Cloud platform computing power resource performance monitoring and real-time scheduling optimization method

Abstract

The invention relates to the technical field of cloud platform monitoring analysis, in particular to a cloud platform computing power resource performance monitoring and real-time scheduling optimization method, which comprises the following steps of: acquiring real-time performance data of heterogeneous computing nodes; performing multi-dimensional index fusion on the real-time performance data to generate a fusion performance index set containing time correlation features; constructing a dynamic computing power topological graph based on the fusion performance index set; generating an elastic scaling strategy set according to the dynamic computing power topological graph; mapping a to-be-allocated task to a physical node meeting the topology constraint condition; and carrying out topology conflict detection on the scheduling result through a feedback verification module, and dynamically correcting the parameter weight of the elastic scaling strategy set. According to the method, resource conflicts are reduced, the utilization efficiency of computing power resources is improved, and the system bottleneck problem caused by topology conflicts is avoided, so that the stability and the load balancing capacity of the system are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud platform monitoring and analysis, and in particular to a method for cloud platform computing resource performance monitoring and real-time scheduling optimization. Background Art

[0002] With the rapid development of cloud computing and big data technologies, cloud platforms have become an important infrastructure to support various application services and task processing. Cloud platforms provide powerful computing resources, but with the continuous increase in user demand and the increasing complexity of computing tasks, how to efficiently and stably schedule and manage computing resources in cloud platforms has become a key issue that needs to be urgently addressed in the current cloud computing field.

[0003] In the existing cloud platform computing resource scheduling system, resource allocation is usually based on preset strategies or static load balancing algorithms. This method is prone to resource waste or unbalanced task scheduling when computing task demands vary greatly or resource loads suddenly increase. For example, when the system load is too high, some nodes may face the problem of excessive concentration of computing resources, while other nodes are idle and cannot be fully utilized. This not only reduces resource utilization, but may also lead to a serious decline in system performance. Summary of the invention

[0004] The present invention provides a method for cloud platform computing resource performance monitoring and real-time scheduling optimization.

[0005] The cloud platform computing resource performance monitoring and real-time scheduling optimization method includes the following steps: S1: Collecting real-time performance data of heterogeneous computing nodes, where the real-time performance data includes at least CPU instruction cycle occupancy, GPU memory bandwidth utilization, and NPU matrix operation delay; S2: performing multi-dimensional index fusion on the real-time performance data to generate a fusion performance index set including time-related features; S3: construct a dynamic computing power topology map based on the fusion performance indicator set, wherein the map nodes represent physical computing units, and the edge weights represent normalized values ​​of communication bandwidth and delay between nodes; S4: Generate an elastic scaling strategy set according to the dynamic computing power topology map, wherein the elastic scaling strategy set includes topological constraints of the virtual resource pool and a load mutation probability threshold; S5: Using a priority dynamic adjustment algorithm to perform real-time scheduling, mapping the tasks to be assigned to physical nodes that meet the topology constraints; S6: Perform topology conflict detection on the scheduling result through the feedback verification module, and dynamically correct the parameter weights of the elastic scaling strategy set.

[0006] Optionally, the S1 specifically includes: S11: by installing a performance monitoring agent on each heterogeneous computing node, the CPU instruction cycle occupancy rate is obtained in real time. The instruction cycle occupancy rate is obtained by using a monitoring tool to obtain the ratio of the idle cycle to the busy cycle of each CPU core; S12: Obtain GPU memory bandwidth utilization in real time through a GPU driver interface, wherein the bandwidth utilization records bandwidth usage of GPU memory access through a GPU hardware counter; S13: Obtain the NPU matrix operation delay in real time through the management interface of the NPU hardware platform. The matrix operation delay records the delay of the NPU executing the matrix operation task through a hardware timer.

