Normalized measurement method for heterogeneous computing resource multi-dimensional index system

Abstract

The invention discloses a normalization measurement method for a heterogeneous computing resource multi-dimensional index system. The method comprises the following steps: constructing the multi-dimensional index system including performance, energy efficiency and utilization rate; constructing a performance model HCRN based on the multi-dimensional index system; collecting multi-dimensional index system data in real time through a hardware monitoring tool and a performance counter, and inputting the data into a performance model HCRN after data preprocessing; constructing a reference task library to evaluate the performance of the system performance under different field applications; dynamically adjusting the weight of the performance model HCRN by adopting a reinforcement learning algorithm to realize the self-adaption of the model to task load and hardware state change; and performing performance evaluation verification and iterative optimization on the adjusted model. According to the method, a scientific and quantitative decision basis is provided for resource scheduling, model selection optimization and service deployment, the utilization efficiency of heterogeneous computing power resources is improved, the overall performance of the system is improved, and it is ensured that the system can operate efficiently and stably in different environments.

Description

Technical Field

[0001] The present invention relates to the technical field of heterogeneous computing resource management, and particularly to a normalization measurement method for a multi-dimensional index system of heterogeneous computing resources. Background Art

[0002] With the rapid development of fields such as artificial intelligence, cloud computing, and high-performance computing, the application of heterogeneous computing resources (such as GPUs, TPUs, NPUs, CPU clusters, etc.) has become increasingly widespread. These resources exhibit significant differences in computing architectures, performance characteristics, and energy efficiency performance, which pose great challenges to resource scheduling and utilization. Currently, the performance evaluation of heterogeneous computing resources mainly relies on theoretical performance or single benchmark tests. However, this method has many limitations and cannot comprehensively reflect the performance of these resources in actual business scenarios. Especially in complex tasks and diverse applications, the evaluation of the actual efficiency of resources has become an important problem.

[0003] Specifically, the main challenges currently faced include the following four aspects: First, there are significant differences in the performance of different hardware in key computing tasks (such as floating-point operations and matrix operations). This difference makes it particularly difficult to compare cross-architecture performance, and it is difficult to accurately measure directly through theoretical performance or single benchmark tests.

[0004] Second, due to the diversity of workloads, the performance of hardware optimized for specific algorithms (such as deep learning, data analysis, etc.) is often not universal. This means that hardware that performs well in certain specific application scenarios may not have the same advantages in other application scenarios.

[0005] Third, the performance of computing resources may also be affected by dynamic factors such as temperature changes, power consumption limits, and energy consumption strategies. Changes in these factors will directly affect the performance of the hardware, making performance evaluation more complex and uncertain.

[0006] Finally, although traditional theoretical metrics (such as FLOPS, i.e., floating-point operations per second) are often used as performance references, it is difficult for them to comprehensively reflect the efficiency of heterogeneous computing resources in actual business scenarios. These theoretical metrics often ignore other important dimensions such as energy efficiency and resource utilization, and thus cannot provide a comprehensive and accurate performance evaluation. Summary of the Invention

[0007] The main purpose of this application is to provide a normalization measurement method for a multi-dimensional index system of heterogeneous computing resources, aiming to solve the limitations of existing methods for the performance evaluation of heterogeneous computing resources and realize a method that can comprehensively and accurately reflect the performance of these resources in actual business scenarios.

[0008] To achieve the above object, the present application provides the following solution: a normalization measurement method for a multi-dimensional index system of heterogeneous computing resources, including the following steps:

[0009] Construct a multi-dimensional index system including performance, energy efficiency, and utilization rate;

[0010] Based on the multi-dimensional index system, construct a performance model HCRN, and the performance model is represented by formula (1):

[0011] HCRN = αP + βE + γU (1)

[0012] Wherein, P is the performance completed per unit time, E is the energy efficiency ratio per unit of computing power, U is the resource utilization rate, and α, β, and γ are adjustable weights;

[0013] Collect multi-dimensional index system data in real time through a hardware monitoring tool and a performance counter, and input the data into the performance model HCRN after data preprocessing;

[0014] Construct a benchmark task library to evaluate the performance of the system under different domain applications;

[0015] Adopt a reinforcement learning algorithm to dynamically adjust the weights of the performance model HCRN to achieve the adaptability of the model to changes in task load and hardware status;

[0016] Perform performance evaluation verification and iterative optimization on the adjusted model.

