Calculation force measurement method based on multi-dimensional calculation force perception

Abstract

The invention relates to the field of computer system performance evaluation, and discloses a multi-dimensional computing power perception-based computing power measurement method, which comprises the following steps of S1, acquiring static performance data of computing power nodes; s2, collecting dynamic state data of the computing power node in real time; s3, performing normalization processing on the static performance data and the dynamic state data to eliminate dimensional difference; s4, based on the normalized static performance data and dynamic state data, generating a comprehensive computing power score through weighted calculation; s5, utilizing a classification and regression tree algorithm to perform segmentation grading on the comprehensive computing power score, and dividing computing power nodes into different performance grades; and S6, according to the task demand and the performance level of the computing power node, dynamically allocating computing resources. According to the method, by collecting the static performance data and the dynamic state data, the comprehensiveness and accuracy of data processing are improved, comprehensive computing power score generation is facilitated, and computing power distribution is assisted.

Description

Technical Field

[0001] This invention relates to the field of computer system performance evaluation, and in particular to a computing power method based on multi-dimensional computing power perception. Background Technology

[0002] Computer system performance evaluation refers to the process of quantitatively analyzing and evaluating various performance indicators of a computer system through a series of techniques and methods. It aims to comprehensively understand the operating performance of a computer system under different workloads, including multiple aspects such as computing power, storage capacity, and communication capabilities. Computer system performance evaluation determines whether the system meets expected goals (such as real-time performance, concurrent processing capacity, or energy efficiency ratio) by quantitatively analyzing key system indicators (such as response time, throughput, and resource utilization) and identifies performance bottlenecks. Computing power is a commonly used method for computer system performance evaluation.

[0003] Existing computing power metrics reveal the task's dependence on computing power or memory bandwidth by quantifying the ratio of floating-point operations to data access in a computational task, thereby maximizing hardware resource utilization. However, this approach processes and analyzes static and dynamic data separately, resulting in limited and ambiguous data processing and leading to errors in computing power allocation. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a computing power method based on multi-dimensional computing power perception, which aims to improve the problem of errors in computing power allocation caused by existing computing power methods processing and analyzing static and dynamic data separately.

[0005] In a first aspect, the present invention provides the following technical solution: a computing power method based on multi-dimensional computing power perception, comprising the following steps.

[0006] S1. Collect static performance data of computing nodes, wherein the static performance data includes computing power, memory performance, storage performance, communication capability and energy efficiency ratio;

[0007] S2. Real-time acquisition of dynamic status data of computing nodes, wherein the dynamic status data includes load status, throughput, latency, resource contention, failure rate, and temperature and power consumption.

[0008] S3. Normalize static performance data and dynamic state data to eliminate dimensional differences;

[0009] S4. Based on the normalized static performance data and dynamic status data, a comprehensive computing power score is generated through weighted calculation.

[0010] S5. Use classification and regression tree algorithms to segment and classify the comprehensive computing power score, and divide the computing power nodes into different performance levels.

[0011] S6. Dynamically allocate computing resources based on task requirements and the performance level of computing nodes.

[0012] Preferably, the computing power includes the number of CPU cores, single-core frequency, multi-core performance, and floating-point operation capability; the memory performance includes memory capacity, memory bandwidth, and memory read / write speed; the storage performance includes storage capacity and storage read / write speed; the communication capability includes network bandwidth, network latency, and data transmission rate; and the energy efficiency ratio includes computing power per unit power consumption, memory performance per unit power consumption, and storage performance per unit power consumption.

[0013] Preferably, the load status includes CPU load rate, memory utilization rate, storage utilization rate, and network bandwidth utilization rate; the throughput includes the processing speed of computing tasks, data transmission throughput, and storage read / write throughput; the latency includes computing latency, network latency, and storage access latency; the resource contention includes CPU resource contention, memory resource contention, storage resource contention, and network resource contention; the failure rate includes hardware failure rate, software failure rate, and system crash rate; the temperature includes CPU temperature, memory temperature, and storage device temperature; and the power consumption includes CPU power consumption, memory power consumption, storage device power consumption, and network device power consumption.

[0014] Preferably, the normalization process uses the following formula:

[0015]

[0016] Where P is the current indicator value, P min P is the minimum value of this indicator. max This represents the maximum value of the indicator.

