A method for benchmark energy efficiency tuning of HPC clusters based on joint hardware and software optimization

Abstract

This invention discloses a benchmark energy efficiency tuning method for HPC clusters based on joint software and hardware optimization, belonging to the field of cluster energy efficiency tuning. The method consists of two aspects: software tuning and hardware tuning. On the software side, after obtaining energy efficiency-sensitive parameters under benchmark operation through parameter importance analysis, this invention combines an energy efficiency prediction model to optimize the sampling algorithm in the parameter optimization process, reducing optimization and model training overhead and improving model convergence speed, achieving offline, low-cost parameter optimization. On the hardware side, based on the optimal operating parameters of the cluster, this invention extracts features from the utilization rate of CPU core groups during benchmark operation, trains a clustering model to identify CPU load patterns, and can judge the resource requirements of the load from a global perspective, thereby performing corresponding dynamic frequency adjustment to improve the energy efficiency performance of the cluster.

Description

Technical Field

[0001] This invention relates to the field of benchmark energy efficiency tuning of HPC clusters in benchmark testing scenarios, and particularly to a method for benchmark energy efficiency tuning of HPC clusters based on joint hardware and software optimization. Background Technology

[0002] With the rapid development of computer technology, high-performance computing (HPC) has become an indispensable tool in many scientific, engineering, and commercial fields. However, with the continuous increase in computing tasks and the expansion of data scale, the energy consumption problem of HPC clusters has gradually become prominent. How to improve performance while keeping energy consumption within an acceptable range is also one of the main challenges facing the future development of exascale supercomputers. This has led to an urgent need for energy efficiency tuning of HPC clusters. Among the many HPC applications, HPC benchmarks are usually typical representatives of high-performance applications, representing typical HPC workloads (such as matrix computation, FFT, molecular dynamics simulations, etc.). Optimizing them can provide valuable experience and reference for energy efficiency optimization in practical applications.

[0003] In terms of software tuning, there have been many studies that have optimized specific benchmarks such as HPL by optimizing calculation methods and task load partitioning. There have also been many studies that have optimized and adapted the underlying operator libraries of HPC such as BLAS. Although these studies have achieved many good optimization results, their optimization objectives have certain specificity and may not be universal for other benchmarks. HPC applications rely on highly configurable software environments, making finding their optimal parameterization a complex task. The size of their parameter spaces and the nonlinear behavior of HPC systems often make manual tuning, theoretical modeling, or exhaustive sampling unsuitable. Black-box optimization-based automatic tuning methods eliminate the complexities of manual or theoretical model tuning, finding the optimal combination of algorithm, application, and hardware parameters. Several studies have applied black-box parameter optimization methods to optimize application-related parameters in the HPC field. Robert S et al. proposed an online auto-tuning tool that relies on black-box optimization to find the optimal parameters for an I / O accelerator for a given HPC application within a finite number of iterations, without making any assumptions about the system's behavior. Mittal S et al. applied Bayesian optimization (BO) to jointly tune the HPL benchmark parameters and hardware parameters of multi-GPU clusters, and configured the supercomputing system Kukai based on this tuning method, achieving second place on the 2017 Green500 list. Lin... W et al. proposed a multi-benchmark-driven parameter tuning method for energy efficiency optimization of servers with mixed loads in data centers. This method collects runtime data from multiple benchmarks, establishes a multi-label load classification model to map the actual mixed load to multiple benchmarks, and trains energy efficiency optimization models for different benchmarks to optimize the energy efficiency performance of the actual mixed load. A significant problem with these methods is that while they eliminate the inefficiency of manual optimization, they require multiple runs on real machines to obtain optimal parameters or train multiple different benchmark energy efficiency optimization models. This process can incur significant model training overhead. Therefore, it is necessary to optimize the optimization algorithm and reduce the search space to improve the convergence speed and reduce the model training overhead.