[0007] Optionally, the S2 specifically includes: S21: Normalize the real-time performance data of each heterogeneous computing node to obtain a standardized performance data set ; S22: Align the normalized performance data of multiple heterogeneous computing nodes in time series to obtain a performance data set within the time window ,in, Timestamps representing different moments; S23: The normalized performance data of different heterogeneous computing nodes are integrated using the weighted average method to obtain a comprehensive performance index set. ; S24: Sliding average method based on time series analysis for the fused comprehensive performance index set Perform smoothing to generate a fusion performance indicator set containing time-related features , to capture both long-term trends and short-term fluctuations in performance changes.

[0008] Optionally, the generation of a fusion performance indicator set containing time-related features It is expressed as: ,in, Indicates at time Comprehensive performance indicators at all times, Indicates the size of the sliding window, the length of the time window used for smoothing, Represents each time point in the sliding window, , It is a set of fusion performance indicators after sliding average smoothing, reflecting the average performance of the time points within the window.

[0009] Optionally, the S3 specifically includes: S31: Node definition: Based on the fusion performance indicator set , defining each heterogeneous computing node as a node in the graph , where each node Represents a physical computing unit (CPU, GPU or NPU). The properties of each node include As well as the load of the computing node, available resources, and current task status, Representation Node In time The set of fusion performance indicators at each moment; S32: Edge definition: edges between nodes Representation Node and nodes The communication connection between them, the edge weight Represents the normalized value of communication bandwidth and delay between nodes; S33: Dynamic update: based on real-time performance data and fused performance indicator set ,Dynamically update node attributes and edge weights. Whenever performance data is updated, recalculate the node load, task allocation status, and update the corresponding node attributes; at the same time, update the edge weights according to the real-time communication bandwidth and delay between nodes. .

[0010] Optionally, the edge weight Represents the normalized value of the communication bandwidth and delay between nodes, expressed as: ,in, Representation Node With Node The communication bandwidth between Representation Node With Node The communication delay between and Respectively represent the maximum value of bandwidth and delay between all nodes, used for normalization, is the edge weight, which indicates the performance of the communication path. The higher the value, the worse the bandwidth and delay performance of the path.

[0011] Optionally, the S4 specifically includes: S41, topological constraint definition: based on the edge weights in the dynamic computing power topology graph Based on this, the topological connection relationship between the physical computing units in the virtual resource pool is determined, and the topological constraints are defined, that is, for each node in the virtual resource pool , based on edge weights Identify other nodes with which it can communicate , define the constraint threshold , if the node and nodes The edge weight between Greater than the constraint threshold , then the task is not allowed to be scheduled to the node and nodes On the communication path between S42, load mutation probability calculation and threshold setting: setting load mutation probability threshold , set to 0.3, for each node , calculate the probability of load mutation , if the load mutation probability exceeds the threshold , it means that the risk of load mutation of the corresponding node is high; S43, generating an elastic scaling strategy set: generating an elastic scaling strategy set for the virtual resource pool according to the topology constraints and the load mutation probability threshold , elastic scaling policy set include: Topological constraints of the virtual resource pool: define the load, communication bandwidth, and communication delay constraints of each node to ensure that the topological constraints are followed during task scheduling; Load mutation probability threshold: The load mutation probability threshold set for each node deployment , guiding when to make dynamic adjustments to resources (such as expanding or shrinking the number of nodes in the virtual resource pool, increasing or decreasing resource allocation, etc.).

[0012] Optionally, the S5 specifically includes: S51, Task Priority Assessment: Based on the task's computing requirements, timeliness requirements, and resource consumption factors, each task to be assigned Calculate its task priority , expressed as: ,in, is the deadline for the task, The urgency of the task, is the resource requirement of the task, is the weight factor, which is 0.5 / 0.3 / 0.2 respectively, and is adjusted according to the specific scenario; S52, topology constraint screening: According to the topology constraint, screen out the set of physical nodes that meet the resource requirements and satisfy the topology constraint. , where each node The edge weight The following conditions are met: ; S53, Task Mapping: Based on Task Priority According to the load of physical nodes, the priority dynamic adjustment algorithm is used to adjust the tasks Mapping to physical nodes that satisfy topological constraints .