[0017] Preferably, the formula for calculating the comprehensive computing power score is:

[0018] S=α·p static +β·D dynamic

[0019] Where S is the overall score, p static For static performance scoring, D dynamic For dynamic state scoring, α and β are weight coefficients, and α+β=1. In computationally intensive tasks, the static performance weight α is increased, and in tasks with high real-time requirements, the dynamic state weight β is increased.

[0020] Preferably, step S5 further includes:

[0021] S501. Collect comprehensive scoring data from a large number of computing power nodes to construct a dataset;

[0022] S502. Use the CART algorithm to segment the comprehensive score data, and determine the optimal segmentation point by minimizing the variance within nodes and maximizing the difference between nodes;

[0023] S503 divides computing nodes into high-performance, medium-performance, and low-performance levels.

[0024] Preferably, step S6 further includes:

[0025] For computationally intensive tasks, priority should be given to allocating them to high-performance computing nodes;

[0026] For tasks with high real-time requirements, priority should be given to computing nodes with high dynamic status scores;

[0027] For light-load or test tasks, allocate them to computing nodes with lower performance levels;

[0028] When a computing node experiences resource overload or failure, tasks are redistributed to other available nodes based on real-time dynamic status data.

[0029] Secondly, the present invention provides the following technical solution: a computing power system based on multi-dimensional computing power perception, comprising:

[0030] The static performance acquisition module is used to perform step S1 to acquire static performance data of the computing nodes;

[0031] The dynamic status monitoring module is used to collect dynamic status data of computing nodes in real time during step S2.

[0032] The data processing module is used to perform step S3 to normalize the static performance data and dynamic status data;

[0033] The comprehensive scoring module is used to perform step S4 based on normalized static performance data and dynamic status data;

[0034] The grading module is used to execute step S5, which segments and grades the computing power nodes using the CART algorithm.

[0035] The resource scheduling module is used to execute step S6, which dynamically allocates computing resources based on task requirements and node hierarchical results.

[0036] Thirdly, the invention provides the following technical solution: a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described image recognition method.

[0037] Fourthly, the present invention provides the following technical solution: a readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the above-described image recognition method.

[0038] The present invention has the following beneficial effects:

[0039] 1. In this invention, by collecting static performance data including computing power, memory performance, storage performance, communication capabilities, and energy efficiency ratio, as well as dynamic status data such as load status, throughput, and latency, the performance of computing nodes can be comprehensively measured from multiple perspectives. At the same time, the static and dynamic data are normalized to eliminate the differences in the units of different data types, so that various data can be compared and analyzed under the same standard, ensuring the accuracy and reliability of the results, providing a basis for subsequent resource allocation, thereby improving the comprehensiveness and accuracy of data processing, helping to generate comprehensive computing power scores, and assisting in computing power allocation.

[0040] 2. In this invention, by generating a comprehensive computing power score and flexibly adjusting the static performance weight α and dynamic state weight β according to different task characteristics, the comprehensive score reflects the computing power situation of the task, providing support for the accurate matching of tasks and computing power nodes. The comprehensive computing power score is segmented and graded using classification and regression tree algorithms, thereby classifying computing power nodes into high-performance, medium-performance and low-performance levels, and facilitating the rapid selection of suitable computing power nodes according to task requirements.

[0041] 3. In this invention, the dynamic status monitoring module collects the dynamic status data of the computing nodes in real time. Combined with the resource scheduling module, computing resources can be dynamically allocated according to the real-time status of the nodes and task requirements. This allows for timely responses to changes in node resources and different task requirements, thereby ensuring the stability of the system and the timeliness of task execution. It also avoids task delays or failures caused by unreasonable resource allocation and improves the overall performance of the system.

[0042] 4. In this invention, through the collaborative work of various modules, the entire process from data collection, processing, scoring, and grading to resource allocation is automated. Furthermore, based on scientific algorithms and multi-dimensional data processing, the potential of computing nodes can be tapped and resources can be rationally allocated, thereby improving the system's computing power utilization efficiency and performance, and meeting the diverse computing task requirements. Attached Figure Description

[0043] Figure 1 This is a flowchart of a computing power method based on multi-dimensional computing power perception proposed in this invention.

[0044] Figure 2 This is a system framework diagram of a computing power method based on multi-dimensional computing power perception proposed in this invention. Detailed Implementation

[0045] The technical solutions in 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.

