An energy-saving strategy recommendation system and method applied to computing power energy efficiency evaluation

Abstract

The application discloses an energy-saving strategy recommendation system and method applied to computing power energy efficiency evaluation, relates to the technical field of data analysis, and comprises the following steps: collecting multi-dimensional resource monitoring data of a computing node; calculating the comprehensive load rate of each node at the corresponding moment according to the collected multi-dimensional resource monitoring data; filtering out a node set with a real-time load rate lower than a set threshold value according to real-time multi-dimensional resource data to form a candidate node set, extracting the resource demand of a to-be-assigned task, constructing a resource demand vector, and calculating the load matching degree of the task and each node in the candidate node set at a specific moment; calculating the load matching degree of the candidate node, sorting the load matching degrees in descending order, selecting the highest load matching degree, and comparing the comprehensive load rate of the highest load matching degree with the threshold value; if the comprehensive load rate is lower than the threshold value, the task is assigned; if the comprehensive load rate exceeds the threshold value, the task is marked as degraded, a degradation factor is calculated by using historical data to adjust the load matching degree, the load matching degrees are re-sorted and compared, and the task is assigned to a node meeting the condition until the task is assigned to the node meeting the condition, so that the energy-saving strategy recommendation of the computing power energy efficiency evaluation scene can be realized.

Description

An energy-saving strategy recommendation system and method for computing power energy efficiency assessment Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to an energy-saving strategy recommendation system and method for evaluating computing power energy efficiency. Background Technology

[0002] With the surge in demand for AI computing power and the expansion of cluster scale, enterprises have placed higher demands on the dynamic scheduling, load balancing, and energy efficiency assessment of computing resources. Load balancing capability directly affects the cluster resource utilization, task execution stability, and response efficiency, while energy efficiency assessment is related to cluster operating cost control, energy consumption optimization, and the achievement of green computing power goals. Together, they constitute the core support for the efficient operation of AI computing power clusters.

[0003] However, traditional load balancing and energy efficiency management methods often face the following problems when dealing with complex, multi-dimensional resource scenarios: First, resource monitoring is limited to a single dimension, relying heavily on CPU or memory resource metrics, lacking real-time dynamic collection of multi-dimensional resources such as GPU utilization, disk I / O latency, and network bandwidth. This makes it difficult to comprehensively reflect the actual load status of nodes, leading to distorted load assessments. Second, load balancing and energy efficiency optimization are disconnected, with static strategies. Traditional strategies do not dynamically match task resource requirements with the real-time remaining resources of nodes, but only allocate tasks according to fixed rules, easily causing imbalances such as "high-load nodes still being assigned tasks" or "low-load nodes having idle resources." Third, the matching degree calculation logic is simple, failing to consider whether the remaining resources meet the task requirements, resulting in poor task and node compatibility. In addition, there is a lack of dynamic degradation and adjustment mechanisms. When the node load exceeds the threshold, it cannot intelligently correct the matching degree based on historical data, easily causing task execution failures or delays, and may also lead to abnormally high node energy consumption due to blindly allocating tasks. Summary of the Invention

[0004] The purpose of this invention is to provide an energy-saving strategy recommendation system and method for computing power energy efficiency assessment, so as to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for recommending energy-saving strategies for computing power energy efficiency assessment, the method comprising the following steps:

[0006] The AI ​​computing power intelligent scheduling platform collects multi-dimensional resource monitoring data from computing nodes;

[0007] The comprehensive load rate of each node at the corresponding time is calculated based on the collected multi-dimensional resource monitoring data;

[0008] Based on real-time multidimensional resource data, nodes with real-time load rates below a set threshold are selected to form a candidate node set. The resource requirements of CPU, GPU, disk input / output, memory, and network bandwidth for the task to be assigned are extracted, a resource requirement vector is constructed, and the load matching degree between the task and each node in the candidate node set at a specific time is calculated.

[0009] The system iterates through candidate nodes to calculate the load matching degree of each node. After sorting the nodes by matching degree from high to low, it selects the node with the highest matching degree and compares its overall load rate with the set load threshold. If the overall load rate of the node is lower than the threshold, the task to be assigned is assigned. If it exceeds the threshold, the node is marked as downgraded, and a downgrade factor is calculated based on historical data to adjust its load matching degree. The nodes are then re-sorted by the adjusted matching degree from high to low, and the node with the highest matching degree is selected for comparison again until a node with an overall load rate lower than the threshold is found. The task is assigned to the node, and the task assignment details of the node are recorded synchronously. This record is used as part of the energy-saving strategy and fed back to the AI ​​computing power scheduling platform.