[0013] Optionally, the S6 specifically includes: S61: After the real-time scheduling of S5 is completed, the scheduling results of each task are collected, including the mapping node of the task, the task completion time, and the node load. The performance data obtained through real-time monitoring includes node load, communication delay, and bandwidth utilization; S62: Use the feedback verification module to perform topology conflict detection on the scheduling result to check whether there is any violation of the topology constraint condition. The topology conflict detection includes: Check whether the task mapping causes excessive communication latency between nodes; Check edge weights between nodes Whether the constraint threshold is exceeded , if it exceeds, it is considered that a topological conflict has occurred; By comparing the actual node connection situation with the predefined topology structure, it is determined whether there is a task allocation situation that does not meet the topology constraints; S63: If a topology conflict is detected, the feedback verification module triggers a conflict correction mechanism; S64: Dynamically correct the elastic scaling policy set: During the conflict correction process, adjust the parameter weights of the elastic scaling policy set.

[0014] Optionally, the conflict correction mechanism includes: Re-adjust the task mapping and select new nodes that meet the topological constraints; Adjust the load distribution of nodes to ensure that communication bandwidth and latency meet preset requirements.

[0015] Beneficial effects of the present invention: The present invention, by adopting a dynamic computing power topology map and a dynamic priority adjustment algorithm, accurately maps tasks to the most suitable physical nodes based on real-time performance data and topological constraints. By comprehensively evaluating the priority of tasks, the load status of nodes, and topological constraints, it is possible to optimize the task allocation process, avoid resource waste and excessive load concentration, and significantly improve the scheduling efficiency and accuracy of computing power resources. In addition, the dynamic adjustment algorithm ensures that task allocation can continuously adapt to changes in system status based on real-time feedback and analysis of historical data, further improving scheduling accuracy. The present invention introduces a topology conflict detection and dynamic correction mechanism, which can timely discover and correct topology constraint conflicts in the real-time scheduling process. Through the real-time feedback verification module, the system can quickly detect potential topology conflicts in the task allocation process, such as insufficient communication bandwidth, excessive delay between nodes, etc., and automatically adjust the topology constraint conditions when conflicts are found, and dynamically correct the threshold of edge weights. The introduction of this mechanism effectively reduces resource conflicts, improves the utilization efficiency of computing resources, avoids system bottlenecks caused by topology conflicts, and thus enhances the stability and load balancing capabilities of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 A schematic diagram of a method flow of an embodiment of the present invention; Figure 2 The figure is a schematic diagram of an elastic scaling policy set according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. At the same time, it is explained here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art may also adopt other alternatives to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0019] It should be noted that the references to "one embodiment", "embodiment", "exemplary embodiments", "some embodiments" and the like in the specification indicate that the embodiments described may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in conjunction with an embodiment, it should be within the knowledge of a person skilled in the art to implement such feature, structure or characteristic in conjunction with other embodiments (whether or not explicitly described).

[0020] In general, a term can be understood, at least in part, from its use in context. For example, depending, at least in part, on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending, at least in part, on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0021] like Figure 1-Figure 2 As shown, the cloud platform computing resource performance monitoring and real-time scheduling optimization method includes the following steps: S1: Collect real-time performance data of heterogeneous computing nodes. The real-time performance data includes at least CPU instruction cycle occupancy, GPU memory bandwidth utilization, and NPU matrix operation delay. S2: Perform multi-dimensional index fusion on real-time performance data to generate a fusion performance index set containing time-related features; S3: Build a dynamic computing power topology map based on the fusion performance indicator set. The map nodes represent physical computing units, and the edge weights represent the normalized values ​​of communication bandwidth and latency between nodes. S4: Generate an elastic scaling strategy set based on the dynamic computing power topology map. The elastic scaling strategy set includes the topology constraints of the virtual resource pool and the load mutation probability threshold. S5: Use the priority dynamic adjustment algorithm to perform real-time scheduling and map the tasks to be assigned to the physical nodes that meet the topology constraints; S6: The scheduling results are checked for topological conflicts through the feedback verification module, and the parameter weights of the elastic scaling strategy set are dynamically corrected.