[0004] Regarding hardware optimization, in the optimization schemes for benchmark hardware voltage and frequency adjustment in the HPC field, JinpyoK et al. determined the optimal GPU operating frequency for HPL operation through multiple sets of experiments. They fixed the GPU at this frequency during HPL operation to improve energy efficiency and proposed that the optimal energy efficiency of the cluster's HPL is not achieved at the maximum performance of the CPU / GPU, therefore, appropriate optimization of the GPU / CPU is essential. Schoonhoven R et al. compared the effects of power capping and fixed frequency energy-saving strategies under Generalized Matrix Multiplication Kernel (GEMM) computation, pointing out that fixed frequency provides finer control over power consumption and supports a wider range of power consumption adjustments. They also established a GPU power consumption model to provide the most energy-efficient clock frequency for the GPU, significantly reducing the large adjustment search space. Like many current methods, these two approaches fix the optimal GPU frequency before benchmark operation, without globally considering dynamic load changes and formulating corresponding optimal frequency adjustment strategies, making them more suitable for tasks with relatively stable load levels during operation.

[0005] Specifically, current benchmark hardware and software energy efficiency optimization for HPC clusters typically relies on software optimizations tailored to specific benchmarks (such as HPL) or operator libraries (such as BLAS), lacking universality and exhibiting complex parameter spaces. While black-box optimization methods avoid the complexity of manual tuning and theoretical modeling, they require extensive real-world testing to collect data, resulting in high overhead and slow convergence. On the hardware side, existing methods often employ fixed-frequency or power capping strategies, failing to adequately consider dynamic load changes and hindering real-time hardware parameter adjustments to adapt to different task requirements. Furthermore, they often lack a global perspective, hindering effective optimization of the entire system's energy efficiency. In addition, hardware-software co-optimization is insufficient; existing optimization methods suffer from poor scalability in large-scale systems, and the lack of standardization among optimization tools across different benchmarks limits their widespread applicability and effectiveness in practical applications. Summary of the Invention

[0006] The main objective of this invention is to overcome the shortcomings of existing HPC benchmark energy efficiency optimization techniques, such as insufficient versatility, high model training and data acquisition overhead, and lack of a global optimization perspective. This invention provides a hardware-software co-optimization method for HPC cluster benchmark energy efficiency tuning. On the software side, this invention selects benchmark energy efficiency-sensitive parameters, reducing the parameter search space. This method can adapt to different benchmark input parameters and system kernel parameters. Furthermore, by combining the sampling algorithm in the parameter optimization process with an energy efficiency prediction model, it reduces optimization and model training overhead, achieving low-cost offline parameter optimization. On the hardware side, based on the optimal parameters provided by the software, this invention extracts features from the CPU core utilization rate during benchmark operation. The trained clustering model can guide dynamic frequency adjustment of the benchmark from a global perspective. The combined hardware and software optimization measures effectively improve the overall energy efficiency level of benchmark operation under the cluster.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a baseline energy efficiency tuning method for HPC clusters based on joint software and hardware optimization, comprising the following steps: including software optimization and hardware optimization components, specifically:

[0009] S1. Software Optimization:

[0010] S11. Collect multiple sets of baseline operating energy efficiency values ​​under different optimization parameters using actual equipment, and filter out energy efficiency sensitive parameters in the baseline operating energy efficiency values ​​through parameter importance analysis;

[0011] S12. Use the Bayesian parameter optimization method to perform a real-machine iterative search on the selected energy efficiency sensitive parameters, and record the energy efficiency value under each parameter combination.

[0012] S13. Use a data-driven ensemble learning method to build a parameter-based energy efficiency prediction model, and compare the actual machine-collected values ​​with the model accuracy index to evaluate the model accuracy.

[0013] S14. Combine the energy efficiency prediction model to select and optimize the sampling algorithm of the Bayesian parameter optimization process, and then optimize the actual Bayesian parameter process.

[0014] S15. The optimized Bayesian parameter optimization model is used to guide the tuning of baseline energy efficiency sensitive parameters under the cluster, and enters the offline tuning stage without the need for physical machine operation.

[0015] S2, Hardware Optimization:

[0016] S21. Run and collect the CPU core utilization changes of different nodes in the cluster at fixed time intervals, and extract the features of the cores in the group according to the CPU cores with shared voltage and frequency.

[0017] S22. Draw a scatter plot of the CPU core utilization within the same group during the baseline operation, and determine the main load mode based on the utilization distribution of the CPU core group.

[0018] S23. Use all extracted CPU core group features as multi-dimensional features and use a clustering algorithm to obtain the load patterns that exist during the benchmark operation.