[0046] Example 1

[0047] Reference Figure 1 and Figure 2 In the first embodiment of the present invention, the present invention provides a computing power method based on multi-dimensional computing power perception, comprising the following steps:

[0048] S1. Collect static performance data of computing nodes, including computing power, memory performance, storage performance, communication capability and energy efficiency ratio;

[0049] S2. Real-time acquisition of dynamic status data of computing nodes, including load status, throughput, latency, resource contention, failure rate, temperature and power consumption.

[0050] S3. Normalize static performance data and dynamic state data to eliminate dimensional differences;

[0051] S4. Based on the normalized static performance data and dynamic status data, a comprehensive computing power score is generated through weighted calculation.

[0052] S5. Use the Classification and Regression Tree (CART) algorithm to segment and classify the comprehensive computing power score, and divide the computing power nodes into different performance levels;

[0053] S6. Dynamically allocate computing resources based on task requirements and the performance level of computing nodes.

[0054] Specifically, S1, computing power acquisition focuses on the following aspects: For GPUs: obtaining peak FLOPS (floating-point operations per second) data at FP32 or FP16 precision through GPU device drivers; For NPUs: obtaining their core TOPS (tensor operations per second) metric in AI inference tasks using the NPU's hardware description file and accompanying development tools; For CPUs: obtaining SPEC benchmark scores by running the SPEC benchmark program to quantify CPU performance in general computing tasks. Additionally, MIPS (million-point operations per second) can also be obtained from the CPU's device drivers or operating system system information. Data related to millions of instructions; memory performance is assessed in the following ways: memory capacity: obtained from the operating system's memory management module or the memory controller information of the hardware device; memory bandwidth: using specialized memory bandwidth testing tools, a series of read and write operations are performed in memory to measure the data transfer rate, thus determining the memory bandwidth (GB / s); cache performance: the size of L1, L2, and L3 caches can be obtained from the CPU's hardware specifications or through the operating system's hardware information query tool; cache access speed can be evaluated by measuring cache hit time and miss time using specialized cache performance testing tools. Storage performance is assessed in the following ways: storage capacity: for local storage devices, the total capacity of the storage device can be obtained through the operating system's disk management tools or related functions of the file system. Storage bandwidth: Using specialized storage performance testing tools to simulate different I / O load patterns, the bandwidth (GB / s) of the storage device during data read and write operations is measured; IOPS (Input / Output Operations Per Second): Using storage performance testing tools, by setting up random I / O test scenarios, the IOPS performance of the storage system when processing a large number of small data requests is measured; Communication capability acquisition includes the following: Network bandwidth: Using network performance testing tools, such as iperf or nettperf, data transmission tests are performed at both ends of the network link to measure network bandwidth (Gbps); Latency: The ping command or a dedicated network latency testing tool, such as PingPlotter, can be used to send ICMP packets to the target address and measure the round-trip time (ms) to obtain network latency. Energy efficiency ratio (EER) is calculated by measuring the power consumption and computing power of computing devices during operation. For power consumption measurement, a power meter is directly connected to the power input of the device to monitor the device's power consumption (W) in real time. Combined with the previously collected computing power data, FLOPS / W (floating-point operations per watt) is calculated, which serves as the core indicator of EER.

[0055] S2. Regarding load status acquisition, CPU / GPU utilization can be obtained through system monitoring tools. In Linux systems, CPU utilization can be calculated using the `top` command or the `proc` filesystem, while GPU utilization is obtained through the NVIDIASMI tool. Memory utilization is calculated through the operating system's memory management module. In Linux systems, this is done by reading the ` / proc / meminfo` file, while in Windows systems, it is obtained through a performance monitor. Storage utilization is calculated through disk management tools or storage management software. For throughput acquisition, data throughput can be measured using network traffic monitoring tools such as `nload` or `iftop` to measure network interface speeds, while at the storage device level, it is measured using tools like `IOMeter` or `FIO`. Task throughput is recorded by the task scheduling system, showing the number of tasks completed per unit of time. For late-event data collection, task latency is recorded by the task scheduling system as the time interval between submission and completion; network latency can be measured using the ping command or PingPlotter tool, or precisely measured using packet timestamps; for resource contention data collection, cache contention is analyzed using the Cachegrind tool to monitor program cache usage; memory contention is detected using the Valgrind tool to detect memory access conflicts in multi-threaded or multi-process environments; failure rate data collection records the number of failures per unit time through system logs and hardware failure monitoring modules; for temperature and power consumption data collection, real-time temperature is obtained through hardware sensors, with Linux systems reading temperature data through the sysfs file system and Windows systems obtaining it through hardware monitoring software; current power consumption can be measured using a power meter or obtained through hardware monitoring functions.