[0010] The AI ​​computing power intelligent scheduling platform collects multi-dimensional resource monitoring data from computing nodes. The specific steps include:

[0011] The AI ​​computing power intelligent scheduling platform obtains the computing node resource dataset N={N1,N2,...,N} through lightweight monitoring agents deployed on each computing node. i ,...,N p}; where N represents the acquired set of computing node resource data, N i This represents the multidimensional resource data of the i-th computing node, where i = 1, 2, ..., p, and p represents the total number of computing nodes in the cluster. For the i-th node, the collected multidimensional resource data includes Ni = {CPU...} i GPU i DISK i MEM i ,NET i}; where CPU i This represents the CPU utilization rate collected by the i-th node, and the GPU utilization rate. i DISK represents the GPU utilization rate collected by the i-th node. i MEM represents the disk input / output latency collected by the i-th node within the detection period T. i Represents the memory utilization rate collected by the i-th node, NET i This represents the network bandwidth utilization rate collected by the i-th node.

[0012] The comprehensive load rate of each node at the corresponding time is calculated based on the collected multi-dimensional resource monitoring data. The specific steps include:

[0013] The collected multidimensional resource data is normalized, and the normalized load of the multidimensional resource data is calculated using the maximum benchmark value of the cluster as the denominator. The load of the normalized multidimensional resource data is then weighted and summed using dynamic weighting coefficients to obtain the comprehensive load rate Li(t) of node i at time t. Here, the cluster represents the set of hardware resources formed by the interconnection of nodes that the AI ​​computing power intelligent scheduling platform can schedule, including CPU, GPU, disk input / output, memory and network bandwidth. The maximum benchmark value of the cluster represents the resource type for each type of hardware resource set. Based on the historical running data of the mainstream configuration nodes in the cluster, the upper limit of resource usage for stable operation is determined. Li(t) represents the comprehensive load rate of node i at time t.

[0014] The overall load rate reflects the stability of a node's multi-dimensional resource utilization. Among them, disk input / output latency is inversely related to the overall load rate, while CPU utilization, GPU utilization, memory utilization, and network bandwidth utilization are positively related to the overall load rate.

[0015] The maximum baseline value of the cluster represents the upper limit of resource usage for stable operation for each resource type in the hardware resource set, determined based on the historical operating data of the mainstream configuration nodes in the cluster. The mainstream configuration node refers to the node that accounts for a proportion of all nodes, where a is a value between 0.5 and 1. Stable operation means that the node has a task success rate greater than or equal to b under this upper limit, where b represents the pre-set minimum threshold for task success rate.

[0016] Based on real-time multidimensional resource data, nodes with real-time load rates below a set threshold are selected to form a candidate node set. The resource requirements for CPU, GPU, disk I / O, memory, and network bandwidth for the task to be assigned are extracted, a resource requirement vector is constructed, and the load matching degree between the task and each node in the candidate node set at a specific time is calculated. Specific steps include:

[0017] A candidate node set NC={n1,n2,...,nm} is generated based on real-time multidimensional resource data. For the task to be assigned tx, its resource requirement vector Rx=[r1,r2,r3,r4,r5] is extracted. Here, candidate nodes represent nodes with a real-time load rate less than a set threshold, NC represents the candidate node set generated based on real-time multidimensional resource data, n1,n2,...,nm represent the 1st, 2nd,...,mth candidate nodes, tx represents the xth task to be assigned, Rx represents the resource requirement vector of the xth task to be assigned, and r1,r2,r3,r4,r5 represent the resource requirements of CPU, GPU, disk I / O, memory, and network bandwidth, respectively.

[0018] The load matching degree between the task to be assigned tx and node nm at time t is calculated and defined as follows:

[0019] When the resource requirement of the k-th dimension of the task to be assigned is greater than or equal to the resource remaining amount of the k-th dimension of the task to be assigned, the load matching degree between the task tx and the node nm at time t is recorded as 0, where the resource remaining amount is calculated based on the total resource capacity minus the real-time resource usage.

[0020] Otherwise, the load matching degree between task tx and node nm at time t is defined as follows:

[0021] ;

[0022] Where M(x,m,t) represents the load matching degree between task tx and node nm at time t, R(m,k,t) represents the real-time usage of the k-th dimension resource by node nm at time t, r(x,k) represents the k-th dimension resource requirement of task tx to be assigned, and C m,k Let wk represent the total resource capacity of node nm in the k-th dimension, where k represents the enumerated dimension, k=1,2,...,5, representing CPU, GPU, disk input / output, memory and network bandwidth respectively, and wk represents the resource weight in the k-th dimension.

[0023] The process involves iterating through candidate nodes, calculating the load matching degree for each node, sorting them from highest to lowest matching degree, selecting the node with the highest matching degree, and comparing its overall load rate with a set load threshold. If the node's overall load rate is lower than the threshold, the tasks to be assigned are allocated; if it exceeds the threshold, the node is marked for demotion, and a demotion factor is calculated based on historical data to adjust its load matching degree. The nodes are then re-sorted from highest to lowest matching degree, and the node with the highest matching degree is selected for comparison again, until a node with an overall load rate lower than the threshold is found. Tasks are then assigned to this node, and the task allocation details for this node are recorded synchronously. This record is incorporated into the energy-saving strategy and fed back to the AI ​​computing power scheduling platform. The specific steps include:

[0024] Traverse the candidate node set to calculate the load matching degree of the candidate nodes in the candidate node set, sort the candidate nodes in descending order according to the calculated load matching degree, and select the node with the highest load matching degree.