[0019] S24. Combining the load patterns obtained from analysis and clustering, a filtering layer is introduced to optimize the performance of the clustering model, and the clustering model provides the load patterns of different CPU core groups.

[0020] S25. Select a load mode with energy-saving potential and formulate frequency regulation actions to guide feedback energy saving.

[0021] As a preferred technical solution, in step S11:

[0022] The discrete or continuous parameter combination for optimization is generated by random parameter search. The energy efficiency value of the benchmark under the parameter combination is collected on the actual machine. The relationship between parameters and energy efficiency is learned by random forest model and the feature importance index of the influence of each parameter on energy efficiency is output. Energy efficiency sensitive parameters are selected to narrow down the parameter search space. The energy efficiency sensitive parameters are those with high importance.

[0023] As a preferred technical solution, in step S12...

[0024] A new parameter space is searched using a Bayesian parameter tuning and optimization method based on Gaussian processes. During the process, pruning methods are used to accelerate the convergence of Bayesian parameter optimization. The baseline energy efficiency value and the historical best energy efficiency value are recorded as the number of search iterations increases. When the historical best energy efficiency value no longer increases or tends to stabilize, the parameter optimization is considered to have converged.

[0025] As a preferred technical solution, in step S13...

[0026] When establishing an energy efficiency prediction model, multiple sets of results obtained from the Bayesian parameter optimization process are used to build a data-driven energy efficiency prediction model. The effectiveness of various ensemble learners is evaluated using model accuracy evaluation metrics and the accuracy differences between the prediction parameter combinations and the actual model. The model with the best overall performance is selected as the energy efficiency prediction model. The model accuracy evaluation metrics include MSE, MAE, and R. 2 .

[0027] As a preferred technical solution, in step S14...

[0028] After establishing the energy efficiency prediction model, the energy efficiency value output by the energy efficiency prediction model based on the given parameter combination is used as the objective function of Bayesian parameter optimization to perform low-overhead parameter optimization based on the prediction model. Under this condition, the convergence speed and convergence accuracy of the sampling algorithms used in the Bayesian parameter optimization process are compared, and different sampling algorithms suitable for the offline stage and the actual data acquisition stage are selected and fed back to the actual parameter optimization process.

[0029] As a preferred technical solution, in step S21...

[0030] The data acquisition method involves grouping CPU cores based on their shared voltage frequency during frequency tuning. Data is collected on the utilization rates of cores within the same group that share the voltage frequency during benchmark operation. Data is processed for each group, with data acquired every time interval T. The CPU core utilization rates of each group are then used to extract a set of feature values ​​χ based on the mean, range, standard deviation, maximum value, and median. i ={U mean U range U std U max U med}

[0031] As a preferred technical solution, in step S22,

[0032] Based on the data collection time interval T and the core grouping, a scatter plot of the core utilization distribution within each CPU core group during benchmark operation was plotted to identify the main load modes. Frequency adjustment feedback was obtained through experiments to confirm the adjustable and non-adjustable modes. The key to distinguishing between adjustable and non-adjustable modes lies in adjusting the core group under this mode using frequency reduction techniques to improve benchmark energy efficiency while keeping the performance degradation within a set threshold dg. max the following.

[0033] As a preferred technical solution, in step S23...

[0034] By taking multidimensional features as input, the combined effects of various feature values ​​are compared. Clustering performance metrics and WCSS are used to select the feature dimension input with the best clustering effect, and the optimal number of clusters is found based on the elbow principle of the clustering algorithm.

[0035] As a preferred technical solution, in step S24...

[0036] By combining adjustable modes, a filtering layer is added before the output of the clustering model to enhance the discriminative power of the adjustable modes, so that when given an input feature, the frequency modulation action required for the output point can be quickly determined according to the clustering model.

[0037] As a preferred technical solution, in step S24...

[0038] CPU load modes can be categorized as follows, with the latter two being adjustable modes:

[0039] I. The utilization rate of most cores within the group is higher than the set threshold U. min This state can be maintained at the highest frequency.

[0040] II. Most CPUs in the group are under low utilization, but a few are under particularly high utilization. In this case, frequency reduction has a significant impact on performance.