[0017] Preferably, the performance indicators include execution time and throughput, wherein:

[0018] Execution time T exec Represents the time from the start to the completion of a task on a specific hardware platform:

[0019] T exec = T end - T start (2)

[0020] Wherein, T start Is the task start time, and T end Is the task end time;

[0021] Throughput TP represents the number of tasks or the amount of data completed per unit time:

[0022]

[0023] Wherein, D is the amount of data or the number of tasks processed by the task, and T exec Is the task execution time.

[0024] Preferably, the energy efficiency indicator is the energy consumption required per unit of computing power:

[0025]

[0026] Among them, FLPOS is the number of floating-point operations per second, and PC is the power consumption, that is, the power consumption of the system during operation.

[0027] Preferably, the utilization rate indicators include computing resource utilization rate, storage resource utilization rate, and communication bandwidth utilization rate, where:

[0028] Computing resource utilization rate U compute represents the actual usage ratio of the computing unit during task execution:

[0029]

[0030] where T busy is the busy time of the computing resource;

[0031] Storage resource utilization rate U storage represents the degree of storage resource used during task execution:

[0032]

[0033] where S used is the storage capacity used during task execution, and S total is the total storage capacity;

[0034] Communication bandwidth utilization rate U bandwidth represents the ratio of the bandwidth used for data transmission during task execution to the total bandwidth:

[0035]

[0036] where B used is the bandwidth consumed during task execution, and B total is the total bandwidth.

[0037] Preferably, the benchmark task library includes:

[0038] General benchmark tasks, used to evaluate the performance of heterogeneous computing resources in common general computing scenarios;

[0039] Specialized benchmark tasks, designed for specific fields or specific types of computing tasks to deeply evaluate the performance of heterogeneous computing resources in specific application scenarios;

[0040] Mixed workload test tasks, which simulate real complex concurrent scenarios by combining multiple different types of tasks to more comprehensively evaluate the performance of heterogeneous computing resources under actual workloads.

[0041] Preferably, the performance model HCRN weights are dynamically adjusted using a reinforcement learning algorithm, which specifically includes the following steps:

[0042] Define the state space S, which is a set of real-time parameters including performance, energy efficiency, and utilization;

[0043] Define the action space A, which is the possible value range of the adjustable weights α, β, and γ;

[0044] Define the reward function R(s,a), which is the reward obtained after taking action a in state s. The reward function is represented by formula (8):

[0045] R(s,a) = -k * HCRN (8)

[0046] where k is a constant;

[0047] When HCRN decreases, the reward increases, encouraging the algorithm to find a strategy that minimizes HCRN.

[0048] Preferably, the Q-Learning algorithm is used in the reinforcement learning process. This process first initializes a Q-table and sets the initial weight combinations α0, β0, and γ0. Then, the algorithm enters the iterative process and performs the following steps at each time step t:

[0049] Observe the current state s t , including performance, energy efficiency, and utilization;

[0050] Select an action a t , that is, select a new weight combination;

[0051] Calculate the new HCRN value according to the selected weight combination and obtain the reward R(s t ,a t ) = -HCRNR;

[0052] Update the Q value in the Q-table using the Q-Learning update formula:

[0053] Q(s t ,a t ) ← (1 - α)Q(s t ,a t ) + α[R(s t ,a t ) + γmax a Q(s t+1 ,a)(9)

[0054] where α is the learning rate, which determines the proportion of new and old information in the update; γ is the discount factor, which controls the influence degree of future rewards on the current decision;

[0055] Repeat the above steps, and the reinforcement learning algorithm gradually converges to the optimal policy, finding the best combination of weights α*, β*, and γ* that minimizes the metric value of the performance model.

