Hash chip-based adaptive energy consumption control method, system, device and medium

Abstract

The application relates to a computing power chip-based adaptive energy consumption control method, system, device and medium. The method comprises the following steps: constructing an optimal prediction regression model based on multi-channel hardware state indexes of a computing power chip at different frequencies, so as to predict performance data of the computing power chip at different frequencies; acquiring to-be-predicted data of the computing power chip under a current working load, predicting the to-be-predicted data based on the optimal prediction regression model, acquiring a target optimal frequency under a preset performance index condition, and the target optimal frequency corresponding to optimal energy consumption; and adaptively adjusting the working frequency of the computing power chip under the current working load based on the target optimal frequency. The application can accurately select the target optimal frequency that meets the performance requirement and has the lowest chip energy consumption, guarantees the chip performance, reduces the energy consumption, and solves the problems of low adjustment accuracy and large energy consumption of the computing power chip in frequency adjustment under different working load performances.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to an adaptive energy consumption control method and system based on a computing power chip, a device and a medium. BACKGROUND

[0002] With the continuous growth of deep learning, artificial intelligence and high-performance computing demand, the energy consumption problem of high-performance computing power chips is increasingly prominent when executing complex artificial intelligence workloads. The prior art has proposed a method that relies on white box information or needs transparent access to application code (such as neural network parameter counting and computing character properties) to determine the optimal frequency, and then performs DVFS (dynamic voltage and frequency scaling) modification. However, the applicability of this method is usually limited by limited data access permissions, which not only increases system overhead and affects system performance, but also has limited scope of application. In addition, a black box DVFS method is also proposed, which does not need to access or modify application-level code, and uses monitorable system-level data to adjust the frequency. However, this method only relies on monitorable system-level data to adjust the frequency, and has the problem of frequency adjustment lag, low adjustment accuracy, and inability to reduce the power consumption of the chip.

[0003] In summary, in the prior art, there are problems of low adjustment accuracy and high energy consumption in maintaining the balance between performance and energy consumption based on frequency adjustment for computing power chips in different workloads. SUMMARY

[0004] The present application provides an adaptive energy consumption control method and system based on a computing power chip to solve the problem of low adjustment accuracy and high energy consumption in maintaining the balance between performance and energy consumption based on frequency adjustment for computing power chips in different workloads in the prior art.

[0005] According to one aspect of an embodiment of the present application, the present application provides an adaptive energy consumption control method based on a computing power chip, which comprises: constructing an optimal prediction regression model based on multi-channel hardware state indicators of the computing power chip at different frequencies to predict performance data of the computing power chip at different frequencies; obtaining predicted data of the computing power chip under the current workload, predicting the predicted data based on the optimal prediction regression model, obtaining a target optimal frequency that meets a preset performance index condition, and the target optimal frequency corresponds to an optimal energy consumption; and adaptively adjusting the working frequency of the computing power chip under the current workload based on the target optimal frequency.

[0006] Optionally, before the step of constructing the optimal prediction regression model based on the multi-channel hardware state indicators of the computing power chip at different frequencies, the method further comprises: monitoring the hardware state of the computing power chip to obtain the multi-channel hardware state indicators of the computing power chip at different frequencies, wherein the multi-channel hardware state indicators comprise performance indicators and GPU state indicators of the computing power chip at different frequencies under different workloads; and aligning the performance indicators and the GPU state indicators of the computing power chip at different frequencies under each server based on timestamps to obtain a plurality of feature vectors to be trained.

[0007] Optionally, the performance indicators comprise memory utilization, and the GPU indicators comprise GPU utilization, and the step of constructing the optimal prediction regression model based on the multi-channel hardware state indicators of the computing power chip at different frequencies comprises: constructing a regression model to be trained by using a recurrent neural network; performing model training on the regression model to be trained based on the plurality of feature vectors to be trained obtained to predict a performance data sequence of a continuous frequency value, wherein the feature vectors to be trained comprise original values and relative change values of the memory utilization and the GPU utilization at different frequencies; and obtaining, based on the performance data sequence of the continuous frequency value, a frequency corresponding to which the performance data satisfies the preset performance indicator condition as an output frequency of model training to construct the optimal prediction regression model.

[0008] Optionally, after the step of obtaining, based on the performance data sequence of the continuous frequency value, a frequency corresponding to which the performance data satisfies the preset performance indicator condition as an output frequency of model training to construct the optimal prediction regression model, the method further comprises: constructing an optimization objective of a loss function, wherein the optimization objective is to minimize a difference between a predicted frequency and a minimum actual frequency satisfying the preset performance indicator condition; and optimizing, based on the optimization objective, the output frequency of model training by using the loss function.

[0009] Optionally, the step of obtaining, based on the optimal prediction regression model, the target optimal frequency satisfying the preset performance indicator condition under the current workload comprises: obtaining the to-be-predicted data of the computing power chip under the current workload, wherein the to-be-predicted data comprises original values and relative change values of the memory utilization and the GPU utilization of the computing power chip at different frequencies under the current workload; and taking the original values and the relative change values of the memory utilization and the GPU utilization of the computing power chip at different frequencies under the current workload as input parameters to perform prediction by using the optimal prediction regression model to output the target optimal frequency corresponding to which the performance data of the computing power chip under the current workload satisfies the preset performance indicator condition.

[0010] Optionally, after the obtaining the to-be-predicted data of the computing power chip under the current workload, the method further comprises: based on a preset trigger mechanism, comparing the memory utilization and / or the GPU utilization of the computing power chip history with the memory utilization and / or the GPU utilization of the computing power chip under the current workload, judging whether the utilization change value of any one exceeds a preset utilization change threshold; if at least one utilization change value exceeds the preset utilization change threshold, triggering frequency updating.