[0022] S1 specifically includes: S11: By installing a performance monitoring agent on each heterogeneous computing node, the CPU instruction cycle occupancy rate is obtained in real time. The instruction cycle occupancy rate is obtained by using a monitoring tool to obtain the ratio of the idle cycle to the busy cycle of each CPU core; S12: Obtain GPU memory bandwidth utilization in real time through the GPU driver interface. The bandwidth utilization records the bandwidth usage of GPU memory access through the GPU hardware counter. S13: Obtain the NPU matrix operation delay in real time through the management interface of the NPU hardware platform. The matrix operation delay records the delay of the NPU executing the matrix operation task through the hardware timer.

[0023] 1. Calculation of CPU instruction cycle occupancy: CPU instruction cycle occupancy Calculated by the following formula: ,in, Indicates the number of busy cycles of the CPU core in one cycle. Indicates the total number of cycles of the CPU core in the same cycle (including idle cycles and busy cycles); 2. Calculation of GPU memory bandwidth utilization: GPU memory bandwidth utilization Calculated by the following formula: ,in, Indicates the actual usage of GPU memory bandwidth. express Maximum bandwidth of video memory; 3. Calculation of NPU matrix operation delay: NPU matrix operation delay Calculated by the following formula: ,in, Indicates the timestamp when NPU starts to perform matrix operations. Indicates the timestamp when the NPU completes the matrix operation.

[0024] S2 specifically includes: S21: Normalize the real-time performance data of each heterogeneous computing node to obtain a standardized performance data set , expressed as: ,in, Indicates the actual value of a performance indicator (CPU instruction cycle occupancy, GPU memory bandwidth utilization or NPU matrix operation latency) at a specific time. represents the mean value of the corresponding performance indicator in historical data, Indicates the standard deviation of the corresponding performance indicator in historical data; S22: Align the normalized performance data of multiple heterogeneous computing nodes in time series to obtain a performance data set within the time window ,in, Timestamps representing different moments; S23: The normalized performance data of different heterogeneous computing nodes are integrated using the weighted average method to obtain a comprehensive performance index set. , expressed as: ,in, represents the number of computing nodes, Indicates The weight of a computing node is determined by factors such as the load and reliability of the node. Indicates Normalized performance data of computing nodes; S24: Sliding average method based on time series analysis for the fused comprehensive performance index set Perform smoothing to generate a fusion performance indicator set containing time-related features , to capture both long-term trends and short-term fluctuations in performance changes.

[0025] Generate a fusion performance indicator set containing time-related features It is expressed as: ,in, Indicates at time Comprehensive performance indicators at all times, Indicates the size of the sliding window, the length of the time window used for smoothing, Represents each time point in the sliding window, , It is a set of fusion performance indicators after sliding average smoothing, reflecting the average performance of the time points within the window; Through the above sliding average processing, short-term fluctuations can be effectively smoothed out, long-term trends and time-related characteristics can be captured, and a more stable and time-related performance indicator set can be obtained. , providing smooth and stable performance data for subsequent dynamic scheduling optimization.