[0056] Preferably, the selection of action a t is achieved through the ∈-greedy policy to balance exploration and exploitation.

[0057] Preferably, the performance evaluation and verification of the adjusted model specifically include:

[0058] Accuracy, using the mean squared error to measure the gap between the predicted value and the true value, and the specific calculation method is shown in formula (10):

[0059]

[0060] where y i is the true value, is the model predicted value, and n is the number of samples;

[0061] Stability, which is evaluated by calculating the standard deviation of the performance metrics over a period of time. For a series of performance metrics P = {p1, p2,..., p n}, its standard deviation is shown in formula (11):

[0062]

[0063] where is the average value of the performance metrics.

[0064] Through the above technical solutions, the beneficial effects of the present invention are as follows: The normalization measurement method for the multi-dimensional index system of heterogeneous computing resources of the present invention combines the three core dimensions of performance, energy efficiency, and utilization rate for evaluation, and introduces an adaptive weight control mechanism, which can flexibly respond to different application scenarios and task requirements and provide accurate computing power evaluation. At the same time, the dynamic performance modeling and real-time monitoring mechanism ensure the accuracy of the evaluation results under the conditions of hardware changes and task load changes. The diversified design of the benchmark task library meets the needs of general and industry-specific applications. Coupled with the continuous optimization and verification mechanism, this solution has strong adaptive capabilities, provides scientific support for resource scheduling and optimization decisions, and ensures its long-term effectiveness and efficiency in different environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention, and the embodiments in the drawings do not constitute any limitation to the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0066] Figure 1 A flowchart of a normalization measurement method for a multi-dimensional index system of heterogeneous computing resources provided by an embodiment of the present application;

[0067] Figure 2 A schematic structural diagram of a computer device provided by an embodiment of the present application.

[0068] The realization of the purpose of the present application, functional features and advantages will be further described with reference to the embodiments and the accompanying drawings. Detailed implementation manners

[0069] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0070] Regarding the foregoing and other technical contents, features and effects of the present invention, they will be clearly presented in the following detailed description of the embodiments in conjunction with the accompanying drawings. The structural contents mentioned in the following embodiments are all referenced to the accompanying drawings of the specification. Figure 1-2 In the following detailed description of the embodiments, they will be clearly presented. The structural contents mentioned in the following embodiments are all referenced to the accompanying drawings of the specification.

[0071] The following will describe the exemplary embodiments of the present invention with reference to the accompanying drawings.

[0072] Refer to Figure 1 , A normalization measurement method for a multi-dimensional index system of heterogeneous computing resources provided by an embodiment of the present application. This method can be executed by a computing device with computing functions, and the computing device can be a desktop computer, a laptop computer, etc.

[0073] Step S1: Construct a multi-dimensional index system including performance, energy efficiency and utilization rate;

[0074] Step S2: Based on the multi-dimensional index system, construct a performance model HCRN, and the performance model is represented by formula (1):

[0075] HCRN = αP + βE + γU (1)

[0076] Wherein, P is the performance completed per unit time, E is the energy efficiency ratio per unit computing power, U is the resource utilization rate, and α, β, and γ are adjustable weights;

[0077] Step S3: Collect data of the multi-dimensional index system in real time through hardware monitoring tools and performance counters, and input it into the performance model HCRN after data preprocessing;

[0078] Step S4: Build a benchmark task library to evaluate the performance of the system under different domain applications;

[0079] Step S5: Dynamically adjust the weights of the performance model HCRN using the reinforcement learning algorithm to enable the model to adapt to changes in task load and hardware status;

[0080] Step S6: Conduct performance evaluation verification and iterative optimization on the adjusted model.

[0081] Performance metrics mainly focus on the execution time and throughput of tasks on a given hardware platform. By accurately collecting this data, it can be determined whether resources are effectively utilized. In Step S1, the performance metrics include execution time and throughput, where:

[0082] Execution time T exec represents the time from the start to the completion of a task on a specific hardware platform:

[0083] T exec = T end - T start (2)

[0084] where T start is the task start time, and T end is the task end time;

[0085] Throughput TP represents the number of tasks or the amount of data completed per unit time:

[0086]

[0087] where D is the amount of data processed by the task or the number of tasks, and T exec is the task execution time.