[0011] Optionally, before the obtaining the to-be-predicted data of the computing power chip under the current workload, the method further comprises: obtaining temperature data and power consumption data of the computing power chip under different frequencies; and constructing an optimal prediction regression model based on the temperature data, the power consumption data of the computing power chip under different frequencies, and the multi-channel hardware state indicators, to predict the performance data of the computing power chip under different frequencies.

[0012] According to another aspect of the embodiments of the present application, the present application provides an adaptive energy consumption control system based on a computing power chip, which comprises: a model construction module, configured to construct an optimal prediction regression model based on multi-channel hardware state indicators of the computing power chip under different frequencies, to predict performance data of the computing power chip under different frequencies; a prediction module, configured to obtain to-be-predicted data of the computing power chip under a current workload, to predict the to-be-predicted data based on the optimal prediction regression model, and to obtain a target optimal frequency under a preset performance index condition, the target optimal frequency corresponding to optimal energy consumption; and an adjustment module, configured to adaptively adjust the working frequency of the computing power chip under the current workload based on the target optimal frequency.

[0013] According to another aspect of the embodiments of the present application, the present application provides a computer device, which comprises: a processor, a memory, and a network interface, the memory storing machine readable instructions executable by the processor, when the computer device is running, the processor and the memory communicate through the network interface, the processor executes the machine readable instructions to perform the steps of the adaptive energy consumption control method based on a computing power chip as described above.

[0014] According to another aspect of the embodiments of the present application, the present application provides a computer readable medium having non-volatile program code executable by a processor, the program code causing the processor to perform the steps of the adaptive energy consumption control method based on a computing power chip.

[0015] The above technical solutions provided by the embodiments of the present application have the following advantages compared with related art:

[0016] The application provides an adaptive energy consumption control method based on a computing power chip, which is suitable for energy consumption optimization of the computing power chip under different work load scenarios. The optimal prediction regression model is constructed by pre-analyzing the multi-channel hardware state indicators of the computing power chip under different frequencies when the computing power chip is under different work loads, and then the to-be-predicted data of the computing power chip under the current work load is predicted to determine the target optimal frequency. The prediction process can realize the prediction of the performance trend under different frequencies, so that the target optimal frequency meeting the performance requirement and having the lowest chip energy consumption can be accurately selected, so that the selection of the target optimal frequency can fully reflect the balance between the computing demand of the work load and the available computing resources of the computing power chip. The frequency of the computing power chip is adjusted based on the target optimal frequency, so that the chip performance is guaranteed while the energy consumption is reduced. And by analyzing the multi-channel hardware state indicators of the computing power chip under different frequencies under the work load, the dynamic adjustment of the working frequency of the chip with the work load is realized. This process does not need to access or modify the application program code, greatly improving the flexibility and applicability of deployment. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present application and, together with the specification, serve to explain the principles of the application.

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the application or the related art, the accompanying drawings needed to be used in the embodiments or related art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0019] Figure 1 A flowchart of an optional adaptive energy consumption control method based on a computing power chip according to an embodiment of the application is provided.

[0020] Figure 2 A flowchart of an optional step S102 according to an embodiment of the application is provided.

[0021] Figure 3 A whole architecture diagram of an optional adaptive energy consumption control system based on a computing power chip according to an embodiment of the application is provided.

[0022] Figure 4 A flowchart of an optional step S104 according to an embodiment of the application is provided.

[0023] Figure 5 A module schematic diagram of an optional adaptive energy consumption control system based on a computing power chip according to an embodiment of the application is provided.

[0024] Figure 6 An optional computer device structure schematic diagram is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0026] To solve the problems mentioned in the background art, according to an aspect of the embodiments of the present application, embodiments of an adaptive energy consumption control method based on a computing power chip are provided.

[0027] It should be noted that the adaptive energy consumption control method based on a computing power chip provided by the embodiments of the present application is generally executed by a server and / or a terminal device, and accordingly, the adaptive energy consumption control system based on a computing power chip is generally arranged in the server and / or the terminal device.

[0028] As shown in FIG. 1, a flowchart of an adaptive energy consumption control method based on a computing power chip is provided for the embodiments of the present application. Taking an example that the adaptive energy consumption control method based on a computing power chip is executed by a server, the adaptive energy consumption control method based on a computing power chip includes the following steps: Figure 1 Figure 1 Step S102, constructing an optimal prediction regression model based on multi-channel hardware state indicators of the computing power chip at different frequencies to predict performance data of the computing power chip at different frequencies.

[0029] In the present embodiment, the adaptive energy consumption control method based on a computing power chip is applicable to scenarios in which the computing power chip meets performance requirements and realizes energy consumption optimization based on frequency adjustment under different work load scenarios, including AI (Artificial Intelligence) server, high-performance computing, deep neural network, large language model training, and inference scenarios to optimize the energy efficiency of the computing power chip while keeping the computing performance within an acceptable range.

[0030] In the present embodiment, the adaptive energy consumption control method based on a computing power chip is applicable to scenarios in which the computing power chip meets performance requirements and realizes energy consumption optimization based on frequency adjustment under different work load scenarios, including AI (Artificial Intelligence) server, high-performance computing, deep neural network, large language model training, and inference scenarios to optimize the energy efficiency of the computing power chip while keeping the computing performance within an acceptable range.

[0031] ​In the embodiment, the multi-channel hardware state indicators of the computing power chip at different frequencies can be collected in real time by a black-box DVFS (dynamic voltage and frequency scaling) method. The black-box DVFS is a dynamic voltage and frequency adjustment technology that does not rely on explicit load monitoring or prediction models. It abstracts the underlying hardware state, reduces the dependence of the software layer on hardware details, and uses a preset strategy or heuristic algorithm to indirectly infer the load state according to the system behavior pattern, thereby dynamically adjusting the voltage and frequency.