[0025] The overall load rate Lmn(t) of node mn at time t is compared with the set load threshold. When the overall load rate Lmn(t) of node mn at time t is less than the set load threshold, the task to be assigned is assigned to node mn.

[0026] When the overall load rate Lmn(t) of node mn at time t exceeds the set load threshold, the node is marked for degradation. The degradation factor is calculated by dividing the number of times the task to be processed was successfully executed under the current overall load rate by the total number of times. The degradation factor is multiplied by the load matching degree of node mn to obtain the adjusted load matching degree of node mn. The load matching degrees of candidate nodes are re-sorted in descending order of value. The node with the highest adjusted load matching degree is selected and compared with the set load threshold until the overall load rate Lmn(t) of node mn at time t is less than the set load threshold. At this time, the task to be assigned is assigned to node mn. The task assignment details of the node are recorded synchronously and the record is fed back to the AI ​​computing power scheduling platform as part of the energy saving strategy.

[0027] An energy-saving strategy recommendation system for computing power energy efficiency assessment includes: a resource data acquisition module, a comprehensive load rate calculation module, a node screening module, and a task intelligent allocation module. The resource data acquisition module collects multi-dimensional resource monitoring data of computing nodes. The comprehensive load rate calculation module calculates the comprehensive load rate of each node at a given time based on the collected multi-dimensional resource monitoring data. The node screening module, based on real-time multi-dimensional resource data, selects nodes with real-time load rates below a set threshold to form a candidate node set, extracts the resource requirements of the tasks to be allocated, constructs a resource requirement vector, and calculates the load matching degree between the task and each node in the candidate node set at a specific time. The intelligent task allocation module calculates the load matching degree of candidate nodes, sorts them from highest to lowest, selects the highest, and compares its comprehensive load rate with a threshold. If the load rate is lower, a task is allocated; if it exceeds the threshold, it is marked as downgraded. Historical data is used to calculate a downgrade factor to adjust the matching degree, and the comparison is re-sorted until a node that meets the conditions is found for allocation. The task allocation details of the node are recorded synchronously and used as part of the energy-saving strategy, which is fed back to the AI ​​computing power scheduling platform. The output of the resource data acquisition module is connected to the input of the comprehensive load rate calculation module; the output of the comprehensive load rate calculation module is connected to the input of the node filtering module; and the output of the node filtering module is connected to the input of the intelligent task allocation module.

[0028] The resource data acquisition module includes a monitoring agent unit and a data aggregation unit. The monitoring agent unit means that the AI ​​computing power intelligent scheduling platform collects CPU utilization, GPU utilization, disk I / O latency, memory utilization and network bandwidth utilization data through lightweight monitoring agents deployed on each computing node. The data aggregation unit is used to summarize the data collected by each node into a resource dataset.

[0029] The comprehensive load rate calculation module includes a data normalization unit and a load weighting calculation unit. The data normalization unit is used to normalize multi-dimensional resource data with reference to the maximum benchmark value of the cluster. The load weighting calculation unit is used to sum the normalized load data by weighting it with dynamic weighting coefficients to obtain the comprehensive load rate of the node at a specific time.

[0030] The node filtering module includes a candidate node filtering unit, a task requirement parsing unit, and a load matching degree calculation unit. The candidate node filtering unit is used to filter out nodes below a set threshold based on the real-time load rate, forming a candidate node set. The task requirement parsing unit is used to extract the resource requirements of the task to be assigned, such as CPU, GPU, disk input / output, memory, and network bandwidth, and construct a resource requirement vector. The load matching degree calculation unit is used to calculate the load matching degree between the task and each candidate node at a specific time based on the remaining node resources and the task requirements.

[0031] The intelligent task allocation module includes a matching degree sorting unit, a load threshold judgment unit, an adjustment unit, and a task allocation execution unit. The matching degree sorting unit sorts the load matching degree of candidate nodes from high to low. The load threshold judgment unit compares the overall load rate of the sorted nodes with a set threshold to determine whether they meet the task allocation conditions. The adjustment unit calculates a degradation factor based on historical data, adjusts the load matching degree of nodes, and re-sorts them when the node load rate exceeds the threshold. The task allocation execution unit selects nodes with an overall load rate lower than the threshold, performs task allocation operations, synchronously records the task allocation details of the node, and feeds this record back to the AI ​​computing power scheduling platform as part of the energy-saving strategy.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] 1. By deploying a lightweight monitoring agent, this invention collects multi-dimensional resource data such as CPU utilization, GPU utilization, disk I / O latency, memory utilization, and network bandwidth utilization of computing nodes. Combined with dynamic weighting coefficients, it calculates the comprehensive load rate. Unlike the single resource indicator monitoring in existing technologies, this invention reflects the actual load status of nodes, realizes the evaluation of node load, and thus improves the comprehensiveness of load monitoring.

[0034] 2. By constructing a task resource demand vector, calculating the load matching degree based on the remaining node resources and task demand, and dynamically sorting and allocating tasks according to the matching degree, unlike the static rule allocation in the existing technology, this invention realizes intelligent adaptation between tasks and nodes, avoiding abnormal increase in node energy consumption due to blind task allocation.