[0041] III. Most CPUs are in a low utilization / idle state, while a small number of CPUs are in a medium to high utilization state. In this state, the frequency needs to be reduced. The frequency reduction level can be determined according to the utilization level of the small number of CPUs.

[0042] IV. The utilization rate of the combined cores is low, and the frequency is reduced directly to the lowest level.

[0043] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0044] (1) This invention proposes an input parameter and system kernel parameter optimization method based on Bayesian optimization. It is not designed for a specific benchmark and can be applied to automatic parameter optimization during the operation of various HPC benchmarks. It can provide optimized parameter combinations for benchmark energy efficiency sensitive parameters without human intervention, thereby improving the efficiency of parameter optimization.

[0045] (2) This invention proposes a strategy of combining data-driven training of energy efficiency prediction model in the parameter optimization process. This strategy outputs the benchmark energy efficiency under a specific parameter combination by combining the energy efficiency prediction model established by combining historical sample values. It can be used for offline parameter optimization and sampling algorithm in the optimization process of clusters of the same size, thereby helping to alleviate the power and time cost problem of black box parameter optimization method needing to be performed on large-scale HPC clusters for actual machine optimization, and improving the optimization effect of the optimization algorithm.

[0046] (2) Based on the dynamic characteristics of the benchmark operation, this invention proposes a dynamic frequency adjustment energy efficiency optimization strategy according to the changes in the CPU core group utilization characteristics. By identifying the adjustable load mode of the CPU voltage and frequency group and combining it with the clustering model to guide the dynamic frequency adjustment action, the benchmark operation energy efficiency can be effectively improved from a global perspective. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of a benchmark energy efficiency tuning method for HPC clusters based on joint hardware and software optimization.

[0049] Figure 2 This is a schematic diagram of the Bayesian parameter optimization process during actual data acquisition.

[0050] Figure 3 This diagram illustrates the optimization process of the Bayesian parameter optimization method combined with the energy efficiency prediction model.

[0051] Figure 4 This is an example of a scatter plot showing the distribution of core utilization within a CPU core group over time during benchmark operation. Detailed Implementation

[0052] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0053] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0054] The server configuration in this embodiment includes hardware information configuration, operating system and dependency library configuration, as shown in Table 1.

[0055] Table 1 Server Configuration of Examples

[0056]

[0057] like Figure 1As shown, this embodiment presents a benchmark energy efficiency tuning method for HPC clusters based on joint software and hardware optimization. This method comprises two aspects: software tuning and hardware tuning. On the software side, after obtaining energy efficiency-sensitive parameters under benchmark operation through parameter importance analysis, the method combines an energy efficiency prediction model to optimize the sampling algorithm during parameter optimization, reducing optimization and model training overhead and improving model convergence speed, thus achieving low-cost offline parameter optimization. On the hardware side, based on the optimal operating parameters of the cluster, the method extracts features from the utilization rate of CPU core groups during benchmark operation, trains a clustering model to identify CPU load patterns, and can judge the resource requirements of the load from a global perspective, thereby performing corresponding dynamic frequency adjustments to improve the energy efficiency performance of the cluster. The specific details of this method are as follows:

[0058] S1. Software Optimization:

[0059] First, for the benchmark requiring optimization, its adjustable parameters are determined, including benchmark-specific input parameters and general system kernel parameters. A random search algorithm is used to generate combinations of adjustable parameters collected on a real machine. After obtaining the operating energy efficiency under multiple sets of different adjustable parameter combinations, a random forest algorithm is used to output the importance analysis of the impact of each adjustable parameter on energy efficiency. From this, energy efficiency-sensitive parameters are selected as the parameter space for subsequent optimization. This step narrows the search space of the parameter optimization algorithm, thereby improving the convergence speed of subsequent optimization. In the example, the energy efficiency-sensitive parameters obtained using this algorithm under the HPL-AI (HPL-derived benchmark adapted to mixed precision scenarios) benchmark are shown in Table 2.

[0060] Table 2. Energy efficiency sensitive parameters obtained after parameter importance analysis.