[0032] The multi-channel hardware state indicators can include, but are not limited to, chip computing unit utilization, memory utilization, computing frequency, power consumption, temperature, bandwidth occupancy, etc. The chip computing unit utilization can be GPU (Graphics Processing Unit) utilization. The multi-channel hardware state indicators can be used to evaluate the performance of the computing power chip. In the model training phase, the multi-channel hardware state indicators of the computing power chip at different frequencies collected above refer to historical data of the computing power chip working at different workloads, which are used to constitute model training samples for model training.

[0033] As a possible implementation, the process of collecting the multi-channel hardware state indicators of the computing power chip at different frequencies based on the black-box DVFS can include: signal collection, collecting multi-channel signals inside the computing power chip; signal processing, performing weighted processing on the collected signals to obtain a load signal to reflect the influence degree of different signals on the load; load prediction, sending the load signal to an EMA (Exponential Moving Average) module to perform an exponential moving average algorithm, and predicting the load demand in the next time period through the EMA module according to historical load data and the current load signal; frequency and voltage adjustment, comparing the result obtained by the EMA module with a pre-set threshold value, if the predicted performance demand is higher than the upper limit, requesting to increase the frequency and voltage, otherwise, if the predicted performance demand is lower than the lower limit, requesting to decrease the frequency and voltage.

[0034] Further, after collecting a required number of multi-channel hardware state indicators of the computing power chip at different frequencies, a recurrent neural network (RNN) can be used for modeling to learn the computing performance trend at different frequency settings and capture the corresponding relationship between the frequency and the performance data. During the training process, a mapping relationship of power consumption-computing capability-frequency is mainly constructed to extract key features, for example, the performance loss trend when the frequency is reduced, the computing bottleneck features (memory limited, computing limited, etc.), and the power consumption characteristics under the load mode. Through continuous model training, when the number of training times or model convergence is met, the corresponding model is used as the optimal prediction regression model, and based on the optimal prediction regression model, the predicted data of the computing power chip under different workloads can be predicted. In the process of predicting the performance data of the computing power chip at different frequencies through the optimal prediction regression model, the performance data corresponding to one or more frequencies can be used as the optimal frequency selection under the condition of meeting the preset performance index.

[0035] In step S104, the predicted data of the computing power chip under the current workload is obtained, the predicted data is predicted based on the optimal prediction regression model, the target optimal frequency meeting the preset performance index condition is obtained, and the target optimal frequency corresponds to the optimal energy consumption.

[0036] In this embodiment, the predicted data of the computing power chip under the current workload can also be obtained through the black box DVFS method. The predicted data includes but is not limited to GPU utilization, memory utilization, computing frequency, power consumption, temperature, bandwidth occupancy, etc. Then, the trained optimal prediction regression model is used to predict the predicted data sampled at different frequency points according to the state of the current workload. In the prediction process, the frequency meeting the preset performance index condition can be selected as the target optimal frequency output, and the preset performance index condition can include that the performance of the current frequency ≥ 90% of the maximum performance of the maximum frequency. The target optimal frequency can include the smallest frequency in the frequency range predicted by the model under the condition of meeting the preset performance index condition, that is, the performance requirement is guaranteed, and it is also the frequency that makes the computing power chip under the current workload reach the lowest energy consumption.

[0037] In some examples, the sample data for performance evaluation is obtained in the following manner: for performance evaluation of the workload, the performance of the chip can be represented by GFLOPS (floating point operations per second). In particular for AI deep learning, performance evaluation by GFLOPS can measure the performance of each AI benchmark at different execution stages, which is obtained by adding the calculation amount and calculation time of each layer of the deep network through the interface in the thop library of pytorch, and the ratio is the calculated GFLOPS. Forward propagation and backward propagation in AI model training often show significant performance differences, so the performance needs to be measured at the granularity of forward and backward operations for each layer.

[0038] In step S106, the working frequency of the computing power chip under the current workload is adaptively adjusted based on the target optimal frequency.

[0039] In the embodiment, based on the DVFS control method, the target optimal frequency output by the model in real time can be dynamically and adaptively adjusted to adjust the working frequency of the computing power chip, so that when the working frequency of the computing power chip under the current workload is adjusted to the target optimal frequency, the optimization target of the minimum power consumption and the maximum energy efficiency can be met, so as to reduce unnecessary energy loss; at the same time, the computing performance of the computing power chip is ensured, so that the performance loss does not exceed the preset performance index condition; in addition, the adaptive frequency adjustment strategy can intelligently switch the frequency under different workload scenarios, rather than working at a fixed frequency under each workload scenario, so that the balance between power consumption and computing capability can be optimized.

[0040] In the embodiment of the application, the multi-channel hardware state indicators are obtained by the black box DVFS method, without modifying the upper application program code, and are suitable for various artificial intelligence computing loads, so as to ensure the transparency and low overhead of the system; the multi-channel hardware state indicators of the computing power chip at each frequency under the workload are analyzed in advance, an optimal prediction regression model is constructed based on an RNN model, and then the target optimal frequency is determined by predicting the to-be-predicted data of the computing power chip under the current workload. The prediction process can predict the performance trend at different frequencies, so that the target optimal frequency that meets the performance requirement and has the lowest chip energy consumption can be accurately selected, so that the selection of the target optimal frequency can fully reflect the balance between the computing demand of the workload and the available computing resources of the computing power chip, the working frequency of the computing power chip is dynamically adjusted based on the target optimal frequency, the chip performance is ensured, and the energy consumption is reduced, so that the flexibility and applicability of deployment are greatly improved.

[0041] In some optional embodiments, before step S102, the method further comprises:

[0042] The hardware state of the computing power chip is monitored to obtain the multi-channel hardware state indicators of the computing power chip at different frequencies, and the multi-channel hardware state indicators include performance indicators and GPU state indicators of the computing power chip at different frequencies under different workloads;

[0043] Based on the timestamps, the performance indicators and the GPU state indicators of the computing power chip at different frequencies under each server are aligned to obtain a plurality of feature vectors to be trained.