[0088] Energy efficiency refers to the energy consumed by a system or device under a certain computing power (computing ability), and specifically measures the energy consumption required per unit of computing power, as shown in formula (4):

[0089]

[0090] where FLPOS is the number of floating-point operations per second, and PC is the power consumption, that is, the power consumption of the system during operation.

[0091] Utilization metrics include computing resource utilization, storage resource utilization, and communication bandwidth utilization, where:

[0092] Computing resource utilization U compute Indicates the actual usage ratio of the computing unit during task execution:

[0093]

[0094] Where T busy Is the busy time of the computing resource;

[0095] Storage resource utilization U storage Indicates the degree of storage resource usage during task execution:

[0096]

[0097] Where S used Is the storage capacity used during task execution, and S total Is the total storage capacity;

[0098] Communication bandwidth utilization U bandwidth Indicates the ratio of the bandwidth used for data transmission during task execution to the total bandwidth:

[0099]

[0100] Where B used Is the bandwidth consumed during task execution, and B total Is the total bandwidth.

[0101] In step S2, based on the multi-dimensional index system, a unified performance model HCRN (Heterogeneous Computing Resource Normalization) is constructed, combining specific indicators of performance, energy efficiency, and resource utilization. The specific model is as follows:

[0102] HCRN = α(f(T exec , TP)) + βE + γ(U compute + U storage + U bandwidth )

[0103] In step S3, a variety of hardware monitoring tools and performance counters are used to collect multi-dimensional data such as power consumption and resource utilization in real time. For example, a power sensor is used to monitor the hardware power consumption, a performance counter records the execution status of the computing resource, a bandwidth monitor obtains the usage rates of memory and network bandwidth, and the hardware status is detected through sensors such as temperature and frequency. At the same time, the performance monitoring software perf can provide information such as the usage of the CPU and memory, task execution time, and throughput. The specific data collection is shown in the following table.

[0104] Table 1 Data Collection Information Table

[0105]

[0106]

[0107] The collected data is preprocessed through cleaning, transformation, and feature engineering, removing outliers, filling in missing values, removing duplicates, performing unit standardization, timestamp formatting, normalizing or standardizing features with different dimensions, using smoothing and noise reduction methods to reduce noise in the data. Finally, the data from different resources is aligned according to timestamps to ensure that the data input into the model is accurate, consistent, and reliable, providing support for accurate dynamic performance evaluation and optimization.

[0108] To ensure the wide applicability of the evaluation method, in step S4, the benchmark task library designs a diverse task set, specifically including general benchmark tasks, dedicated benchmark tasks, and mixed workload tests, aiming to comprehensively evaluate the performance of the system under different domain applications. The following is a detailed description of each type of task:

[0109] General benchmark tasks are used to evaluate the performance of heterogeneous computing resources in common general computing scenarios; this task focuses on key metrics such as basic operation performance, response speed, throughput, and latency, such as tasks like data processing, file input / output, and network communication, which helps evaluate the basic operation ability of the system and provides a reference for performance benchmarks in different application fields.

[0110] Dedicated benchmark tasks are designed for specific domains or specific types of computing tasks to deeply evaluate the performance of heterogeneous computing resources in specific application scenarios; for example, in fields such as image processing, machine learning model training, and scientific computing, it can more accurately measure performance and help analyze the efficiency and bottlenecks of the system under specific loads.

[0111] Mixed workload test tasks simulate real complex concurrent scenarios by combining multiple different types of tasks to more comprehensively evaluate the performance of heterogeneous computing resources under actual workloads. Mixed workload tests can not only test the stability and response ability of the system under complex loads but also explore the adaptive ability of the system under different loads by dynamically adjusting the load ratio.