[0035] 3. A dynamic degradation and adjustment mechanism is introduced. When the overall load rate of a node exceeds the threshold, the degradation factor is calculated based on historical data to adjust the matching degree and re-sort. Unlike the lack of a load overload response mechanism in the existing technology, this invention selects stable nodes to undertake tasks, reducing task failures or delays caused by resource fluctuations. Attached Figure Description

[0036] Figure 1 is a flowchart illustrating an energy-saving strategy recommendation method for computing power energy efficiency assessment according to the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] In an embodiment: As shown in Figure 1, the present invention provides a technical solution, a method for recommending energy-saving strategies for computing power energy efficiency assessment, the method comprising the following steps:

[0039] The AI ​​computing power intelligent scheduling platform collects multi-dimensional resource monitoring data from computing nodes;

[0040] The comprehensive load rate of each node at the corresponding time is calculated based on the collected multi-dimensional resource monitoring data;

[0041] Based on real-time multidimensional resource data, nodes with real-time load rates below a set threshold are selected to form a candidate node set. The resource requirements of CPU, GPU, disk input / output, memory, and network bandwidth for the task to be assigned are extracted, a resource requirement vector is constructed, and the load matching degree between the task and each node in the candidate node set at a specific time is calculated.

[0042] The system iterates through candidate nodes to calculate the load matching degree of each node. After sorting the nodes by matching degree from high to low, it selects the node with the highest matching degree and compares its overall load rate with the set load threshold. If the overall load rate of the node is lower than the threshold, the task to be assigned is assigned. If it exceeds the threshold, the node is marked as downgraded, and a downgrade factor is calculated based on historical data to adjust its load matching degree. The nodes are then re-sorted by the adjusted matching degree from high to low, and the node with the highest matching degree is selected for comparison again until a node with an overall load rate lower than the threshold is found. The task is assigned to the node, and the task assignment details of the node are recorded synchronously. This record is used as part of the energy-saving strategy and fed back to the AI ​​computing power scheduling platform.

[0043] The AI ​​computing power intelligent scheduling platform collects multi-dimensional resource monitoring data from computing nodes. The specific steps include:

[0044] The AI ​​computing power intelligent scheduling platform obtains the computing node resource dataset N={N1,N2,...,N} through lightweight monitoring agents deployed on each computing node. i ,...,N p}; where N represents the acquired set of computing node resource data, N i This represents the multidimensional resource data of the i-th computing node, where i = 1, 2, ..., p, and p represents the total number of computing nodes in the cluster. For the i-th node, the collected multidimensional resource data includes Ni = {CPU...} i GPU i DISK i MEM i ,NET i}; where CPU i This represents the CPU utilization rate collected by the i-th node, and the GPU utilization rate. i DISK represents the GPU utilization rate collected by the i-th node. i MEM represents the disk input / output latency collected by the i-th node within the detection period T. i Represents the memory utilization rate collected by the i-th node, NET i This represents the network bandwidth utilization rate collected by the i-th node.

[0045] Specifically, each node deploys a lightweight monitoring agent with a monitoring cycle of 5 minutes, collecting multi-dimensional resource data: Node 1's CPU utilization = 30%, GPU utilization = 20%, disk I / O latency = 10ms, memory utilization = 40%, and network bandwidth utilization = 25%;

[0046] Node 2 has the following CPU utilization rates: 70%, GPU utilization: 60%, disk I / O latency: 30ms, memory utilization: 75%, and network bandwidth utilization: 80%.

[0047] Node 3 has the following CPU utilization rates: 20%, GPU utilization: 10%, disk I / O latency: 8ms, memory utilization: 30%, and network bandwidth utilization: 20%.

[0048] Node 4 has the following CPU utilization rates: 50%, GPU utilization: 40%, disk I / O latency: 15ms, memory utilization: 50%, and network bandwidth utilization: 50%.

[0049] Node 5 has a CPU utilization of 65%, a GPU utilization of 55%, a disk I / O latency of 25ms, a memory utilization of 60%, and a network bandwidth utilization of 70%.

[0050] The comprehensive load rate of each node at the corresponding time is calculated based on the collected multi-dimensional resource monitoring data. The specific steps include:

[0051] The collected multidimensional resource data is normalized, and the normalized load of the multidimensional resource data is calculated using the maximum benchmark value of the cluster as the denominator. The load of the normalized multidimensional resource data is then weighted and summed using dynamic weighting coefficients to obtain the comprehensive load rate Li(t) of node i at time t. Here, the cluster represents the set of hardware resources formed by the interconnection of nodes that the AI ​​computing power intelligent scheduling platform can schedule, including CPU, GPU, disk input / output, memory and network bandwidth. The maximum benchmark value of the cluster represents the resource type for each type of hardware resource set. Based on the historical running data of the mainstream configuration nodes in the cluster, the upper limit of resource usage for stable operation is determined. Li(t) represents the comprehensive load rate of node i at time t.