[0044] In this embodiment, before the offline modeling process, the performance indicators and the GPU state indicators of the computing power chip under different frequency settings under different server benchmark tests need to be collected first. The selected benchmark tests include computation-intensive, memory access-intensive and mixed-intensive models, etc. In these benchmark tests, the performance indicators and the corresponding GPU state indicators are recorded, and the indicators are aligned based on timestamps to obtain the performance data of each model unit, and the performance indicators and the GPU state indicators under different frequency settings are captured. The time synchronization protocol can be used to ensure the alignment of the timestamps, and the alignment of the timestamps can avoid data misplacement. For example, if the timestamps are not aligned, the model may mistakenly believe that the high GPU utilization at a certain time is caused by the low frequency. The performance indicators and the GPU state indicators of each model unit constitute an analysis sample of a model, so a plurality of feature vectors to be trained can be obtained.

[0045] Further, the hardware state of the computing power chip can be monitored by a black box DVFS method. In the monitoring process, the multi-channel hardware state indicators include hardware features and real-time features. The hardware features are obtained by using the interfaces in the standard library, such as the nvml library of the nvidia GPU. The real-time features include the graphics computing frequency, the GPU utilization, the memory utilization, the power, etc.

[0046] In this embodiment, the above-mentioned GPU indicators include the GPU utilization and the memory utilization, etc. The GPU utilization and the memory utilization can more effectively represent the use of hardware resources. Therefore, in this embodiment, NVML is used as the main tool for reading the GPU hardware counter, and the command line tool nvidia-smi is used to adjust the GPU frequency and collect the required indicators, including the GPU utilization, the memory utilization, etc.

[0047] In the embodiment, by monitoring the hardware state of the computing chip and constructing the feature vector, high-quality data input can be provided for the subsequent model, thereby improving the model performance and prediction accuracy. Aligning the performance indicators with the GPU state indicators based on the timestamps can avoid the data misplacement problem caused by time deviation. Combining the performance indicators with the GPU state indicators enables the multiple trained feature vectors to contain multidimensional information, which can more comprehensively reflect the running state of the computing chip and provide richer input for the model to improve the expression ability of the model.

[0048] In some optional embodiments, in combination with Figure 2 As shown in the above step S102, comprising:

[0049] S1021, constructing a trained regression model by a recurrent neural network;

[0050] S1022, based on the multiple trained feature vectors constructed, performing model training on the trained regression model to predict a performance data sequence of a continuous frequency value, the trained feature vectors including original values and relative change values of the memory utilization and the GPU utilization at different frequencies;

[0051] S1023, based on the performance data sequence of the continuous frequency value, obtaining a frequency corresponding to a condition that the performance data meets the preset performance indicators as an output frequency of the model training, to construct the optimal prediction regression model.

[0052] In combination with Figure 3 As shown in the above, in the offline modeling process, input features need to be collected. Among the input features, Freq1, Freq2, Freq3, …, represent frequency points, and at each frequency point, data GPU_util representing GPU utilization and data Mem_util representing memory utilization are collected. In the modeling process, RNN is used as the modeling basis, and frequency is used as the label. In the model training process, the loss function can also be used for frequency optimization, so that the optimal frequency output by the training can not only meet the performance (Performance) requirement, but also minimize the energy consumption.

[0053] In some possible embodiments, because the frequency interval is between the default frequency and the maximum frequency when the frequency interval is empty, the gradient is collected from low to high frequency, and therefore, the sequences of the GPU utilization and the memory utilization at different frequencies are ordered and have a dependence relationship between the frequencies. For example, Figure 3As shown, the load starts to calculate saturation as the frequency increases, reaches the bandwidth bottleneck, and the performance interval corresponding to the bandwidth bottleneck is the frequency range that can be optimized, that is, the performance part corresponding to the part above the optimal frequency. This dynamic behavior is sequence-specific, so the method provided in this embodiment is applicable to RNN models, and a regression model to be trained can be constructed through a recurrent neural network.

[0054] Further, in combination with Figure 3 As shown, the input features of the regression model to be trained, GPU utilization and memory utilization, are divided into two categories: original values: including the GPU utilization and memory utilization corresponding to each frequency point Freqn, reflecting the resource consumption of the application at each frequency point, for example, the original values of the GPU utilization and memory utilization corresponding to the frequency point Freq1 are 50% and 30% respectively; the original values of the GPU utilization and memory utilization corresponding to the frequency point Freq2 are 60% and 40% respectively, and the original values of the GPU utilization and memory utilization corresponding to the frequency point Freq3 are 66% and 30% respectively. Relative change value: the utilization at the maximum frequency, the percentage change of the GPU utilization and memory utilization corresponding to each frequency point Freqn collected, indicating the increasing or decreasing trend of resource utilization as the frequency changes, for example, the original values of the GPU utilization and memory utilization at the maximum frequency 1Ghz of the computing chip can reach 90% and 95% respectively, the original values of the GPU utilization and memory utilization corresponding to the frequency point Freq1 are 50% and 30% respectively, then the relative change value of the GPU utilization is (50%-90%) / 90%=-44%, and the relative change value of the memory utilization is (30%-95%) / 95%=-68%. These inputs together constitute a utilization feature sequence under multiple frequencies, which is essentially a time series structure in the frequency dimension. For the label required for offline modeling, the optimal frequency point is used as the reference, and if there is one or more frequencies that can meet the preset performance indicator, they can be selected as the output optimal frequency of the model, and the optimal frequency selected for model training is the smallest frequency in the interval, that is, the frequency that can achieve the lowest power consumption under the condition of guaranteeing performance requirements.

[0055] Further, the regression model to be trained above is used to predict which frequency the load is suitable for running at under different frequency hardware state sequences, that is, to obtain the approximation form of the frequency optimization function through RNN modeling regression, in order to find the frequency corresponding to the lowest power consumption, as shown in the following formula (1).