[0026] S3 specifically includes: S31: Node definition: Based on the fusion performance indicator set , defining each heterogeneous computing node as a node in the graph , where each node Represents a physical computing unit (CPU, GPU or NPU). The properties of each node include As well as the load of the computing node, available resources, and current task status, Representation Node In time The set of fusion performance indicators at each moment; S32: Edge definition: edges between nodes Representation Node and nodes The communication connection between them, the edge weight Represents the normalized value of communication bandwidth and delay between nodes; S33: Dynamic update: based on real-time performance data and fused performance indicator set ,Dynamically update node attributes and edge weights. Whenever performance data is updated, recalculate the node load, task allocation status, and update the corresponding node attributes; at the same time, update the edge weights according to the real-time communication bandwidth and delay between nodes. , ensuring that the graph reflects the real-time computing power topology. The core idea is to dynamically adjust the properties of nodes and edges in the computing power topology graph by updating real-time performance data to ensure that the graph can accurately reflect the computing resource status of the current cloud platform, thereby providing real-time support for scheduling optimization; 1. Each computing node (such as CPU, GPU or NPU) has a corresponding node in the topology map. These nodes are not just abstract representations of computing resources. Their properties will also be adjusted according to real-time performance data changes, including load and task allocation status, where: Load: As tasks are executed and load changes, the load of a node will continue to change. For example, if a node is executing a large number of computing tasks, its load may be very high; if the node is idle, the load is low. Load information is an important indicator that directly affects the decision-making of subsequent task scheduling. Task allocation status: The task execution status of each node will also affect the node's properties. For example, some nodes may have completed their tasks, while others are waiting for tasks or are busy performing specific operations. Dynamically updating this information helps to understand the current status of each node in real time.

[0027] 2. The edges in the graph represent the communication paths between two nodes, and the weights of these edges are closely related to the communication bandwidth and latency; Bandwidth and latency: The communication bandwidth and latency between nodes are important factors affecting the efficiency of task allocation. For example, when the bandwidth between two nodes is high and the latency is low, tasks can transmit data faster and more efficiently, while connections with low bandwidth or high latency may lead to performance bottlenecks. Real-time updates: As network conditions change (such as an increase or decrease in network traffic), bandwidth and latency will also change. When these data change, the corresponding edge weights need to be recalculated to reflect the latest communication status. For example, if the bandwidth increases, the edge weight may decrease because the communication performance of the path is better; while an increase in latency will increase the weight, indicating that the performance of the path has deteriorated.

[0028] 3. Through the above updates, the computing power topology map always maintains a real-time mapping of the computing resources of the cloud platform. Whenever the performance data (such as load, bandwidth, latency, etc.) changes, the nodes and edges in the map will be updated accordingly to ensure that the map accurately reflects the current status of the cloud platform. In this way, the scheduling system can make decisions based on the latest topology map, such as allocating tasks to nodes with lower load and better communication performance, thereby optimizing resource utilization and system performance.

[0029] In summary, the core goal of this part is to enable the computing power topology map to "reflect in real time" the changes in computing resources, ensuring that scheduling decisions can be made based on the most accurate current resource status rather than relying on outdated data, thereby improving the efficiency and reliability of the overall system.

[0030] Edge Weight Represents the normalized value of the communication bandwidth and delay between nodes, expressed as: ,in, Representation Node With Node The communication bandwidth between Representation Node With Node The communication delay between and Respectively represent the maximum value of bandwidth and delay between all nodes, used for normalization, is the edge weight, which indicates the performance of the communication path. The higher the value, the worse the bandwidth and delay performance of the path.

[0031] S4 specifically includes: S41, topological constraint definition: based on the edge weights in the dynamic computing power topology graph Based on this, the topological connection relationship between the physical computing units in the virtual resource pool is determined, and the topological constraints are defined, that is, for each node in the virtual resource pool , based on edge weights Identify other nodes with which it can communicate , define the constraint threshold , if the node and nodes The edge weight between Greater than the constraint threshold , then the task is not allowed to be scheduled to the node and nodes on the communication path between them, thus avoiding performance degradation; In high performance computing scenarios, constraint thresholds Take 0.3 to ensure that only high-bandwidth, low-latency communication paths are selected; In general load or resource scheduling scenarios, constraint thresholds Take 0.6 to allow a certain degree of communication delay and bandwidth bottleneck; S42, load mutation probability calculation and threshold setting: setting load mutation probability threshold , set to 0.3, for each node , calculate the probability of load mutation , if the load mutation probability exceeds the threshold , it means that the risk of load mutation of the corresponding node is high, and the resource quota of the node needs to be adjusted. Used to determine whether the virtual resource pool needs to be elastically scaled to avoid excessive load or resource waste; Load mutation probability The calculations include: Collecting Load Data: Get Node At consecutive time points , the load data ; Calculate the load variation range: Calculate the load variation range between adjacent time points, that is: ; Setting mutation threshold :According to the load fluctuation range tolerated by the system, set a load change mutation threshold , Set to ; Calculate the load mutation frequency: Calculate at consecutive time points The load change exceeds the threshold Number of times: ; Load mutation probability Calculated as: ,in, is the total number of load data points, indicating the time span of the data; Load mutation probability threshold Set by system requirements, reflecting tolerance to load fluctuations: 1. For load mutations with high tolerance, the load mutation probability threshold Set to a higher value (0.3 or 0.4); 2. If tolerance to load variations is low, set to a lower value (0.1 or 0.2).