[0056] S3. Normalize the static performance data and dynamic status data to eliminate differences in units. Static performance data (computing power, memory performance, storage performance, communication capability, and energy efficiency ratio) are normalized using a formula. Processing is performed, where P is the current indicator value, P min P is the minimum value of this indicator. max This represents the maximum value of the indicator. Dynamic status data (load status, throughput, latency, resource contention, failure rate, temperature, and power consumption) are also normalized using the same formula. Through normalization, indicators with different dimensions are converted into dimensionless values ​​in the 0-1 range, which facilitates subsequent comprehensive calculation and comparison.

[0057] S4. Based on the normalized static performance data and dynamic state data, a comprehensive computing power score is generated through weighted calculation. The weights can be determined using the entropy weight method (based on the reciprocal normalization of information entropy). Let the static performance weight be α, the dynamic state weight be β, and α + β = 1. The static performance score p static The dynamic state score D is obtained by weighted summation of normalized indicators such as computing power, memory performance, storage performance, communication capability, and energy efficiency ratio. dynamicThe comprehensive computing power score S = α·p is obtained by weighted summation of normalized indicators such as load status, throughput, latency, resource contention, failure rate, temperature, and power consumption. static +β·D dynamic This weighted calculation method combines static performance and dynamic status data to generate a score that reflects the overall performance of the computing nodes.

[0058] S5. The Classification and Regression Tree (CART) algorithm is used to segment and classify the comprehensive computing power score and divide the performance level of computing power nodes. First, a large amount of comprehensive computing power score data and task performance data of computing power nodes are collected to build a dataset covering different hardware configurations and application scenarios. Then, the CART algorithm is trained using this dataset. By continuously splitting the dataset in two, the optimal split point is found to segment the comprehensive computing power score and realize the performance classification of computing power nodes.

[0059] Computing power includes the number of CPU cores, single-core frequency, multi-core performance, and floating-point arithmetic capability; memory performance includes memory capacity, memory bandwidth, and memory read / write speed; storage performance includes storage capacity and storage read / write speed; communication power includes network bandwidth, network latency, and data transfer rate; energy efficiency includes computing power per unit power consumption, memory performance per unit power consumption, and storage performance per unit power consumption.

[0060] Specifically, when evaluating computing power, the number of CPU cores determines the potential for parallel processing, single-core frequency affects instruction execution speed, multi-core performance needs to be comprehensively evaluated through professional tests (such as CINEBENCH) to assess core collaboration efficiency, and floating-point operation power (FLOPS) directly affects the performance of professional scenarios such as scientific computing; in terms of memory performance, capacity determines the scale of data processing, and bandwidth (GB / s) and read / write speed affect data transmission efficiency, which is especially crucial for high-frequency read / write applications such as databases; in storage performance, capacity ensures data storage needs, and read / write speed (such as SSD IOPS) improves the efficiency of scenarios such as big data analysis; communication capabilities are centered on bandwidth (Gbps) and latency (ms), high bandwidth reduces the transmission time of distributed computing, and low latency ensures the experience of real-time applications; energy efficiency ratio is measured by indicators such as FLOPS / W, reflecting the energy utilization rate of computing equipment and helping to reduce data center operating costs.

[0061] Load status includes CPU load rate, memory utilization, storage utilization, and network bandwidth utilization; throughput includes the processing speed of computing tasks, data transfer throughput, and storage read / write throughput; latency includes computing latency, network latency, and storage access latency; resource contention includes CPU resource contention, memory resource contention, storage resource contention, and network resource contention; failure rate includes hardware failure rate, software failure rate, and system crash rate; temperature includes CPU temperature, memory temperature, and storage device temperature; power consumption includes CPU power consumption, memory power consumption, storage device power consumption, and network device power consumption.