[0056] (1)

[0057] In formula (1), f represents the frequency; P(f) represents the power consumption with respect to the frequency f, represents the performance of the chip running at frequency f ; Perf f ( f ) represents the performance about frequency function, which represents the performance index (including FLOPS, throughput, delay, etc.) of the chip running at frequency f , generally, the performance increases with the increase of frequency; a represents the coefficient of performance index condition, which represents the proportion of the lower limit of performance allowed to the maximum performance; Perf max represents the maximum performance value.

[0058] Based on the above formula (1), it can be known that the model learns the implicit mapping from the hardware resource response trend of the computing chip → performance satisfaction → optimal frequency, as follows formula (2):

[0059] (2)

[0060] In formula (2), the input data set X is composed of a series of feature vectors x f to be trained, each vector x f corresponds to a hardware state index at frequency f x f , which contains the original value and the relative change value of the memory utilization and GPU utilization of the computing chip at frequency f f , the value range of which is from the minimum frequency f min to the maximum frequency f max ; the output f ∗ represents the optimal frequency, that is, under the premise of meeting the performance index condition, the lowest frequency is selected; Perf f represents the performance index at frequency f ; Perf f max represents the maximum performance at frequency f max ; wherein, the performance index condition is: Perf f ≥ 0.9Perf f max , a = 0.9, that is, the performance is at least 90% of the maximum performance; argmin f represents the minimum frequency f among all frequencies that meet the performance index condition.

[0061] ​​Further, for each layer mechanism of the regression model to be trained, the input layer: takes the feature vector to be trained as input in order from low to high frequency, and each time step in the sequence of the feature vector to be trained represents the state at a frequency point. Each time step input includes: GPU utilization and memory utilization, GPU utilization can be core computing resource consumption GPU utilization; Memory utilization can refer to memory bandwidth pressure Memory utilization, including original value and relative change value, to reveal the potential "response slope" and resource saturation point, so as to enhance the model's perception ability to nonlinear behavior changes. This layer converts the hardware behavior sequence in the discrete frequency space into a time sequence tensor acceptable to the RNN, maintaining the order of the physical change process.

[0062] Hidden state propagation layer: composed of a plurality of same cycle units arranged in frequency sequence, each unit receives the input features of the current frequency point and the hidden state of the last frequency point, and performs formula (3):

[0063] h t =σ( W x t + U h t-1 + b )(3)

[0064] wherein, h t represents the hidden state of the current time step t , which is a vector, represents the state of the model hidden layer at time step t ; σ represents an activation function, including a sigmoid function, a tanh function, etc.; W represents an input weight matrix for mapping the input to the hidden state space; x t represents the input vector of the current time step t ; U represents a hidden state weight matrix for mapping the hidden state of the previous time step h t-1 to the current hidden state space; b represents a bias vector for adjusting the input of the activation function, increasing the flexibility of the model, wherein W, U and b are known quantities.

[0065] wherein each hidden state h tIn fact, the dynamic dependence relationship between the historical resource response and the current state at the current frequency point is actually coded, and through cyclic propagation, a global memory of the overall frequency response trend is formed. This mechanism does not assume a fixed response function, but rather lets the network automatically learn which frequency bands are sensitive to resource changes and at which points the performance converges. Therefore, this layer has the ability to model "saturation points" and "non-linear response inflection points".

[0066] Final hidden state reading layer: In this layer, the last hidden state hT of the RNN is regarded as a compressed representation of the entire frequency sequence behavior, that is, a result extraction of the overall performance resource response trend from low frequency to high frequency. This vector is not just some numerical transformation, but encodes cross-frequency behavior patterns such as performance change intervals in the sequence, frequency bands where bandwidth bottlenecks occur, etc. through time structure. Regression output layer: This layer performs linear regression mapping on the final hidden state and outputs a performance data sequence of continuous frequency values, representing the minimum adaptation frequency of the workload at hand under the premise of keeping the performance ≥90% of the maximum performance, i.e., the energy-optimal point. That is, the frequency corresponding to the performance data that meets the preset performance indicator condition is obtained as the output frequency of the model training. This layer learns the implicit objective function approximation form from the performance data sequence trend to the optimal point, rather than explicitly modeling the polynomial relationship chain of "frequency → power consumption → performance", thereby enhancing the generalization ability of the model. Based on the above method, the optimal prediction regression model is obtained through training, and the output result of the training is the output frequency described above.

[0067] In this embodiment, the regression model to be trained is constructed by RNN, the input vector fuses the original values of the memory utilization rate and the GPU utilization rate of the computing power chip at different frequencies and their relative change values, the dynamic correlation between the hardware state and the frequency is captured by using the time series modeling capability of RNN, so that the performance data sequence corresponding to the continuous frequency value can be predicted; during the model training process, the frequency meeting the preset performance indicator condition is selected as the output target based on the predicted performance data sequence, the model is accurately fitted to the relationship between the optimal frequency and the hardware state through iterative optimization, and finally the optimal model capable of efficiently outputting the optimal frequency configuration according to real-time hardware data is generated, which significantly improves the energy efficiency performance and performance stability of the computing power chip under dynamic load.

[0068] In some optional embodiments, after step S102, the method further includes:

[0069] An optimization target of the loss function is constructed, and the optimization target is to minimize the difference between the predicted frequency and the minimum actual frequency meeting the preset performance indicator condition;

[0070] Based on the optimization target, the output frequency of the model training is optimized by the loss function.

[0071] In this embodiment, to improve the accuracy of frequency prediction, an optimization objective can be set for the frequency. This objective can be: minimizing the difference between the predicted frequency and the minimum actual frequency required to meet performance requirements. Based on the optimization objective, the output frequency of the model training is optimized using a loss function. This loss function can include, but is not limited to, mean squared error (MSE) loss function, mean absolute error (MAE) loss function, Huber Loss, and bounding loss.

