively. The recursive least squares online update structure is kept in a continuously enabled state during the model operation. When each time index position is reached, the sparse fusion state vector corresponding to the time index position is read, and the management output variable corresponding to the time index position is read synchronously. The sparse fusion state vector is used as the update input of the current time index position, and the management output variable is used as the update reference of the current time index position. The update process and the state generation process maintain a one-to-one correspondence on the time index. Using the data at the current time index position as a single update sample, recursive least squares update calculations are performed on the first parameter subset and the second parameter subset respectively. The update calculation is based only on the data at the current time index position and the parameter state and covariance information retained at the previous time index position, without introducing the data sequence of historical time index positions for batch backtracking calculations. During the recursive least squares update process, the corresponding gain vectors are calculated for the first and second parameter subsets respectively, and the parameters are corrected based on the gain vectors. The parameter correction is written at the current time index position. After the parameter correction is completed, the updated parameter subset is retained as the initial parameter state for the recursive update at the next time index position. After the parameter update is completed at any time index position, the output layer linear mapping structure immediately performs a linear mapping operation on the sparse fused state vector at the time index position using the updated first parameter subset and second parameter subset, and manages the output to be continuously generated as the time index advances; During the online update process, the update processes of the first parameter subset and the second parameter subset are executed independently. No coupling relationship is established across parameter subsets during the parameter update phase. Moreover, the parameter update operation is performed only once at any time index position. The entire update process proceeds continuously along the time index direction, thus forming a time-index-based online parameter update mechanism based on the recursive least squares algorithm.