[0062] Specifically, load status includes CPU load rate (obtained through Task Manager or the `top` command, reflecting the percentage of time the CPU is busy with tasks), memory usage rate (monitored through Resource Monitor or system functions, reflecting the percentage of memory used), storage usage rate (viewed through the `df` command or storage management system, reflecting the percentage of storage used), and network bandwidth usage rate (obtained through tools such as Nagios, reflecting the percentage of network bandwidth used). Throughput covers computing task processing speed (calculated by the task scheduling system, reflecting task processing capacity), data transfer throughput (measured through tools such as Wireshark, reflecting data transfer efficiency), and storage read / write throughput (obtained through tools such as Iometer, reflecting storage read / write efficiency). Latency includes computation latency (measured through tools such as SPECCPU, affected by various factors), network latency (affecting the experience of real-time applications), and storage access latency (measuring latency through various tools such as SPECCPU, affected by various factors). Measured using tools such as fio (affected by factors such as storage device type), resource contention includes CPU resource contention (analyzed using tools such as oprofile), memory resource contention (detected using tools such as Valgrind), storage resource contention (monitored using tools such as iostat), and network resource contention (discovered using tools such as Snort). Failure rates include hardware failure rate (statistically calculated using tools such as OpenManage), software failure rate (recorded through error logs), and system crash rate (statistically calculated using tools such as Zabbix). Regarding temperature, CPU temperature, memory temperature, and storage device temperature can be monitored in real time using sensors and related monitoring software. Excessive temperature can affect performance and stability. Regarding power consumption, CPU power consumption, memory power consumption, storage device power consumption, and network device power consumption are measured using relevant tools and product specifications. Power consumption monitoring helps optimize energy utilization and reduce costs.

[0063] The normalization process uses the following formula:

[0064]

[0065] Where P is the current indicator value, P min P is the minimum value of this indicator. max This represents the maximum value of the indicator.

[0066] Specifically, in the normalization process, the formula standardizes performance indicators with different dimensions and value ranges to the [0,1] interval, thereby achieving unified comparison and comprehensive calculation between indicators. This normalization method eliminates the incomparability caused by differences in dimensions and value ranges, providing a basis for the reasonable weighted calculation of various indicators in the comprehensive computing power scoring model.

[0067] The formula for calculating the comprehensive computing power score is as follows:

[0068] S=α·p static +β·D dynamic

[0069] Where S is the overall score, p static For static performance scoring, D dynamic For dynamic state scoring, α and β are weight coefficients, and α+β=1. In computationally intensive tasks, the static performance weight α is increased, and in tasks with high real-time requirements, the dynamic state weight β is increased.

[0070] Specifically, the comprehensive computing power score calculation formula includes a static performance score that normalizes and weights the static performance indicators mentioned above, such as computing power, memory performance, storage performance, communication capability, and energy efficiency ratio, according to their respective weights. Similarly, the dynamic status score is calculated by normalizing and weighting the dynamic status indicators, such as load status, throughput, latency, resource contention, failure rate, temperature, and power consumption, according to their respective weights. Through this calculation method, the comprehensive score can fully and reasonably reflect the overall capabilities of computing nodes in different scenarios, providing a quantitative basis for the classification and task allocation of computing nodes. Please summarize this content using synonyms and integrate it into a single paragraph.

[0071] Step S5 also includes:

[0072] S501. Collect comprehensive scoring data from a large number of computing power nodes to construct a dataset;

[0073] S502. Use the CART algorithm to segment the comprehensive score data, and determine the optimal segmentation point by minimizing the variance within nodes and maximizing the difference between nodes;

[0074] S503 divides computing nodes into high-performance, medium-performance, and low-performance levels.

