Solid state disk power consumption optimization method and system based on load prediction

Abstract

The application discloses a solid state disk power consumption optimization method and system based on load prediction, and particularly relates to the technical field of solid state disks, which comprises the following steps: S1. data acquisition; S2. load prediction model construction and training; S3. load prediction; S4. power consumption optimization: according to the load prediction result, a multi-objective optimization dynamic power consumption adjustment algorithm is used to optimize the power consumption of the solid state disk; S5. feedback adjustment: the actual power consumption and performance parameters of the optimized solid state disk are monitored in real time, and the load prediction model and the dynamic power consumption adjustment algorithm are adjusted in combination with a reinforcement learning algorithm. The solid state disk power consumption optimization method and system based on load prediction form a complete and efficient solid state disk power consumption optimization scheme from the load prediction model construction, the dynamic power consumption adjustment algorithm design to the feedback adjustment mechanism optimization, and are expected to provide a new technical thought and practical direction for the solid state disk in terms of energy consumption reduction and performance improvement.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of solid state disks, more specifically, the present application relates to a solid state disk power consumption optimization method and system based on load prediction. BACKGROUND

[0002] With the rapid development of information technology, solid state disks have been widely used in computer, mobile device, data center and other fields due to their high speed reading and writing, low power consumption, anti-vibration and other advantages. However, in the context of large-scale data storage and processing, the power consumption problem of solid state disks is increasingly prominent, and the high power consumption not only increases the operating cost of the device, but also causes the device to heat up, affecting its stability and service life.

[0003] At present, the existing solid state disk power consumption optimization method is mostly based on real-time load passive adjustment, that is, when the load changes, the power consumption is adjusted. This method has certain hysteresis and cannot make power consumption adjustment in advance before the load changes, resulting in unsatisfactory power consumption optimization effect. At the same time, the existing load prediction model has low prediction accuracy, which is difficult to accurately grasp the trend of load change, further affecting the effect of power consumption optimization;

[0004] Therefore, in view of the above problems, a solid state disk power consumption optimization method and system based on load prediction are proposed. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the present application provides a solid state disk power consumption optimization method and system based on load prediction to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides the following technical scheme: a solid state disk power consumption optimization method based on load prediction, the method comprising the following steps:

[0007] S1. Data acquisition: collect historical load data and corresponding power consumption data of solid state disk, the historical load data including I / O request quantity, data transmission quantity, request response time, request type;

[0008] S2. Load prediction model construction and training: based on the collected historical load data, a bidirectional LSTM-Transformer load prediction model is constructed, and the model is trained by using the historical load data, so as to obtain the trained load prediction model;

[0009] S3. Load prediction: input the real-time collected current load data of solid state disk into the trained load prediction model to obtain the load prediction result in the future preset time period;

[0010] S4. Power consumption optimization: according to the load prediction result, a multi-objective optimization dynamic power consumption adjustment algorithm is used to optimize the power consumption of the solid state disk, the dynamic power consumption adjustment algorithm includes a voltage adjustment sub-algorithm, a frequency adjustment sub-algorithm, a storage unit activation number adjustment sub-algorithm and a cache strategy adjustment sub-algorithm based on the load prediction result;

[0011] S5. Feedback adjustment: real-time monitoring of the actual power consumption and performance parameters of the optimized solid state disk, and adjusting the load prediction model and dynamic power consumption adjustment algorithm combined with a reinforcement learning algorithm;

[0012] The specific steps of the feedback adjustment in S5 are:

[0013] S51. Constructing a real-time monitoring parameter system;

[0014] The real-time monitoring parameter system adopts a three-layer monitoring architecture, and core parameters are collected every 5ms, including:

[0015] Power consumption parameters: core chip power supply voltage, working current, flash array power consumption proportion;

[0016] Performance parameters: random read / write IOPS, continuous read / write bandwidth, average response time, read / write delay jitter value;

[0017] Prediction deviation parameters: instantaneous error of load prediction value and actual value, i.e. Sliding window error mean, i.e. , wherein represents the instantaneous error of the load prediction value at time t and the actual value, is the load prediction value at time t, is the load actual value at time t, represents the sliding window error mean, and the window size , i.e. the instantaneous errors of the previous 20 time points including the current time are averaged;

[0018] All parameters are transmitted to the monitoring database through an encrypted channel, stored in a time series database, and the latest 72 hours of high-frequency data and 90 days of minute-level summary data are retained;

[0019] S52. Reinforcement learning algorithm deployment;