[0052] The overall load rate reflects the stability of a node's multi-dimensional resource utilization. Among them, disk input / output latency is inversely related to the overall load rate, while CPU utilization, GPU utilization, memory utilization, and network bandwidth utilization are positively related to the overall load rate.

[0053] The maximum baseline value for the cluster represents the upper limit of resource usage for stable operation for each resource type in the hardware resource set, determined based on historical operating data of the mainstream configuration nodes in the cluster. Mainstream configuration nodes are defined as nodes representing a percentage (a) of all nodes, where a is a value between 0.5 and 1. Stable operation means that the task success rate of a node is greater than or equal to b, where b represents a pre-set minimum threshold for task success rate, within this upper limit. Specifically, the maximum baseline value for the cluster (based on historical data of mainstream configuration nodes, a = 0.6, i.e., 3 mainstream nodes) is as follows: CPU = 80%, GPU = 70%, Disk I / O latency = 50ms, Memory = 80%, Network bandwidth = 90%.

[0054] Set dynamic weighting coefficients, and perform weighted summation on the multidimensional resource data load based on the set weighting coefficients to calculate the comprehensive load rate of nodes 1 to 5 respectively.

[0055] Based on real-time multidimensional resource data, nodes with real-time load rates below a set threshold are selected to form a candidate node set. The resource requirements for CPU, GPU, disk I / O, memory, and network bandwidth for the task to be assigned are extracted, a resource requirement vector is constructed, and the load matching degree between the task and each node in the candidate node set at a specific time is calculated. Specific steps include:

[0056] A candidate node set NC={n1,n2,...,nm} is generated based on real-time multidimensional resource data. For the task to be assigned tx, its resource requirement vector Rx=[r1,r2,r3,r4,r5] is extracted. Here, candidate nodes represent nodes with a real-time load rate less than a set threshold, NC represents the candidate node set generated based on real-time multidimensional resource data, n1,n2,...,nm represent the 1st, 2nd,...,mth candidate nodes, tx represents the xth task to be assigned, Rx represents the resource requirement vector of the xth task to be assigned, and r1,r2,r3,r4,r5 represent the resource requirements of CPU, GPU, disk I / O, memory, and network bandwidth, respectively.

[0057] The load matching degree between the task to be assigned tx and node nm at time t is calculated and defined as follows:

[0058] When the resource requirement of the k-th dimension of the task to be assigned is greater than or equal to the resource remaining amount of the k-th dimension of the task to be assigned, the load matching degree between the task tx and the node nm at time t is recorded as 0, where the resource remaining amount is calculated based on the total resource capacity minus the real-time resource usage.

[0059] Otherwise, the load matching degree between task tx and node nm at time t is defined as follows:

[0060] ;

[0061] Where M(x,m,t) represents the load matching degree between task tx and node nm at time t, R(m,k,t) represents the real-time usage of the k-th dimension resource by node nm at time t, r(x,k) represents the k-th dimension resource requirement of task tx to be assigned, and C m,k Let wk represent the total resource capacity of node nm in the k-th dimension, where k represents the enumerated dimension, k=1,2,...,5, representing CPU, GPU, disk input / output, memory and network bandwidth respectively, and wk represents the resource weight in the k-th dimension.

[0062] Specifically, set the load threshold to 60% and filter nodes with an overall load rate of <60%: candidate node set NC={n1,n3,n4} (nodes 1, 3, and 4 meet the conditions).

[0063] The resource requirement vector for task tx is Rx=[20%,30%,5ms,15%,20%] (CPU requires 20%, GPU requires 30%, disk I / O latency requires ≤5ms, memory requires 15%, and bandwidth requires 20%). Calculate the load matching degree for nodes 1, 3, and 4 respectively.

[0064] The process involves iterating through candidate nodes, calculating the load matching degree for each node, sorting them from highest to lowest matching degree, selecting the node with the highest matching degree, and comparing its overall load rate with a set load threshold. If the node's overall load rate is lower than the threshold, the tasks to be assigned are allocated; if it exceeds the threshold, the node is marked for demotion, and a demotion factor is calculated based on historical data to adjust its load matching degree. The nodes are then re-sorted from highest to lowest matching degree, and the node with the highest matching degree is selected for comparison again, until a node with an overall load rate lower than the threshold is found. Tasks are then assigned to this node, and the task allocation details for this node are recorded synchronously. This record is incorporated into the energy-saving strategy and fed back to the AI ​​computing power scheduling platform. The specific steps include:

[0065] Traverse the candidate node set to calculate the load matching degree of the candidate nodes in the candidate node set, sort the candidate nodes in descending order according to the calculated load matching degree, and select the node with the highest load matching degree.

[0066] The overall load rate Lmn(t) of node mn at time t is compared with the set load threshold. When the overall load rate Lmn(t) of node mn at time t is less than the set load threshold, the task to be assigned is assigned to node mn.