[0112] Table 2 Task Library Construction List

[0113]

[0114]

[0115] Due to the dynamic adjustment of task load, hardware platform, and resource utilization over time and in different environments, traditional static performance models struggle to maintain high efficiency and accuracy when faced with changing systems. Therefore, dynamic performance models need to be able to adaptively update to respond to these changes in real time. To achieve this goal, in step S5, an online learning method is adopted to continuously adjust the model parameters using real-time data, thereby improving the accuracy and adaptability of the model.

[0116] Reinforcement Learning (RL) is a machine learning method that can interact with the environment and adjust strategies based on feedback. It can continuously optimize task scheduling and resource allocation according to real-time task execution results and resource status, thus achieving the adaptive update of the performance model. The specific steps for dynamically adjusting the weights of the performance model HCRN using the reinforcement learning algorithm are as follows:

[0117] Define the state space S, which is a set of real-time parameters including performance, energy efficiency, and utilization rate;

[0118] Define the action space A, which is the possible value range of the adjustable weights α, β, and γ;

[0119] Define the reward function R(s,a), which is the reward obtained after taking action a in state s. The reward function is represented by formula (8):

[0120] R(s,a) = -k * HCRN (8)

[0121] Where k is a constant to ensure that the reward value appropriately reflects the change in HCRN;

[0122] When HCRN decreases, the reward increases, encouraging the algorithm to find a strategy that minimizes HCRN.

[0123] In the reinforcement learning process, the Q-Learning algorithm is adopted. This process first initializes a Q-table and sets the initial weight combinations α0, β0, and γ0. Then, the algorithm enters an iterative process and performs the following steps at each time step t:

[0124] Observe the current state s t , including performance, energy efficiency, and utilization rate;

[0125] Select action a t , that is, select a new weight combination;

[0126] Calculate the new HCRN value according to the selected weight combination, and obtain the reward R(s t ,a t ) = -HCRNR;

[0127] Update the Q-values in the Q-table using the update formula of Q-Learning:

[0128] Q(s t ,a t )←(1-α)Q(s t ,a t )+α[R(s t ,a t )+γmax a Q(s t+1 ,a)(9)

[0129] where α is the learning rate, determining the proportion of new and old information in the update; γ is the discount factor, controlling the impact of future rewards on the current decision-making;

[0130] Repeat the above steps, and the reinforcement learning algorithm gradually converges to the optimal policy, finding the best weight combination α*, β*, and γ* that minimizes the performance model metric value. During this process, as the number of iterations increases, the system can automatically adjust the weights to adapt to different task requirements and environmental changes, thereby achieving dynamic performance optimization. Finally, after multiple iterations, the algorithm will output a set of final weights, which are used to guide resource allocation and performance tuning in practical applications.

[0131] In the above, the selection of action a t is achieved through the ε-greedy policy, which specifically includes the following steps:

[0132] Set a small probability value ε (0 < ε < 1), representing the probability of choosing a random action;

[0133] At each time step t, generate a random number rand, whose value ranges from 0 to 1;

[0134] If rand is less than ε, then adopt the exploration strategy, that is, randomly select an action as the current action a t , to try different weight combinations and explore possible new states;

[0135] If rand is greater than or equal to ε, then adopt the exploitation strategy, that is, according to the current Q-table, select the action corresponding to the highest Q-value as the current action a t , to utilize the known information and select the currently considered optimal weight combination;

[0136] Execute the selected action a t , and update the system state according to this action, entering the iterative process of the next time step.

[0137] Through the ε-greedy strategy, the algorithm can achieve a balance between exploration and exploitation. It can not only try new weight combinations to discover potentially better strategies but also utilize known information to select the currently considered optimal actions, thereby gradually optimizing the normalization measurement method for the multi-dimensional index system of heterogeneous computing resources.