[0032] S43, generating an elastic scaling strategy set: generating an elastic scaling strategy set for the virtual resource pool according to the topology constraints and the load mutation probability threshold , elastic scaling policy set include: Topological constraints of the virtual resource pool: define the load, communication bandwidth, and communication delay constraints of each node to ensure that the topological constraints are followed during task scheduling; Load mutation probability threshold: The load mutation probability threshold set for each node deployment , guiding when to make dynamic adjustments to resources (such as expanding or shrinking the number of nodes in the virtual resource pool, increasing or decreasing resource allocation, etc.).

[0033] S5 specifically includes: S51, Task Priority Assessment: Based on the task's computing requirements, timeliness requirements, and resource consumption factors, each task to be assigned Calculate its task priority , expressed as: ,in, is the deadline for the task, The urgency of the task, is the resource requirement of the task, is the weight factor, which is 0.5 / 0.3 / 0.2 respectively, and is adjusted according to the specific scenario; S52, topology constraint screening: According to the topology constraint, screen out the set of physical nodes that meet the resource requirements and satisfy the topology constraint. , where each node The edge weight The following conditions are met: ; S53, Task Mapping: Based on Task Priority According to the load of physical nodes, the priority dynamic adjustment algorithm is used to adjust the tasks Mapping to physical nodes that satisfy topological constraints , the mapping rules are: Select the task with the highest priority ; In the physical node collection In the example, select the node with the lightest current load and that satisfies the topological constraints. , so that the task The resource requirements of the physical nodes are met and the load of the physical nodes does not exceed their own capacity threshold.

[0034] S6 specifically includes: S61: After the real-time scheduling of S5 is completed, the scheduling results of each task are collected, including the mapping node of the task, the task completion time, and the node load. The performance data obtained through real-time monitoring includes node load, communication delay, and bandwidth utilization; S62: Use the feedback verification module to perform topology conflict detection on the scheduling result to check whether there is any violation of the topology constraint condition. The topology conflict detection includes: Check whether the task mapping causes excessive communication latency between nodes; Check edge weights between nodes Whether the constraint threshold is exceeded , if it exceeds, it is considered that a topological conflict has occurred; By comparing the actual node connection situation with the predefined topology structure, it is determined whether there is a task allocation situation that does not meet the topology constraints; S63: If a topology conflict is detected, the feedback verification module triggers a conflict correction mechanism; S64: Dynamically modify the elastic scaling policy set: During the conflict modification process, the parameter weights of the elastic scaling policy set are adjusted. The adjustment steps include: Adjust constraint thresholds in topology constraints based on conflict detection results , in order to adapt to the new task allocation mode, if the current system resource constraints are too strict, causing the communication between most nodes to exceed the constraint threshold, you can appropriately increase To increase the flexibility of the system and allow more task allocation; if conflict detection shows that the system is too dependent on certain high-bandwidth or low-latency communication links, resulting in excessive edge weights between some nodes, which may cause excessive resource competition or load imbalance, you can appropriately reduce , to strengthen the control of the communication quality between nodes.

[0035] Recalculate load mutation probability .