[0067] When the overall load rate Lmn(t) of node mn at time t exceeds the set load threshold, the node is marked for degradation. The degradation factor is calculated by dividing the number of times the task to be processed was successfully executed under the current overall load rate by the total number of times. The degradation factor is multiplied by the load matching degree of node mn to obtain the adjusted load matching degree of node mn. The load matching degrees of candidate nodes are re-sorted in descending order of value. The node with the highest adjusted load matching degree is selected and compared with the set load threshold until the overall load rate Lmn(t) of node mn at time t is less than the set load threshold. At this time, the task to be assigned is assigned to node mn. The task assignment details of the node are recorded synchronously and the record is fed back to the AI ​​computing power scheduling platform as part of the energy saving strategy.

[0068] Specifically, based on the calculation, node 4 is a valid candidate node; load threshold judgment: the overall load rate of node 4 = 55% < 60% (threshold), and the task tx is directly assigned to node 4.

[0069] An energy-saving strategy recommendation system for computing power energy efficiency assessment includes: a resource data acquisition module, a comprehensive load rate calculation module, a node screening module, and a task intelligent allocation module. The resource data acquisition module collects multi-dimensional resource monitoring data of computing nodes. The comprehensive load rate calculation module calculates the comprehensive load rate of each node at a given time based on the collected multi-dimensional resource monitoring data. The node screening module, based on real-time multi-dimensional resource data, selects nodes with real-time load rates below a set threshold to form a candidate node set, extracts the resource requirements of the tasks to be allocated, constructs a resource requirement vector, and calculates the load matching degree between the task and each node in the candidate node set at a specific time. The intelligent task allocation module calculates the load matching degree of candidate nodes, sorts them from highest to lowest, selects the highest, and compares its comprehensive load rate with a threshold. If the load rate is lower, a task is allocated; if it exceeds the threshold, it is marked as downgraded. Historical data is used to calculate a downgrade factor to adjust the matching degree, and the comparison is re-sorted until a node that meets the conditions is found for allocation. The task allocation details of the node are recorded synchronously and used as part of the energy-saving strategy, which is fed back to the AI ​​computing power scheduling platform. The output of the resource data acquisition module is connected to the input of the comprehensive load rate calculation module; the output of the comprehensive load rate calculation module is connected to the input of the node filtering module; and the output of the node filtering module is connected to the input of the intelligent task allocation module.

[0070] The resource data acquisition module includes a monitoring agent unit and a data aggregation unit. The monitoring agent unit means that the AI ​​computing power intelligent scheduling platform collects CPU utilization, GPU utilization, disk I / O latency, memory utilization and network bandwidth utilization data through lightweight monitoring agents deployed on each computing node. The data aggregation unit is used to summarize the data collected by each node into a resource dataset.

[0071] The comprehensive load rate calculation module includes a data normalization unit and a load weighting calculation unit. The data normalization unit is used to normalize multi-dimensional resource data with reference to the maximum benchmark value of the cluster. The load weighting calculation unit is used to sum the normalized load data by weighting it with dynamic weighting coefficients to obtain the comprehensive load rate of the node at a specific time.

[0072] The node filtering module includes a candidate node filtering unit, a task requirement parsing unit, and a load matching degree calculation unit. The candidate node filtering unit is used to filter out nodes below a set threshold based on the real-time load rate, forming a candidate node set. The task requirement parsing unit is used to extract the resource requirements of the task to be assigned, such as CPU, GPU, disk input / output, memory, and network bandwidth, and construct a resource requirement vector. The load matching degree calculation unit is used to calculate the load matching degree between the task and each candidate node at a specific time based on the remaining node resources and the task requirements.

[0073] The intelligent task allocation module includes a matching degree sorting unit, a load threshold judgment unit, an adjustment unit, and a task allocation execution unit. The matching degree sorting unit sorts the load matching degree of candidate nodes from high to low. The load threshold judgment unit compares the overall load rate of the sorted nodes with a set threshold to determine whether they meet the task allocation conditions. The adjustment unit calculates a degradation factor based on historical data, adjusts the load matching degree of nodes, and re-sorts them when the node load rate exceeds the threshold. The task allocation execution unit selects nodes with an overall load rate lower than the threshold, performs task allocation operations, synchronously records the task allocation details of the node, and feeds this record back to the AI ​​computing power scheduling platform as part of the energy-saving strategy.