[0138] In the process of performance evaluation and iterative optimization, first, based on the results of the preliminarily trained model, a comprehensive evaluation is conducted from aspects such as accuracy, stability, and efficiency, and the test set and benchmark task library are used to quantify the model performance. In step S6, the model uses accuracy and stability as the main evaluation indicators:

[0139] For accuracy, the mean squared error (MSE) is used to measure the gap between the predicted value and the true value, and the specific calculation method is shown in formula (10):

[0140]

[0141] where, y i is the true value, is the model predicted value, and n is the number of samples;

[0142] For stability, it is evaluated by calculating the standard deviation of the performance metrics over a period of time. For a series of performance metrics P = {p1, p2,..., p n}, its standard deviation is shown in formula (11):

[0143]

[0144] where, is the average value of the performance metrics.

[0145] During the evaluation process, once it is found that the model performs poorly under specific conditions, it will return to the data collection and preprocessing stage to further obtain relevant data and fine-tune or retrain the model. At the same time, the model structure, hyperparameter settings, and feature engineering strategies are adjusted to optimize the performance. Through this iterative and continuously improving process, it is ensured that the model has good generalization ability and high resource utilization rate, and remains robust and efficient in practical applications.

[0146] In summary, the normalization measurement method of the heterogeneous computing resource multi-dimensional index system of the present invention evaluates by combining the three core dimensions of performance, energy efficiency, and utilization rate, comprehensively covering the key aspects of resource evaluation. Compared with the traditional single-index evaluation, it can more accurately reflect the comprehensive performance in actual applications. The introduced adaptive weight regulation mechanism can flexibly adjust the weights of each dimension index according to different scenarios and task requirements, provide accurate computing power evaluation, and meet diverse usage needs. The dynamic performance modeling and real-time monitoring mechanism can track the changes of hardware and task loads in real time, adjust the evaluation results in a timely manner, and the online learning and weight adaptive update mechanism further enhances the system's adaptability. The diversified benchmark task library covers general and industry-specific applications, meets the performance evaluation needs of different fields and tasks, and improves the versatility and practicality. Finally, this method provides a scientific and quantitative decision-making basis for resource scheduling, selection optimization, and business deployment, helps improve the utilization efficiency of heterogeneous computing resources, enhances the overall performance of the system, and ensures that the system can operate efficiently and stably in different environments.

[0147] In an exemplary embodiment, a computer device is provided. The computer device can be a server or a terminal, and its internal structure diagram can be as Figure 2 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store heterogeneous computing resource multi-dimensional index system data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a normalization measurement method of a heterogeneous computing resource multi-dimensional index system.

[0148] Those skilled in the art can understand that Figure 2 the structure shown in

[0149] In an exemplary embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0150] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0151] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0152] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0153] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Among them, any reference to a memory, database, or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0154] In each of the embodiments provided in this application, the database involved may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., without limitation. In each of the embodiments provided in this application, the processor may be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without limitation.

[0155] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0156] Specific examples are used in this article to elaborate on the principles and implementation manners of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A normalized measurement method for a multi-dimensional index system of heterogeneous computing resources, characterized in that: The steps include: Construct a multi-dimensional indicator system including performance, energy efficiency and utilization; A performance model HCRN is constructed based on the multi-dimensional indicator system. The performance model is expressed by formula (1): HCRN=αP+βE+γU (1) Among them, P is the performance completed per unit time, E is the energy efficiency ratio per unit computing power, U is the resource utilization rate, and α, β, and γ are adjustable weights; The multi-dimensional indicator system data is collected in real time through hardware monitoring tools and performance counters, and input into the performance model HCRN after data preprocessing; Build a benchmark task library to evaluate the performance of the system in different application fields; A reinforcement learning algorithm is used to dynamically adjust the weight of the performance model HCRN to achieve the model's self-adaptation to changes in task load and hardware status; The adjusted model is subjected to performance evaluation, verification and iterative optimization.

2. The normalized measurement method for the multidimensional index system of heterogeneous computing resources according to claim 1 is characterized in that: The performance indicators include execution time and throughput, where: Execution time T exec Indicates the time from task start to completion on a specific hardware platform: T exec =T end -T start (2) Where, T start is the task start time, T end The task end time; Throughput TP represents the number of tasks or data completed per unit time: Where D is the amount of data or number of tasks to be processed, T exec The task execution time.