[0075] Specifically, in step S501, when collecting comprehensive score data from a large number of computing nodes to construct the dataset, it is necessary to cover data from different types of computing units (such as CPU, GPU, NPU, etc.), different operating environments (such as data centers with different temperatures and humidity, and edge computing nodes with different network bandwidths), and different load states (from idle to full load) to ensure the comprehensiveness and representativeness of the data. In step S502, the CART algorithm is used to segment the comprehensive score data. The algorithm iterates through all possible segmentation points, calculates the intra-node variance and inter-node variance (such as the Gini index) after the segmentation, finds the segmentation point that minimizes the intra-node variance and maximizes the inter-node variance, and recursively... The dataset is divided into multiple subsets until a stopping condition is met (such as the number of child node data or variance being less than a set threshold). In step S503, based on the split points determined by the CART algorithm, the computing power nodes are divided into high-performance, medium-performance, and low-performance levels. High-performance nodes (high overall score) are suitable for handling computationally intensive tasks, such as large-scale AI model training; medium-performance nodes (medium overall score) meet routine business needs, such as daily office work and general data analysis; low-performance nodes (low overall score) are used for light-load tasks or testing environments, such as web application testing and software functional testing. This hierarchical approach enables refined management and rational utilization of computing power nodes.

[0076] Step S6 also includes:

[0077] For computationally intensive tasks, priority should be given to allocating them to high-performance computing nodes;

[0078] For tasks with high real-time requirements, priority should be given to computing nodes with high dynamic status scores;

[0079] For light-load or test tasks, allocate them to computing nodes with lower performance levels;

[0080] When a computing node experiences resource overload or failure, tasks are redistributed to other available nodes based on real-time dynamic status data.

[0081] Specifically, in step S6, resources are allocated based on task type and computing node performance: computationally intensive tasks (such as large-scale numerical simulations) are preferentially allocated to high-performance nodes, as they have high core counts, high-frequency CPUs, and high-performance GPUs, enabling them to quickly process a large number of computational tasks and shorten execution time; tasks with high real-time requirements (such as online games and video live streaming transcoding) are preferentially allocated to nodes with high dynamic status scores, as these nodes have low latency and high throughput, enabling them to quickly respond to user requests and ensure a smooth experience; light-load tasks (such as text processing and small database queries) and test tasks (such as functional testing and algorithm verification) are allocated to low-performance nodes to avoid resource waste and save costs. When a node is overloaded (such as CPU load exceeding 90%) or fails, the system reallocates tasks to available nodes with lower loads based on real-time dynamic status data to ensure smooth task execution.

[0082] Example 2:

[0083] Referring to the figures, in a second embodiment of the present invention, the present invention provides an image recognition system, comprising:

[0084] The static performance acquisition module is used to perform step S1 to acquire static performance data of the computing nodes;

[0085] The dynamic status monitoring module is used to collect dynamic status data of computing nodes in real time during step S2.

[0086] The data processing module is used to perform step S3 to normalize the static performance data and dynamic status data;

[0087] The comprehensive scoring module is used to perform step S4 based on normalized static performance data and dynamic status data;

[0088] The grading module is used to execute step S5, which segments and grades the computing power nodes using the CART algorithm.

[0089] The resource scheduling module is used to execute step S6, which dynamically allocates computing resources based on task requirements and node hierarchical results.

[0090] Specifically, the static performance acquisition module collects computing power data (such as GPU FLOPS, NPU TOPS, CPU SPEC score and MIPS), memory performance (capacity, bandwidth, cache size and access speed), storage performance (capacity, bandwidth, IOPS), communication capabilities (network bandwidth and latency), and energy efficiency ratio data from computing nodes using various tools; the dynamic status monitoring module collects real-time load status (CPU / GPU utilization, memory utilization, storage utilization), throughput, latency, resource contention, failure rate, temperature, and power consumption data; the data processing module processes the collected static data... The system normalizes dynamic data to eliminate dimensional differences; the comprehensive scoring module calculates the comprehensive computing power score based on the normalized data and the weights determined by the entropy weight method; the grading module uses the CART algorithm to segment and grade computing power nodes into high-performance, medium-performance, and low-performance levels; the resource scheduling module dynamically allocates resources according to task requirements and node grading results, prioritizing the allocation of computationally intensive tasks to high-performance nodes, real-time tasks to nodes with high dynamic status scores, and light-load and test tasks to low-performance nodes, and reallocates tasks when nodes are overloaded or fail, in order to optimize resource utilization and ensure task execution.

[0091] Example 3

[0092] In a third embodiment of the present invention, based on the same inventive concept, the present invention proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the image recognition method described in the above embodiments.

[0093] Example 4

[0094] According to the fourth embodiment of the present invention, based on the same inventive concept, the present invention proposes a computer device comprising: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory to perform the image recognition method of the above embodiment.