[0020] A deep Q network (DQN) is used as the core framework of the reinforcement learning, specifically:

[0021] State space definition: , wherein represents​​​ The state vector at any given time integrates multiple key parameters related to the power consumption and operating status of the solid-state drive. The power consumption at the current moment. The reference power consumption is a preset baseline value used to measure the power consumption at the current moment. Normalization is performed. This is the historical average load. This is the frequency adjustment amount. For voltage regulation, For current performance, Due to performance constraints, As the reference frequency, The reference voltage;

[0022] Motion space design: Discretized motion sets The strategy corresponding to each action is:

[0023] Load prediction model parameter adjustment: dynamically adjust the learning rate, support gradient update amplitude adjustment of ±10% or ±20%, and optimize model training efficiency;

[0024] Dynamic power consumption algorithm parameter calibration: Adjust key coefficients to support parameter fluctuations of ±5% or ±10% to achieve a dynamic balance between power consumption and performance;

[0025] Reward function optimization: , , These are the weighting coefficients;

[0026] Experience replay mechanism: Maintain an experience pool with a capacity of 10,000, and randomly select 32 samples for training in each iteration. - Greedy strategy, and It decreases linearly from 0.9 to 0.1;

[0027] S53. Dynamic adjustment of the load forecasting model;

[0028] When there are 5 consecutive monitoring cycles, i.e., 25ms At that time, model fine-tuning is triggered:

[0029] Attention weight matrix The updated formula is: , To enhance the learning rate of learning output, Represents the loss function Regarding the attention weight matrix The gradient;

[0030] LSTM Gating Parameter Correction: Forget Gate Bias Added to: ;

[0031] S54. Adaptive adjustment of dynamic power consumption algorithm;

[0032] Voltage regulation sub-algorithm correction:

[0033] Real-time calculation of voltage regulation error , is the actual voltage, is the target voltage, when , , wherein, is the voltage regulation coefficient, is the corrected voltage regulation coefficient;

[0034] Frequency regulation coefficient optimization:

[0035] Based on the energy consumption-performance ratio , dynamic adjustment: wherein is the preset optimal energy consumption-performance ratio, is the input-output operation rate of the solid state disk at time, is the actual power consumption of the solid state disk at time, is the corrected frequency regulation coefficient, .

[0036] Further, the specific process of constructing the bidirectional LSTM-Transformer load prediction model with fusion attention mechanism in S2 is:

[0037] S21. Use bidirectional LSTM network to extract features from historical load data, capture bidirectional time series dependence of data;

[0038] S22. Input the features output by the bidirectional LSTM network into the encoder of the Transformer model, and weight the features through the attention mechanism to highlight important features;

[0039] S23. Input the features output by the encoder into the decoder of the Transformer model to obtain the load prediction result.

[0040] Further, the forward propagation and backward propagation formulas of the bidirectional LSTM network are as follows:

[0041] Forward propagation formula:

[0042] ,

[0043] ,

[0044] ,

[0045] ,

[0046] ;

[0047] Backpropagation formula:

[0048] ,

[0049] ,

[0050] ,

[0051] ,

[0052] ;

[0053] in, , , Forward propagation Input gate, forget gate, and output gate at any time. For positive propagation Cellular state at any given moment This represents the hidden state from the previous moment in the forward propagation. The hyperbolic tangent activation function is used. , , , The inputs during forward propagation are respectively The weight matrices for the input gate, forget gate, cell state update, and output gate; , , , These are the hidden states from the previous time step during forward propagation. The weight matrices for the input gate, forget gate, cell state update, and output gate; , , , These are the bias terms for the input gate, forget gate, cell state update, and output gate in forward propagation, respectively.

[0054] , , In reverse propagation Input gate, forget gate, and output gate at any time. For backpropagation Cellular state at any given moment This represents the hidden state in the next time step during backpropagation. , , , are weight matrices of input gate, forget gate, cell state update, output gate in back propagation, respectively; , , , , are weight matrices of next time hidden state, input gate, forget gate, cell state update, output gate in back propagation, respectively, , , , , are bias terms of input gate, forget gate, cell state update, output gate in back propagation, respectively, superscript represents forward, superscript represents backward, is a sigmoid function, is element multiplication, is input data at time ; bidirectional LSTM output feature ;

[0055] The attention mechanism formula is: wherein, are query matrix, key matrix, value matrix, respectively, is feature dimension, is feature importance weight matrix, is matrix transpose symbol, represents transposing the key matrix, represents the attention mechanism formula; the decoder output layer adopts linear transformation and softmax function to obtain the final prediction value: wherein, is output weight matrix, is output bias, is decoder output; in the model training process, an improved loss function is adopted, combining mean square error and prediction bias penalty term: wherein is prediction value, is actual value, is penalty coefficient, represents mean square error, and the model parameters are adjusted through the back propagation algorithm.