[0074] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for recommending energy-saving strategies for computing power energy efficiency assessment, characterized in that: The method includes the following steps: An AI computing power intelligent scheduling platform collects multi-dimensional resource monitoring data of computing nodes; calculates the comprehensive load rate of each node at a given time based on the collected multi-dimensional resource monitoring data; based on real-time multi-dimensional resource data, selects nodes with real-time load rates below a set threshold to form a candidate node set; extracts the resource requirements of CPU, GPU, disk input / output, memory, and network bandwidth for the task to be assigned, constructs a resource requirement vector, and calculates the load matching degree between the task and each node in the candidate node set at a specific time; iterates through the candidate nodes to calculate the load matching degree of each node, and sorts them from high to low matching degree. The process begins by selecting the node with the highest overall load rate and comparing it to a set load threshold. If the node's overall load rate is below the threshold, the tasks to be assigned are allocated. If it exceeds the threshold, the node is marked for demotion, and a demotion factor is calculated based on historical data to adjust its load matching degree. The nodes are then re-sorted from highest to lowest according to the adjusted matching degree, and the node with the highest load rate is selected for comparison again. This process continues until a node with an overall load rate below the threshold is found. The tasks to be assigned are then scheduled to this node, and the task allocation details for this node are recorded synchronously. This record is incorporated into the energy-saving strategy and fed back to the AI ​​computing power scheduling platform. The process then iterates through the candidate node set to perform calculations. The load matching degree of candidate nodes in the candidate node set is calculated and sorted in descending order. The node with the highest load matching degree is selected. The overall load rate Lmn(t) of node mn at time t is compared with a set load threshold. If the overall load rate Lmn(t) of node mn at time t is less than the set load threshold, the task to be assigned is assigned to node mn. If the overall load rate Lmn(t) of node mn at time t exceeds the set load threshold, the node is marked for demotion. Historical data is used to determine the task to be processed under the current overall load rate. The degradation factor is calculated by dividing the number of successful task executions by the total number of executions. The degradation factor is multiplied by the load matching degree of node mn to obtain the adjusted load matching degree of node mn. The load matching degrees of candidate nodes are re-sorted in descending order of value. The node with the highest adjusted load matching degree is selected and compared with the set load threshold until the comprehensive load rate Lmn(t) of node mn at time t is less than the set load threshold. At this time, the task to be assigned is assigned to node mn. The task assignment details of the node are recorded synchronously and the record is used as part of the energy saving strategy and fed back to the AI ​​computing power scheduling platform.

2. The energy-saving strategy recommendation method for computing power energy efficiency assessment according to claim 1, characterized in that: The AI ​​computing power intelligent scheduling platform collects multi-dimensional resource monitoring data from computing nodes. The specific steps include: the AI ​​computing power intelligent scheduling platform obtains the computing node resource dataset N={N1,N2,...,N...} through lightweight monitoring agents deployed on each computing node. i ,...,N p }; where N represents the acquired set of computing node resource data, N i This represents the multidimensional resource data of the i-th computing node, where i = 1, 2, ..., p, and p represents the total number of computing nodes in the cluster. For the i-th node, the collected multidimensional resource data includes Ni = {CPU...} i GPU i DISK i MEM i ,NET i }; where CPU i This represents the CPU utilization rate collected by the i-th node, and the GPU utilization rate. i DISK represents the GPU utilization rate collected by the i-th node. i MEM represents the disk input / output latency collected by the i-th node within the detection period T. i Represents the memory utilization rate collected by the i-th node, NET i This represents the network bandwidth utilization rate collected by the i-th node.

3. The energy-saving strategy recommendation method for computing power energy efficiency assessment according to claim 2, characterized in that: The comprehensive load rate of each node at time t is calculated based on the collected multidimensional resource monitoring data. The specific steps include: normalizing the collected multidimensional resource data, calculating the normalized load of the multidimensional resource data using the cluster's maximum benchmark value as the denominator, and weighting and summing the normalized multidimensional resource data loads using dynamic weighting coefficients to obtain the comprehensive load rate Li(t) of node i at time t. Here, the cluster represents a set of hardware resources formed by interconnected nodes that the AI ​​computing power intelligent scheduling platform can schedule, including CPU, GPU, disk input / output, memory, and network bandwidth. The cluster's maximum benchmark value represents the load rate of each resource type within the hardware resource set. Based on the historical operating data of the mainstream configuration nodes in the cluster, a stable operating load rate is determined. The resource usage limit of a node is defined as follows: Li(t) represents the overall load rate of node i at time t. The overall load rate reflects the stability of the node's multi-dimensional resource usage. Among them, disk I / O latency is inversely related to the overall load rate, while CPU utilization, GPU utilization, memory utilization, and network bandwidth utilization are positively related to the overall load rate. The maximum baseline value of the cluster represents the resource usage limit for stable operation for each resource type of the hardware resource set, determined based on the historical running data of the mainstream configuration nodes in the cluster. The mainstream configuration node refers to the node that accounts for a proportion of all nodes, where a is a value between 0.5 and 1. Stable operation means that the node's task success rate is greater than or equal to b under this limit, where b represents the pre-set minimum threshold for task success rate.