[0095] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0096] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A computational power method based on multi-dimensional computational power perception, characterized in that, Includes the following steps: S1. Collect static performance data of computing nodes, wherein the static performance data includes computing power, memory performance, storage performance, communication capability and energy efficiency ratio; S2. Real-time acquisition of dynamic status data of computing nodes, wherein the dynamic status data includes load status, throughput, latency, resource contention, failure rate, and temperature and power consumption. S3. Normalize static performance data and dynamic state data to eliminate dimensional differences; S4. Based on the normalized static performance data and dynamic status data, a comprehensive computing power score is generated through weighted calculation. S5. Use classification and regression tree algorithms to segment and classify the comprehensive computing power score, and divide the computing power nodes into different performance levels. S6. Dynamically allocate computing resources based on task requirements and the performance level of computing nodes.

2. The computing power method based on multi-dimensional computing power perception according to claim 1, characterized in that, The computing power includes the number of CPU cores, single-core frequency, multi-core performance, and floating-point operation capability; the memory performance includes memory capacity, memory bandwidth, and memory read / write speed; the storage performance includes storage capacity and storage read / write speed; the communication capability includes network bandwidth, network latency, and data transmission rate; and the energy efficiency ratio includes computing power per unit power consumption, memory performance per unit power consumption, and storage performance per unit power consumption.

3. The computing power method based on multi-dimensional computing power perception according to claim 1, characterized in that, The load status includes CPU load rate, memory utilization rate, storage utilization rate, and network bandwidth utilization rate; the throughput includes the processing speed of computing tasks, data transmission throughput, and storage read / write throughput; the latency includes computing latency, network latency, and storage access latency; the resource contention includes CPU resource contention, memory resource contention, storage resource contention, and network resource contention; the failure rate includes hardware failure rate, software failure rate, and system crash rate; the temperature includes CPU temperature, memory temperature, and storage device temperature; and the power consumption includes CPU power consumption, memory power consumption, storage device power consumption, and network device power consumption.

4. The computing power method based on multi-dimensional computing power perception according to claim 1, characterized in that, The normalization process uses the following formula: Where P is the current indicator value, P min P is the minimum value of this indicator. max This represents the maximum value of the indicator.

5. The computing power method based on multi-dimensional computing power perception according to claim 1, characterized in that, The formula for calculating the comprehensive computing power score is as follows: S=α·p static +β·D dynamic Where S is the overall score, p static For static performance scoring, D dynamic For dynamic state scoring, α and β are weight coefficients, and α+β=1. In computationally intensive tasks, the static performance weight α is increased, and in tasks with high real-time requirements, the dynamic state weight β is increased.

6. The computing power method based on multi-dimensional computing power perception according to claim 1, characterized in that, Step S5 further includes: S501. Collect comprehensive scoring data from a large number of computing power nodes to construct a dataset; S502. Use the CART algorithm to segment the comprehensive score data, and determine the optimal segmentation point by minimizing the variance within nodes and maximizing the difference between nodes; S503 divides computing nodes into high-performance, medium-performance, and low-performance levels.

7. The computing power method based on multi-dimensional computing power perception according to claim 1, characterized in that, Step S6 further includes: For computationally intensive tasks, priority should be given to allocating them to high-performance computing nodes; For tasks with high real-time requirements, priority should be given to computing nodes with high dynamic status scores; For light-load or test tasks, allocate them to computing nodes with lower performance levels; When a computing node experiences resource overload or failure, tasks are redistributed to other available nodes based on real-time dynamic status data.

8. A computing power system based on multi-dimensional computing power perception, characterized in that, The system for the image recognition method according to any one of claims 1-7 comprises: The static performance acquisition module is used to perform step S1 to acquire static performance data of the computing nodes; The dynamic status monitoring module is used to collect dynamic status data of computing nodes in real time during step S2. The data processing module is used to perform step S3 to normalize the static performance data and dynamic status data; The comprehensive scoring module is used to perform step S4 based on normalized static performance data and dynamic status data; The grading module is used to execute step S5, which segments and grades the computing power nodes using the CART algorithm. The resource scheduling module is used to execute step S6, which dynamically allocates computing resources based on task requirements and node hierarchical results.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the computing power method based on multi-dimensional computing power perception as described in any one of claims 1 to 7.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the image recognition method as described in any one of claims 1 to 7.

Images