[0072] In some alternative embodiments, combined with Figure 4 As shown, step S104 above includes:

[0073] S1041. Obtain the data to be predicted from the computing power chip under the current workload, wherein the data to be predicted includes the original values ​​and relative changes of the memory utilization and the GPU utilization sampled by the computing power chip at different frequencies under the current workload.

[0074] S1042. Using the original values ​​and relative changes of the memory utilization and GPU utilization sampled by the computing chip at different frequencies under the current workload as input parameters, the optimal prediction regression model is used to predict and output the target optimal frequency corresponding to the performance data of the computing chip under the current workload meeting the preset performance index conditions.

[0075] In this embodiment, combined with Figure 3 As shown, the real-time GPU utilization (Gpu Util.) and memory utilization (Mem Util.) of the computing chip under the current workload can be obtained by sampling with a counter, and the original values ​​and relative changes are used as the data to be predicted and input into the trained model. It should be noted that the frequency corresponding to the data to be predicted must be consistent with the frequency used during RNN training to ensure data consistency in frequency and comparability.

[0076] Furthermore, during the decision-making phase, the optimal predictive regression model outputs the target optimal frequency at which the performance data of the computing chip under the current workload meets the preset performance index conditions, based on the original values ​​and relative changes of the input GPU utilization and memory utilization. To reduce data collection overhead, this embodiment can use only the memory utilization and GPU utilization at three frequency points for prediction.

[0077] In the embodiment, the original values and the relative change values of the memory utilization and the GPU utilization of the computing power chip sampled at different frequencies under the current workload are taken as the input parameters of the optimal prediction regression model, so that the target optimal frequency of the performance data of the computing power chip under the current workload satisfying the preset performance index condition can be predicted, the target optimal frequency corresponding to the lowest power consumption can be found while the performance demand is guaranteed, and the working frequency of the computing power chip is adjusted to the target optimal frequency, so that the performance is guaranteed while the power consumption is reduced.

[0078] In some optional embodiments, after step S1041, the method further includes:

[0079] Based on the preset trigger mechanism, the memory utilization and / or the GPU utilization of the computing power chip in the history are compared with the memory utilization and / or the GPU utilization of the computing power chip under the current workload, and it is determined whether the change value of any utilization exceeds the preset utilization change threshold value;

[0080] If at least one utilization change value exceeds the preset utilization change threshold value, the frequency is updated.

[0081] In the embodiment, according to the target optimal frequency output by the optimal prediction regression model, the system can further perform frequency updating. In the embodiment, the frequency can be updated only when a certain index or multiple indexes of the workload change significantly, including: the utilization change value of the memory utilization of the computing power chip under the current workload and the historical memory utilization used in the model training exceeds the preset utilization change threshold value, such as Figure 3 Δ(Mem Util.)>Threshold); or the utilization change value of the GPU utilization of the computing power chip under the current workload and the historical GPU utilization used in the model training exceeds the preset utilization change threshold value, such as Figure 3 Δ(GPU Util.)>Threshold); or the utilization change values of both of them and the historical utilization of both of them used in the model training all exceed the preset utilization change threshold value, that is, Δ(Mem Util.)>Threshold) and Δ(GPU Util.)>Threshold) are satisfied at the same time. For example, when the utilization change value of the GPU utilization of the computing power chip under the current workload relative to the GPU utilization used in the model training exceeds the 20% utilization change threshold value A, the frequency is updated.

[0082] In the embodiment, the efficiency and pertinence of the optimization process are ensured through the above mechanism. The historical memory utilization rate and the historical GPU utilization rate can be compared by taking values in the form of average to reduce the calculation amount. In the embodiment, the efficiency and pertinence of the optimization process are ensured through the setting of the triggering mechanism.

[0083] In some optional embodiments, before step S104, the method further comprises:

[0084] Obtaining temperature data and power consumption data of the computing power chip at different frequencies;

[0085] Constructing an optimal prediction regression model based on the temperature data, the power consumption data of the computing power chip at different frequencies, and the multi-channel hardware state indicators, to predict the performance data of the computing power chip at different frequencies.

[0086] In some examples, temperature sensing and real-time power consumption monitoring can also be added, temperature and power consumption feature data can be introduced into the original RNN model, and the optimal prediction regression model can be constructed based on the multi-channel hardware state indicators of the computing power chip. Based on this, the performance data of the computing power chip at different frequencies can be predicted. By introducing temperature data and power consumption data for model training and prediction, the safety and stability of the system under extreme load or high temperature environment can be further improved.

[0087] It should be noted that when the temperature data and power consumption data are introduced into the RNN model for training, the temperature data and power consumption data of the computing power chip under the current workload to be predicted can also be included in the predicted data, which is based on this to predict, so as to consider the case of extreme load and high temperature environment in performance prediction, and select the target optimal frequency in combination with the extreme load and high temperature environment, to further improve the safety and stability of the system under extreme load or high temperature environment.

[0088] According to another aspect of the embodiments of the present application, as Figure 5 shown, corresponding to the adaptive energy consumption control method based on the computing power chip in the above embodiment, the present embodiment provides an adaptive energy consumption control system based on the computing power chip, which comprises:

[0089] A model construction module 501 is configured to construct an optimal prediction regression model based on the multi-channel hardware state indicators of the computing power chip at different frequencies, to predict the performance data of the computing power chip at different frequencies.

[0090] A prediction module 503 is configured to obtain to-be-predicted data of the computing power chip under a current workload, to predict the to-be-predicted data based on the optimal prediction regression model, to obtain a target optimal frequency that meets a preset performance index condition, and the target optimal frequency corresponds to optimal energy consumption.

[0091] An adjusting module 505 is configured to adaptively adjust the working frequency of the computing power chip under the current working load based on the target optimal frequency.