[0061]

[0062] Subsequently, a Bayesian parameter optimization algorithm is used to optimize the energy efficiency-sensitive parameters. The Bayesian parameter optimization process requires defining the parameter search space, step size, sampling algorithm, and objective function. The parameter search space is given by the combination of energy efficiency-sensitive parameters determined in the previous step. The sampling algorithm can initially use the default sampling algorithm, and the objective function is... Figure 2 The energy efficiency data acquisition module shown collects data. During the optimization process, the recorded historical sample points are used as training samples for the energy efficiency prediction model. A data-driven ensemble learning algorithm is used to train the energy efficiency prediction model, and the model accuracy obtained from various different ensemble learners (XGBoost, LightGBM, CatBoost, RandomForest, GradientBoost) is compared. The model accuracy is evaluated based on the model accuracy metrics (MSE, MAE, R). 2The optimal ensemble learner for building an energy efficiency prediction model is selected by considering the differences in accuracy between the predicted parameter combination energy efficiency and the actual model.

[0063] As the Bayesian parameter optimization process progresses, the accuracy of the energy efficiency prediction model continuously improves. To minimize the number of actual samples, this method evaluates the impact of the number of training samples on the model's prediction accuracy. As shown in Table 3, this evaluation method reveals that when the number of samples is around 60, an R0 can be obtained. 2 The energy efficiency prediction model has a value of 0.9327 and a prediction accuracy of 93.4%. Moreover, as the number of optimization samples continues to increase, its prediction accuracy still has room for improvement. This indicates that the method used can achieve high prediction accuracy with a small number of samples, thereby alleviating the problem that traditional black-box parameter optimization requires a large number of real-world operations to collect data, resulting in high optimization costs.

[0064] Table 3. Performance of Energy Efficiency Prediction Model as the Recorded Sample Interval Increases

[0065] Sampling sample interval (0,20) (0,40) (0,60) (0,80) (0,100) <![CDATA[Prediction accuracy R 2 > -0.5302 0.8748 0.9327 0.9495 0.9512 Predicting optimal energy efficiency (Gflops / W) 1.344 1.639 1.8108 1.8584 1.8685 Error (%) compared to actual value 30.5 15.3 6.4 3.9 3.4

[0066] After obtaining a high-accuracy energy efficiency prediction model, it can be combined with a Bayesian parameter optimization process. At this point, offline Bayesian parameter optimization can be performed. The objective function in the Bayesian parameter optimization process is given by the energy efficiency prediction model. The Bayesian parameter optimization process combined with the energy efficiency prediction model is as follows: Figure 3 As shown, this model can also be used to quickly iterate and compare sampling algorithms for the Bayesian parameter optimization process with extremely low overhead, thereby selecting the sampling algorithm more suitable for the current optimization situation. The convergence results of different sampling algorithms (TPE, CMAES, GPEI, GPPI, GPLCB) obtained using this model are shown in Table 4. For these five different algorithms, the overhead of TPE and CMAES is much smaller than that of the latter three algorithms based on the GP model. The GP model algorithm improves faster in the early stages (especially the GPPI model) and can quickly converge to a better value, but its solution time has a cumulative effect as time increases. As the number of sampling rounds increases, the GP-based model needs to use more historical data for prediction. The sampling involves calculating the inverse of the covariance matrix, and the increase in data volume will lead to a significant increase in computational complexity. Its time complexity is usually O(n^2). 3Therefore, it is not suitable as a sampler for a large number of rounds. GPPI is suitable as a sampling algorithm during the data acquisition phase when real-world operation is required. However, once a highly accurate energy efficiency model is established, the less expensive TPE sampling algorithm should be chosen for multiple rounds of sampling. While TPE's initial improvement speed is slower than GPPI-based models, its convergence speed is second only to the GPPI sampler, and its time overhead is much less than that of GPPI-based sampling algorithms. After obtaining the comparison results of the sampling algorithms, the most suitable sampling algorithm can be selected to improve optimization efficiency during the subsequent real-world and offline operation phases of this benchmark.