3. The normalized measurement method for the multidimensional index system of heterogeneous computing resources according to claim 2 is characterized in that: The energy efficiency index is the energy consumption per unit computing power: Among them, FLPOS is the number of floating-point operations per second, and PC is the power consumption, that is, the power consumption of the system when it is running.

4. The normalized measurement method for the multi-dimensional index system of heterogeneous computing resources according to claim 3 is characterized in that: The utilization index includes computing resource utilization, storage resource utilization and communication bandwidth utilization, where: Compute resource utilization U compute Indicates the actual usage ratio of computing units during task execution: Among them, T busy To calculate the busy time of resources; Storage resource utilization U storage Indicates the extent to which a task uses storage resources during execution: Among them, S used is the storage capacity used during task execution, S total is the total storage capacity; Communication bandwidth utilization U bandwidth Indicates the ratio of bandwidth used for data transmission to total bandwidth during task execution: Among them, B used is the bandwidth consumed during task execution, B total is the total bandwidth.

5. The normalized measurement method for the multi-dimensional index system of heterogeneous computing resources according to claim 1 is characterized in that: The benchmark task library includes: General benchmark tasks, used to evaluate the performance of heterogeneous computing resources in common general computing scenarios; Specialized benchmark tasks are designed for specific fields or types of computing tasks to deeply evaluate the performance of heterogeneous computing resources in specific application scenarios; Mixed load testing tasks simulate real complex concurrent scenarios by combining multiple different types of tasks to more comprehensively evaluate the performance of heterogeneous computing resources under actual workloads.

6. The normalized measurement method for the multi-dimensional index system of heterogeneous computing resources according to claim 5 is characterized in that: The method of dynamically adjusting the weight of the performance model HCRN by using a reinforcement learning algorithm specifically includes the following steps: Define the state space S, which is a set of real-time parameters including performance, energy efficiency and utilization; Define the action space A, which is the possible value range of the adjustable weights α, β, and γ; Define the reward function R(s,a), which is the reward obtained after taking action a in state s. The reward function is expressed by formula (8): R(s,a)=-k*HCRN (8) Where k is a constant; As HCRN decreases, the reward increases, encouraging the algorithm to find a policy that minimizes HCRN.

7. The normalized measurement method for the multi-dimensional index system of heterogeneous computing resources according to claim 6 is characterized in that: The reinforcement learning process adopts the Q-Learning algorithm, which first initializes a Q-table and sets the initial weight combination α0, β0 and γ0; then, the algorithm enters an iterative process and performs the following steps at each time step t: Observe the current state t , including performance, energy efficiency and utilization; Select action a t , that is, select a new weight combination; Calculate the new HCRN value based on the selected weight combination, and get the reward R(s) accordingly t ,a t )=-HCRNR; Update the Q value in the Q-table and use the Q-Learning update formula: Q(s t ,a t )←(1-α)Q(s t ,a t )+α[R(s t ,a t )+γmax a Q(s t+1 ,a) (9) Among them, α is the learning rate, which determines the proportion of new and old information in the update; γ is the discount factor, which controls the influence of future rewards on current decisions; By repeating the above steps, the reinforcement learning algorithm gradually converges to the optimal strategy and finds the best weight combination α*, β* and γ* that minimizes the performance model metric.

8. The normalized measurement method for the multi-dimensional index system of heterogeneous computing resources according to claim 7 is characterized in that: The selection action a t This is achieved through a ∈-greedy strategy to strike a balance between exploration and exploitation.

9. The normalized measurement method for the multi-dimensional index system of heterogeneous computing resources according to claim 1, characterized in that: The performance evaluation and verification of the adjusted model specifically includes: Accuracy uses the mean square error to measure the difference between the predicted value and the true value. The specific calculation method is shown in formula (10): Among them, y i is the true value, is the model prediction value, n is the number of samples; Stability is evaluated by calculating the standard deviation of the performance index over a period of time. For a series of performance indicators P = {p1, p2, ..., p n }, and its standard deviation is shown in formula (11): in, is the average value of the performance indicator.

Images