[0092] It should be noted that in this embodiment, the model constructing module 501 can be configured to perform step S102 in the embodiments of the present application, the prediction module 503 in this embodiment can be configured to perform step S104 in the embodiments of the present application, and the adjusting module 505 in this embodiment can be configured to perform step S106 in the embodiments of the present application.

[0093] In some optional embodiments, the system further comprises a data acquisition module configured to monitor the hardware state of the computing power chip, and acquire the multi-channel hardware state indicators of the computing power chip under different frequencies, wherein the multi-channel hardware state indicators comprise performance indicators and GPU state indicators of the computing power chip under different frequencies and different workloads; and a data processing module configured to align the performance indicators and the GPU state indicators of the computing power chip under each server under different frequencies based on timestamps, so as to construct a plurality of feature vectors to be trained.

[0094] In some optional embodiments, the model constructing module 501 comprises a model constructing submodule configured to construct a regression model to be trained by using a recurrent neural network; a model training submodule configured to perform model training on the regression model to be trained based on the plurality of feature vectors to be trained constructed, and predict a performance data sequence of continuous frequency values, wherein the feature vectors to be trained comprise original values and relative change values of the memory utilization rate and the GPU utilization rate under different frequencies; and a frequency output submodule configured to acquire a frequency corresponding to the performance data satisfying the preset performance indicator condition as an output frequency of model training based on the performance data sequence of continuous frequency values, so as to construct the optimal prediction regression model.

[0095] In some optional embodiments, the system further comprises an optimization target constructing module configured to construct an optimization target of a loss function, wherein the optimization target is to minimize the difference between the predicted frequency and the minimum actual frequency satisfying the preset performance indicator condition; and an optimization module configured to optimize the output frequency of model training by using the loss function based on the optimization target.

[0096] In some optional embodiments, the prediction module 503 includes: a data acquisition submodule for acquiring the data to be predicted of the computing chip under the current workload, wherein the data to be predicted includes the original values ​​and relative changes of the memory utilization and GPU utilization sampled at different frequencies under the current workload; and a model prediction submodule for using the original values ​​and relative changes of the memory utilization and GPU utilization sampled at different frequencies under the current workload as input parameters, and performing prediction through the optimal prediction regression model to output the target optimal frequency corresponding to the performance data of the computing chip under the current workload satisfying the preset performance index conditions.

[0097] In some optional embodiments, the system further includes: a judgment submodule, configured to, based on a preset triggering mechanism, compare the historical memory utilization and / or GPU utilization of the computing chip with the memory utilization and / or GPU utilization of the computing chip under the current workload, and determine whether any of the utilization changes exceeds a preset utilization change threshold; and a triggering submodule, configured to, if at least one utilization change exceeds the preset utilization change threshold, trigger a frequency update.

[0098] In some optional embodiments, the system further includes: a data acquisition module, further configured to acquire temperature data and power consumption data of the computing chip at different frequencies; and a model building module, further configured to build an optimal predictive regression model based on the temperature data, power consumption data and multi-channel hardware status indicators of the computing chip at different frequencies, in order to predict the performance data of the computing chip at different frequencies.

[0099] It should be noted that the examples and application scenarios implemented by the above modules, sub-modules, and corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. It should also be noted that the above modules and sub-modules, as part of the system, can run in the hardware environment of the adaptive energy consumption control system based on computing chips, and can be implemented through software or hardware.

[0100] According to another aspect of the embodiments of this application, this application provides a computer device, such as... Figure 6 As shown, it includes a memory 601, a processor 603, a communication interface 605, and a communication bus 607. The memory 601 stores a computer program that can run on the processor 603. The memory 601 and the processor 603 communicate through the communication interface 605 and the communication bus 607. When the processor 803 executes the computer program, it implements the steps of the above-mentioned adaptive energy consumption control method based on computing power chip.

[0101] The memory, the processor in the computer device above communicate through a communication bus and a communication interface. The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0102] The memory above can include a Random Access Memory (RAM) and can also include a non-volatile memory, such as at least one disk memory. Optionally, the memory can also be at least one storage device located away from the processor above.

[0103] The processor above can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0104] According to another aspect of the embodiments of the present application, a computer program product or computer program is also provided, which includes computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps of the adaptive energy consumption control method based on the computing chip in any of the embodiments above.

[0105] Optionally, in the embodiments of the present application, the computer readable medium is configured to store program codes for the processor to execute the steps of the adaptive energy consumption control method based on the computing chip in the embodiments above, and the steps of the adaptive energy consumption control method based on the computing chip specifically include:

[0106] S102, based on the multi-channel hardware state indicators of the computing chip at different frequencies, an optimal prediction regression model is constructed to predict the performance data of the computing chip at different frequencies.

[0107] S104, acquire the to-be-predicted data of the computing power chip under the current workload, perform prediction on the to-be-predicted data based on the optimal prediction regression model, acquire a target optimal frequency under a preset performance index condition, and the target optimal frequency corresponds to optimal energy consumption;

[0108] S106, perform self-adaptive adjustment on the working frequency of the computing power chip under the current workload based on the target optimal frequency.

[0109] Optionally, specific examples in the embodiments can refer to the examples described in the above-described embodiments, and the embodiments will not be described here again. When the embodiments are implemented, reference can be made to the above-described embodiments, and the corresponding technical effects are achieved.

[0110] It can be understood that the embodiments described herein can be realized by hardware, software, firmware, middleware, microcode or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), general purpose processors, controllers, micro-controllers, microprocessors, other electronic units for performing the functions described in the present application, or a combination thereof. For software implementation, the units described herein can be implemented by units performing the functions described herein. The software code can be stored in a memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0111] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be described here again.

[0112] In the embodiments provided by the present application, it should be understood that the disclosed apparatus and method can be implemented in other manners. For example, the division of the modules is merely logical function division, and there can be another division manner in actual implementation. For example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different modules can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0113] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the embodiments of the present application essentially or the parts that make contributions to the prior art or the parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application. The foregoing storage medium includes: U disk, mobile hard disk, ROM, RAM, magnetic disk or optical disk, and various other media that can store program codes.