[0056] Further, the specific steps of the voltage adjustment sub-algorithm based on the load prediction result in the S4 are: S411. Based on the load prediction value and the performance constraint, the target voltage value is calculated through a voltage optimization formula, and the voltage optimization formula is: wherein, is a load prediction value, is a performance constraint, is a reference voltage, is a voltage adjustment coefficient, is a performance compensation coefficient, is a current performance;

[0057] S412. Adjust the working voltage of the solid state disk to the target voltage value according to the preset voltage adjustment step length, so that the voltage change rate does not exceed the voltage change rate threshold, and the adjustment formula is: wherein, is a preset voltage adjustment step length, is a voltage change rate threshold, and are the voltages at and moments, respectively, is a sign function.

[0058] Further, the specific steps of the frequency adjustment sub-algorithm based on the load prediction result in S4 are:

[0059] S421. According to the load prediction result, the target working frequency is calculated by the frequency optimization formula, and the frequency optimization formula is: wherein, is a load prediction result, is a target working frequency, is a reference frequency, is a reference load, is a frequency adjustment coefficient, is an energy consumption coefficient, is a current power consumption, is an optimal power consumption;

[0060] S422. The current working frequency of the solid state disk is adjusted to the target working frequency in a smooth transition manner, and the frequency change formula is: wherein, is a smooth transition coefficient, and are the frequencies at and moments, respectively.

[0061] Further, the specific steps of the storage unit activation quantity adjustment sub-algorithm based on the load prediction result in S4 are:

[0062] S431. According to the load prediction result and the storage unit performance parameter, the required activated storage unit quantity is calculated by the storage unit quantity calculation formula, and the storage unit quantity calculation formula is: wherein, the number of storage units, a preset time, a single storage unit processing capacity, a redundancy coefficient;

[0063] S432. Activate the corresponding number of storage units, and close the redundant storage units.

[0064] Further, the specific steps of the S4 cache strategy adjustment sub-algorithm are:

[0065] S441. Determine the data access hotspots according to the load prediction results;

[0066] S442. Based on the data access hotspots and the cache capacity, calculate the cache allocation ratio of different data through the cache allocation formula, and the cache allocation formula is: wherein: is the cache allocation ratio of the th data, is the access probability of the th data, is the sum of all data access probabilities, is the cache capacity;

[0067] S443. Adjust the cache strategy according to the calculated cache allocation ratio, and preferentially cache high access probability data.

[0068] Further, the solid state disk power consumption optimization system based on load prediction, the system comprises:

[0069] a data acquisition module: for collecting historical load data and corresponding power consumption data of the solid state disk, and collecting current load data, actual power consumption and performance parameters of the solid state disk after optimization in real time;

[0070] a load prediction module: for constructing a bidirectional LSTM-Transformer load prediction model based on the historical load data collected by the data acquisition module, and training, and for inputting the real-time collected current load data into the trained model to obtain the load prediction result;

[0071] a power consumption optimization module: for optimizing the power consumption of the solid state disk by using a multi-objective optimization dynamic power consumption adjustment algorithm according to the load prediction result obtained by the load prediction module;

[0072] a feedback adjustment module: for adjusting the load prediction model in the load prediction module and the dynamic power consumption adjustment algorithm in the power consumption optimization module according to the actual power consumption and performance parameters after optimization collected by the data acquisition module, combined with a reinforcement learning algorithm.

[0073] The technical effects and advantages of the present application are as follows:

[0074] Compared with the prior art, the solid state disk power consumption optimization method and system based on load prediction form a complete and efficient solid state disk power consumption optimization scheme from load prediction model construction, dynamic power consumption adjustment algorithm design to feedback adjustment mechanism optimization, which is expected to provide new technical ideas and practical directions for reducing energy consumption and improving performance of solid state disks, and promote further development and application of solid state disk technology in the field of low power consumption.

[0075] The LSTM-Transformer load prediction model with fusion attention mechanism is used for load prediction. This model combines the advantages of LSTM in capturing time-dependent relationships and Transformer self-attention mechanism in highlighting important features, improving the accuracy of load prediction and enabling more accurate grasp of the trend of solid state disk load changes.