[0067] Table 4 Comparison of convergence rounds and convergence time for different sampling algorithms

[0068] Sampling algorithm type CMA-ES TPE GPEI GPPI GPLCB Number of convergence rounds 111 73 77 55 75 Convergence time (s) 15.88 7.05 28.89 9.7 15.47

[0069] S2, Hardware Optimization:

[0070] After obtaining the optimal operating parameters for the baseline through software tuning, these optimal parameters are used as the baseline operating parameters for hardware tuning. During the hardware tuning phase, it is necessary to run and collect data on the CPU utilization changes of different nodes within the cluster at fixed time intervals. Data is collected once every time interval T. The CPU cores are grouped according to the shared voltage frequency during frequency tuning, and the CPU core utilization of each group is used to extract a set of utilization feature values ​​χ based on the average, range, standard deviation, maximum, and median. i ={U mean U range U std U max U med The feature value is used to train the k-means clustering model. All the collected feature points are used as inputs and clustering performance indicators (Silhouette Score and WCSS) are used to select the feature dimension input with the best clustering effect. The optimal number of clusters is found according to the elbow principle of the clustering algorithm.

[0071] In the load pattern identification step, the collected data is used to plot a scatter plot of the distribution of core utilization within the CPU core group over time during benchmark operation (e.g., ...). Figure 4 The main load modes were identified (shown by different colored dashed boxes), and frequency tuning feedback was obtained through experiments to confirm the adjustable and non-adjustable modes. In the adjustable mode, adjusting the core components under this mode via frequency reduction improved baseline energy efficiency while keeping performance degradation within a set threshold (dg). max The following describes the CPU load modes during the operation of the HPL-AI benchmark in the example, with the latter two being adjustable modes:

[0072] I. The utilization rate of most cores within the group is higher than the set threshold U. min This state can be maintained at the highest frequency.

[0073] II. Most CPUs in the group are in a low utilization state, but a small number (1-2) have particularly high utilization. In this state, frequency reduction has a significant impact on performance.

[0074] III. Most CPUs are in a low utilization / idle state, and a small number of CPUs are in a medium to high utilization (10%-70%). In this state, the frequency needs to be reduced. The frequency reduction level can be determined according to the utilization of the small number of CPUs.

[0075] IV. The utilization rate of the combined cores is low, so it can be directly reduced to the lowest frequency.

[0076] After obtaining the load adjustable mode, a feature filtering layer can be written to filter the CPU core utilization characteristics during the baseline operation based on the characteristics of the adjustable mode. Then, a clustering model is used to cluster and obtain the load classification. The frequency adjustment action is formulated by combining the load mode and the classification. The energy efficiency optimization results of the frequency adjustment strategy combined with the load mode grouping in the embodiment are shown in Table 5. The energy efficiency after combining software parameter optimization and dynamic frequency adjustment is 9.84% higher than that under the default energy saving mode of the server.

[0077] Table 5 shows the energy efficiency improvement achieved by combining the sub-mode dynamic frequency modulation method with parameter optimization methods.

[0078]

[0079] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0080] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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, they should be considered to be within the scope of this specification.

[0081] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for benchmark energy efficiency tuning of HPC clusters based on joint hardware and software optimization, characterized in that, The process includes the following steps: software optimization and hardware optimization, specifically: S1. Software Optimization: S11. Collect multiple sets of baseline operating energy efficiency values ​​under different optimization parameters using actual equipment, and filter out energy efficiency sensitive parameters in the baseline operating energy efficiency values ​​through parameter importance analysis; S12. Use the Bayesian parameter optimization method to perform a real-machine iterative search on the selected energy efficiency sensitive parameters, and record the energy efficiency value under each parameter combination. S13. Use a data-driven ensemble learning method to build a parameter-based energy efficiency prediction model, and compare the actual machine-collected values ​​with the model accuracy index to evaluate the model accuracy. S14. Combine the energy efficiency prediction model to select and optimize the sampling algorithm of the Bayesian parameter optimization process, and then optimize the actual Bayesian parameter process. S15. The optimized Bayesian parameter optimization model is used to guide the tuning of baseline energy efficiency sensitive parameters under the cluster, and enters the offline tuning stage without the need for physical machine operation. S2, Hardware Optimization: S21. Run and collect the CPU core utilization changes of different nodes in the cluster at fixed time intervals, and extract the features of the cores in the group according to the CPU cores with shared voltage and frequency. S22. Draw a scatter plot of the CPU core utilization within the same group during the baseline operation, and determine the load mode based on the utilization distribution of the CPU core group. S23. Use all extracted CPU core group features as multi-dimensional features and use a clustering algorithm to obtain the load patterns that exist during the benchmark operation. S24. Combining the load patterns obtained from analysis and clustering, a filtering layer is introduced to optimize the performance of the clustering model, and the clustering model provides the load patterns of different CPU core groups. S25. Select a load mode with energy-saving potential and formulate frequency regulation actions to guide feedback energy saving.