4. The energy-saving strategy recommendation method for computing power energy efficiency assessment according to claim 3, characterized in that: Based on real-time multidimensional resource data, nodes with real-time load rates below a set threshold are selected to form a candidate node set. The resource requirements of the task to be assigned (CPU, GPU, disk I / O, memory, and network bandwidth) are extracted, and a resource requirement vector is constructed. The load matching degree between the task and each node in the candidate node set at a specific time is calculated. Specific steps include: generating a candidate node set NC={n1,n2,...,nm} based on real-time multidimensional resource data; for the task to be assigned tx, extracting its resource requirement vector Rx=[r1,r2,r3,r4,r5]; where candidate nodes represent nodes with real-time load rates below a set threshold, NC represents the candidate node set generated based on real-time multidimensional resource data, and n1, n2,...,nm respectively... Let represent the 1st, 2nd, ..., mth candidate nodes, tx represent the xth task to be assigned, Rx represent the resource requirement vector of the xth task to be assigned, and r1, r2, r3, r4, r5 represent the resource requirements of CPU, GPU, disk I / O, memory, and network bandwidth, respectively. The load matching degree between the task to be assigned tx and node nm at time t is calculated as follows: If the resource requirement of the k-th dimension of the task to be assigned tx is greater than or equal to the remaining resource quantity of the k-th dimension of the task to be assigned tx, then the load matching degree between task tx and node nm at time t is recorded as 0, where the remaining resource quantity is calculated based on the total resource capacity minus the real-time resource usage; otherwise, the load matching degree between task tx and node nm at time t is defined as follows: Where M(x,m,t) represents the load matching degree between task tx and node nm at time t, R(m,k,t) represents the real-time resource usage of node nm in the k-th dimension at time t, r(x,k) represents the resource requirement in the k-th dimension of task tx to be assigned, and C m,k wk represents the total resource capacity of node nm in the k-th dimension, where k represents the enumerated dimension, k=1,2,...,5, representing CPU, GPU, disk input / output, memory and network bandwidth respectively, and wk represents the resource weight in the k-th dimension.

5. An energy-saving strategy recommendation system for computing power energy efficiency assessment, applied to the energy-saving strategy recommendation method for computing power energy efficiency assessment as described in any one of claims 1-4, characterized in that: The system includes: a resource data acquisition module, a comprehensive load rate calculation module, a node filtering module, and a task intelligent allocation module. The resource data acquisition module collects multi-dimensional resource monitoring data from computing nodes. The comprehensive load rate calculation module calculates the comprehensive load rate of each node at a given time based on the collected multi-dimensional resource monitoring data. The node filtering module selects nodes with real-time load rates below a set threshold based on real-time multi-dimensional resource data to form a candidate node set, extracts the resource requirements of the tasks to be assigned, constructs a resource requirement vector, and calculates the load matching degree between the task and each node in the candidate node set at a specific time. The task intelligent allocation module calculates the load matching degree of candidate nodes, sorts them from highest to lowest, selects the highest value, and compares its comprehensive load rate with the threshold. If the ratio is lower, a task is assigned; if it exceeds, it is marked as downgraded. Historical data is used to calculate a downgrade factor to adjust the matching degree, and the nodes are re-sorted and compared until a node that meets the criteria is found for allocation. The task allocation details for that node are recorded synchronously, and this record is used as part of the energy-saving strategy and fed back to the relevant department. The AI ​​computing power scheduling platform has the following components: the output of the resource data acquisition module is connected to the input of the comprehensive load rate calculation module; the output of the comprehensive load rate calculation module is connected to the input of the node filtering module; the output of the node filtering module is connected to the input of the intelligent task allocation module; the intelligent task allocation module includes a matching degree sorting unit, a load threshold judgment unit, an adjustment unit, and a task allocation execution unit. The matching degree sorting unit sorts the load matching degree of candidate nodes from highest to lowest; the load threshold judgment unit compares the comprehensive load rate of the sorted nodes with a set threshold to determine whether they meet the task allocation conditions; the adjustment unit calculates a degradation factor based on historical data, adjusts the node load matching degree, and re-sorts the nodes when the node load rate exceeds the threshold; the task allocation execution unit selects nodes with a comprehensive load rate lower than the threshold, performs task allocation operations, synchronously records the task allocation details of the node, and feeds this record back to the AI ​​computing power scheduling platform as part of the energy-saving strategy.

6. The energy-saving strategy recommendation system for computing power energy efficiency assessment according to claim 5, characterized in that: The resource data acquisition module includes a monitoring agent unit and a data aggregation unit. The monitoring agent unit means that the AI ​​computing power intelligent scheduling platform collects CPU utilization, GPU utilization, disk I / O latency, memory utilization and network bandwidth utilization data through lightweight monitoring agents deployed on each computing node. The data aggregation unit is used to summarize the data collected by each node into a resource dataset.

7. The energy-saving strategy recommendation system for computing power energy efficiency assessment according to claim 6, characterized in that: The comprehensive load rate calculation module includes a data normalization unit and a load weighting calculation unit. The data normalization unit is used to normalize multi-dimensional resource data with reference to the maximum benchmark value of the cluster. The load weighting calculation unit is used to sum the normalized load data by weighting it with dynamic weighting coefficients to obtain the comprehensive load rate of the node at a specific time.

8. The energy-saving strategy recommendation system for computing power energy efficiency assessment according to claim 7, characterized in that: The node filtering module includes a candidate node filtering unit, a task requirement parsing unit, and a load matching degree calculation unit. The candidate node filtering unit is used to filter out nodes below a set threshold based on the real-time load rate to form a candidate node set. The task requirement parsing unit is used to extract the CPU, GPU, disk input / output, memory, and network bandwidth resource requirements of the tasks to be assigned and construct a resource requirement vector. The load matching degree calculation unit is used to calculate the load matching degree between the task and each candidate node at a specific time based on the remaining node resources and the task requirements.

Images