[0114] It should be noted that, in this document, the terms such as first, second, etc. are merely used to distinguish one entity or operation from another, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the term includes, includes or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such a process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0115] The above description is merely one specific implementation of the present application, enabling one skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for adaptive energy consumption control based on a computing power chip, characterized in that, The method comprises: Based on the multi-channel hardware state indicators of the computing power chip at different frequencies, an optimal prediction regression model is constructed to predict the performance data of the computing power chip at different frequencies; Obtain the to-be-predicted data of the computing power chip under the current workload, predict the to-be-predicted data based on the optimal prediction regression model, obtain the target optimal frequency that meets the preset performance index condition, and the target optimal frequency corresponds to the optimal energy consumption; Based on the target optimal frequency, the working frequency of the computing power chip under the current workload is adaptively adjusted; The multi-channel hardware state indicators include memory utilization and GPU utilization, and the optimal prediction regression model is constructed based on the multi-channel hardware state indicators of the computing power chip at different frequencies, which comprises: constructing a to-be-trained regression model through a recurrent neural network; based on the plurality of to-be-trained feature vectors constructed, the to-be-trained regression model is trained, and a performance data sequence of a continuous frequency value is predicted, wherein the to-be-trained feature vectors include original values and relative change values of the memory utilization and the GPU utilization at different frequencies; based on the performance data sequence of the continuous frequency value, a frequency corresponding to the performance data meeting the preset performance index condition is obtained as an output frequency of model training, so as to construct the optimal prediction regression model.

2. The hash rate chip-based adaptive energy consumption control method according to claim 1, characterized in that, Before the optimal prediction regression model is constructed based on the multi-channel hardware state indicators of the computing power chip at different frequencies, the method further comprises: Monitoring the hardware state of the computing power chip to obtain the multi-channel hardware state indicators of the computing power chip at different frequencies, which include performance indicators and GPU state indicators of the computing power chip running at different frequencies under different workloads; Based on the timestamp, the performance indicators and the GPU state indicators of the computing power chip running at different frequencies under each server are aligned to construct a plurality of to-be-trained feature vectors.

3. The hash chip-based adaptive energy consumption control method of claim 1, wherein, After the optimal prediction regression model is constructed based on the performance data sequence of the continuous frequency value, the frequency corresponding to the performance data meeting the preset performance index condition is obtained as an output frequency of model training, the method further comprises: An optimization objective of a loss function is constructed, and the optimization objective is to minimize the difference between the predicted frequency and the minimum actual frequency meeting the preset performance index condition; Based on the optimization objective, the output frequency of model training is optimized through the loss function.

4. The hash chip-based adaptive energy consumption control method of claim 1, wherein, The to-be-predicted data of the computing power chip under the current workload is obtained, and the to-be-predicted data is predicted based on the optimal prediction regression model to obtain the target optimal frequency that meets the preset performance index condition, which comprises: In the to-be-predicted data of the computing power chip under the current workload, the original values and the relative change values of the memory utilization and the GPU utilization sampled by the computing power chip at different frequencies under the current workload are included; The memory utilization rate and the GPU utilization rate of the computing power chip under the current workload are sampled at different frequencies, and the original values and the relative change values of the memory utilization rate and the GPU utilization rate are taken as input parameters. The optimal prediction regression model is used for prediction to output the target optimal frequency corresponding to the condition that the performance data of the computing power chip under the current workload meets the preset performance index.

5. The hash chip-based adaptive energy consumption control method of claim 4, wherein, After obtaining the to-be-predicted data of the computing power chip under the current workload, the method further comprises: Based on the preset triggering mechanism, the memory utilization rate and / or the GPU utilization rate of the computing power chip history are compared with the memory utilization rate and / or the GPU utilization rate of the computing power chip under the current workload, and it is judged whether the utilization rate change value of any one exceeds the preset utilization rate change threshold value; If at least one utilization rate change value exceeds the preset utilization rate change threshold value, the frequency update is triggered.

6. The hash chip-based adaptive energy consumption control method according to any one of claims 1 to 5, characterized in that, Before obtaining the to-be-predicted data of the computing power chip under the current workload, predicting the to-be-predicted data based on the optimal prediction regression model, and obtaining the target optimal frequency under the condition that the preset performance index is met, the method further comprises: Obtain the temperature data and power consumption data of the computing power chip at different frequencies; Based on the temperature data, power consumption data and the multi-channel hardware state indicators of the computing power chip at different frequencies, an optimal prediction regression model is constructed to predict the performance data of the computing power chip at different frequencies.

7. A hash chip-based adaptive energy consumption control system for implementing the hash chip-based adaptive energy consumption control method according to any one of claims 1 to 6, characterized in that, The system comprises: A model construction module is configured to construct an optimal prediction regression model based on the multi-channel hardware state indicators of the computing power chip at different frequencies to predict the performance data of the computing power chip at different frequencies. A prediction module is configured to obtain to-be-predicted data of the computing power chip under the current workload, predict the to-be-predicted data based on the optimal prediction regression model, and obtain a target optimal frequency under the condition that a preset performance index is met, the target optimal frequency corresponding to optimal energy consumption. An adjustment module is configured to adaptively adjust the working frequency of the computing power chip under the current workload based on the target optimal frequency.

8. A computer device comprising: A processor, a memory and a network interface, the memory storing machine readable instructions executable by the processor, characterized in that when the computer device is running, the processor and the memory communicate through the network interface, and the processor executes the machine readable instructions to perform the steps of the adaptive energy consumption control method based on the computing power chip according to any one of claims 1 to 6.

9. A computer readable medium having a non-transitory program code executable by a processor, the program code comprising instructions for: The program code causes the processor to perform the steps of the adaptive energy consumption control method based on the computing power chip according to any one of claims 1 to 6.

Images