2. The HPC cluster baseline energy efficiency tuning method based on hardware and software joint optimization according to claim 1, characterized in that, In step S11: The discrete or continuous parameter combination for optimization is generated by random parameter search. The energy efficiency value of the benchmark under the parameter combination is collected on the actual machine. The relationship between parameters and energy efficiency is learned by random forest model and the feature importance index of the influence of each parameter on energy efficiency is output. Energy efficiency sensitive parameters are selected to narrow down the parameter search space. The energy efficiency sensitive parameters are those with high importance.

3. The HPC cluster baseline energy efficiency tuning method based on hardware and software joint optimization according to claim 1, characterized in that, In step S12, A new parameter space is searched using a Bayesian parameter tuning and optimization method based on Gaussian processes. During the process, pruning methods are used to accelerate the convergence of Bayesian parameter optimization. The baseline energy efficiency value and the historical best energy efficiency value are recorded as the number of search iterations increases. When the historical best energy efficiency value no longer increases or tends to stabilize, the parameter optimization is considered to have converged.

4. The HPC cluster baseline energy efficiency tuning method based on hardware and software joint optimization according to claim 1, characterized in that, In step S13, When establishing an energy efficiency prediction model, multiple sets of results obtained from the Bayesian parameter optimization process are used to build a data-driven energy efficiency prediction model. The effectiveness of various ensemble learners is evaluated using model accuracy evaluation metrics and the accuracy differences between the prediction parameter combinations and the actual model. The model with the best overall performance is selected as the energy efficiency prediction model. The model accuracy evaluation metrics include MSE, MAE, and R. 2 .

5. The HPC cluster baseline energy efficiency tuning method based on hardware and software joint optimization according to claim 1, characterized in that, In step S14, After establishing the energy efficiency prediction model, the energy efficiency value output by the energy efficiency prediction model based on the given parameter combination is used as the objective function of Bayesian parameter optimization to perform low-overhead parameter optimization based on the prediction model. Under this condition, the convergence speed and convergence accuracy of the sampling algorithms used in the Bayesian parameter optimization process are compared, and different sampling algorithms suitable for the offline stage and the actual data acquisition stage are selected and fed back to the actual parameter optimization process.

6. The HPC cluster baseline energy efficiency tuning method based on hardware and software joint optimization according to claim 1, characterized in that, In step S21, The data acquisition method involves grouping CPU cores that can share voltage frequencies during frequency tuning. The utilization rates of cores sharing voltage frequencies within the same group during benchmark operation are collected and processed as a single group. Data is acquired once every time interval T. The CPU core utilization rates of each group are then used to extract a set of feature values ​​χ based on the mean, range, standard deviation, maximum value, and median. i ={U mean U range U std U max U med } 7. The HPC cluster baseline energy efficiency tuning method based on hardware and software joint optimization according to claim 1, characterized in that, In step S22, Based on the data collection time interval T and the core grouping, a scatter plot of the core utilization distribution within each CPU core group during the benchmark operation was plotted to identify the load modes. Frequency adjustment feedback was obtained through experiments to confirm the adjustable and non-adjustable modes. The key to distinguishing between adjustable and non-adjustable modes lies in adjusting the core group under that load mode using frequency reduction methods to improve benchmark energy efficiency while keeping the performance degradation within a set threshold dg. max the following.

8. The HPC cluster baseline energy efficiency tuning method based on hardware and software joint optimization according to claim 1, characterized in that, In step S23, By taking multidimensional features as input, the combined effects of various feature values ​​are compared. Clustering performance metrics and WCSS are used to select the feature dimension input with the best clustering effect, and the optimal number of clusters is found based on the elbow principle of the clustering algorithm.

9. The HPC cluster baseline energy efficiency tuning method based on hardware and software joint optimization according to claim 1, characterized in that, In step S24, By combining adjustable modes, a filtering layer is added before the output of the clustering model to enhance the discriminative power of the adjustable modes, so that when given an input feature, the frequency tuning action required for the output point can be quickly determined according to the clustering model.

Images