[0036] Conflict correction mechanisms include: Re-adjust the task mapping and select new nodes that meet the topological constraints; Adjust the load distribution of nodes to ensure that communication bandwidth and latency meet preset requirements.

[0037] The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, but those skilled in the art can fully understand the present invention without the description of these details. In addition, in order to avoid unnecessary confusion about the essence of the present invention, well-known methods, processes, procedures, components and circuits are not described in detail.

[0038] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A cloud platform computing resource performance monitoring and real-time scheduling optimization method, characterized in that: The following steps are involved: S1: Collecting real-time performance data of heterogeneous computing nodes, where the real-time performance data includes at least CPU instruction cycle occupancy, GPU memory bandwidth utilization, and NPU matrix operation delay; S2: performing multi-dimensional index fusion on the real-time performance data to generate a fusion performance index set including time-related features; S3: construct a dynamic computing power topology map based on the fusion performance indicator set, wherein the map nodes represent physical computing units, and the edge weights represent normalized values ​​of communication bandwidth and delay between nodes; S4: Generate an elastic scaling strategy set according to the dynamic computing power topology map, wherein the elastic scaling strategy set includes topological constraints of the virtual resource pool and a load mutation probability threshold; S5: Using a priority dynamic adjustment algorithm to perform real-time scheduling, mapping the tasks to be assigned to physical nodes that meet the topology constraints; S6: Perform topology conflict detection on the scheduling result through the feedback verification module, and dynamically correct the parameter weights of the elastic scaling strategy set.

2. The cloud platform computing resource performance monitoring and real-time scheduling optimization method according to claim 1 is characterized in that: The S1 specifically includes: S11: by installing a performance monitoring agent on each heterogeneous computing node, the CPU instruction cycle occupancy rate is obtained in real time. The instruction cycle occupancy rate is obtained by using a monitoring tool to obtain the ratio of the idle cycle to the busy cycle of each CPU core; S12: Obtain GPU memory bandwidth utilization in real time through a GPU driver interface, wherein the bandwidth utilization records bandwidth usage of GPU memory access through a GPU hardware counter; S13: Obtain the NPU matrix operation delay in real time through the management interface of the NPU hardware platform. The matrix operation delay records the delay of the NPU executing the matrix operation task through a hardware timer.

3. The cloud platform computing resource performance monitoring and real-time scheduling optimization method according to claim 1 is characterized in that: The S2 specifically includes: S21: Normalize the real-time performance data of each heterogeneous computing node to obtain a standardized performance data set ; S22: Align the normalized performance data of multiple heterogeneous computing nodes in time series to obtain a performance data set within the time window ,in, Timestamps representing different moments; S23: The normalized performance data of different heterogeneous computing nodes are integrated using the weighted average method to obtain a comprehensive performance index set. ; S24: Sliding average method based on time series analysis for the fused comprehensive performance index set Perform smoothing to generate a fusion performance indicator set containing time-related features , to capture both long-term trends and short-term fluctuations in performance changes.

4. The cloud platform computing resource performance monitoring and real-time scheduling optimization method according to claim 3 is characterized in that: The generation of the fusion performance indicator set containing time-related features is expressed as: ,in, Indicates at time Comprehensive performance indicators at all times, Indicates the size of the sliding window, the length of the time window used for smoothing, Represents each time point in the sliding window, , It is a set of fusion performance indicators after sliding average smoothing, reflecting the average performance of the time points within the window.

5. The cloud platform computing resource performance monitoring and real-time scheduling optimization method according to claim 1 is characterized in that: The S3 specifically includes: S31: Node definition: Based on the fusion performance indicator set , defining each heterogeneous computing node as a node in the graph , where each node Represents a physical computing unit. The properties of each node include As well as the load of the computing node, available resources, and current task status, Representation Node In time The set of fusion performance indicators at each moment; S32: Edge definition: edges between nodes Representation Node and nodes The communication connection between them, the edge weight Represents the normalized value of communication bandwidth and delay between nodes; S33: Dynamic update: based on real-time performance data and fused performance indicator set ,Dynamically update node attributes and edge weights. Whenever performance data is updated, recalculate the node load, task allocation status, and update the corresponding node attributes; at the same time, update the edge weights according to the real-time communication bandwidth and delay between nodes. .

6. The cloud platform computing resource performance monitoring and real-time scheduling optimization method according to claim 5 is characterized in that: The edge weight Represents the normalized value of the communication bandwidth and delay between nodes, expressed as: ,in, Representation Node With Node The communication bandwidth between Representation Node With Node The communication delay between and Respectively represent the maximum value of bandwidth and delay between all nodes, used for normalization, is the edge weight, which indicates the performance of the communication path. The higher the value, the worse the bandwidth and delay performance of the path.

7. The cloud platform computing resource performance monitoring and real-time scheduling optimization method according to claim 6 is characterized in that: The S4 specifically includes: S41, topological constraint definition: based on the edge weights in the dynamic computing power topology graph Based on this, the topological connection relationship between the physical computing units in the virtual resource pool is determined, and the topological constraints are defined, that is, for each node in the virtual resource pool , based on edge weights Identify other nodes with which it can communicate , define the constraint threshold , if the node and nodes The edge weight between Greater than the constraint threshold , then the task is not allowed to be scheduled to the node and nodes On the communication path between S42, load mutation probability calculation and threshold setting: setting load mutation probability threshold , for each node , calculate the probability of load mutation , if the load mutation probability exceeds the threshold , it means that the risk of load mutation of the corresponding node is high; S43, generating an elastic scaling strategy set: generating an elastic scaling strategy set for the virtual resource pool according to the topology constraints and the load mutation probability threshold , elastic scaling policy set include: Topological constraints of the virtual resource pool: define the load, communication bandwidth, and communication delay constraints of each node to ensure that the topological constraints are followed during task scheduling; Load mutation probability threshold: The load mutation probability threshold set for each node deployment , guiding when to make dynamic adjustments to resources.

8. The cloud platform computing resource performance monitoring and real-time scheduling optimization method according to claim 7 is characterized in that: The S5 specifically includes: S51, Task Priority Assessment: Based on the task's computing requirements, timeliness requirements, and resource consumption factors, each task to be assigned Calculate its task priority , expressed as: ,in, is the deadline of the task, The urgency of the task, is the resource requirement of the task, is the weight factor; S52, topology constraint screening: According to the topology constraint, screen out the set of physical nodes that meet the resource requirements and satisfy the topology constraint. , where each node The edge weight The following conditions are met: ; S53, Task Mapping: Based on Task Priority According to the load of physical nodes, the priority dynamic adjustment algorithm is used to adjust the tasks Mapping to physical nodes that satisfy topological constraints .

9. The cloud platform computing resource performance monitoring and real-time scheduling optimization method according to claim 7 is characterized in that: The S6 specifically includes: S61: After the real-time scheduling of S5 is completed, the scheduling results of each task are collected, including the mapping node of the task, the task completion time, and the node load. The performance data obtained through real-time monitoring includes node load, communication delay, and bandwidth utilization; S62: Use the feedback verification module to perform topology conflict detection on the scheduling result to check whether there is any violation of the topology constraint condition. The topology conflict detection includes: Check whether the task mapping causes excessive communication latency between nodes; Check edge weights between nodes Whether the constraint threshold is exceeded ; By comparing the actual node connection situation with the predefined topology structure, it is determined whether there is a task allocation situation that does not meet the topology constraints; S63: If a topology conflict is detected, the feedback verification module triggers a conflict correction mechanism; S64: Dynamically correct the elastic scaling policy set: During the conflict correction process, adjust the parameter weights of the elastic scaling policy set.

10. The cloud platform computing resource performance monitoring and real-time scheduling optimization method according to claim 9 is characterized in that: The conflict correction mechanism includes: Re-adjust the task mapping and select new nodes that meet the topological constraints; Adjust the load distribution of nodes to ensure that communication bandwidth and latency meet preset requirements.

Images