[0076] Based on the load prediction results, dynamic power consumption adjustment algorithm is used for power consumption optimization, including voltage, frequency and storage unit activation number adjustment, realizing multi-dimensional optimization of power consumption. And it can adjust the power consumption in advance according to the prediction results, avoiding the hysteresis of traditional passive adjustment, improving the timeliness and effectiveness of power consumption optimization.

[0077] The feedback adjustment module is set up, which can adjust the load prediction model and dynamic power consumption adjustment algorithm according to the actual running situation, ensuring the stability and continuity of the power consumption optimization effect, so that the system can adapt to different application scenarios and load changes. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 The method flowchart of the present application is shown in the figure;

[0079] Figure 2 The system flowchart of the present application is shown in the figure;

[0080] Figure 3 The flowchart of constructing bidirectional LSTM-Transformer load prediction model with fusion attention mechanism in the present application is shown in the figure;

[0081] Figure 4 The flowchart of voltage adjustment sub-algorithm based on load prediction results in the present application is shown in the figure;

[0082] Figure 5 The flowchart of frequency adjustment sub-algorithm based on load prediction results in the present application is shown in the figure;

[0083] Figure 6 The flowchart of storage unit activation number adjustment sub-algorithm based on load prediction results in the present application is shown in the figure;

[0084] Figure 7 Flowchart of the process of the cache strategy adjustment sub-algorithm in the present application;

[0085] Figure 8 Flowchart of the specific steps of the feedback adjustment in the present application.

[0086] The reference signs are: 1, data acquisition module; 2, load prediction module; 3, power consumption optimization module; 4, feedback adjustment module. DETAILED DESCRIPTION

[0087] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0088] Embodiment 1

[0089] As shown in the load prediction-based solid state disk power consumption optimization method shown in Figure 1 and Figures 3-8 , the load prediction-based solid state disk power consumption optimization method, the method comprising the following steps:

[0090] S1. Data acquisition: collecting historical load data and corresponding power consumption data of a solid state disk, the historical load data including I / O request quantity, data transmission quantity, request response time, request type;

[0091] S2. Load prediction model construction and training: based on the collected historical load data, a bidirectional LSTM-Transformer load prediction model fused with an attention mechanism is constructed, and the model is trained using the historical load data to obtain a trained load prediction model;

[0092] The specific process of constructing the bidirectional LSTM-Transformer load prediction model fused with an attention mechanism in S2 is as follows:

[0093] S21. Feature extraction of historical load data using a bidirectional LSTM network to capture bidirectional time sequence dependency of data;

[0094] S22. Inputting the features output by the bidirectional LSTM network into the encoder of the Transformer model to perform weighted processing on the features through the attention mechanism to highlight important features;

[0095] S23. Inputting the features output by the encoder into the decoder of the Transformer model to obtain a load prediction result.

[0096] The forward propagation and backward propagation formulas of the bidirectional LSTM network are as follows:

[0097] The forward propagation formula is:

[0098] ,

[0099] ,

[0100] ,

[0101] ,

[0102] ;

[0103] The backward propagation formula is:

[0104] ,

[0105] ,

[0106] ,

[0107] ,

[0108] ;

[0109] wherein, , , are the input gate, forget gate, and output gate at time t in the forward propagation, is the cell state at time t in the forward propagation, is the hidden state at the previous time in the forward propagation, is the hyperbolic tangent activation function, , , , , , are the weight matrices of the input to the input gate, forget gate, cell state update, and output gate in the forward propagation; , , , are the weight matrices of the hidden state at the previous time to the input gate, forget gate, cell state update, and output gate in the forward propagation; , , , are the bias terms of the input gate, forget gate, cell state update, and output gate in the forward propagation;

[0110] 、 、 respectively are the input gate, forget gate, output gate at time t in the forward propagation, is the cell state at time t in the forward propagation, is the hidden state at next time t+1 in the forward propagation, 、 、 、 、 、 respectively are the weight matrices of input to input gate, forget gate, cell state update, output gate at next time t+1 in the backward propagation, 、 、 、 respectively are the weight matrices of hidden state to input gate, forget gate, cell state update, output gate at time t in the backward propagation, 、 、 、 respectively are the bias terms of input gate, forget gate, cell state update, output gate in the backward propagation, superscript represents forward, superscript represents backward, is a sigmoid function, is an element multiplication, is the input data at time t; the bidirectional LSTM output feature ; ;

[0111] The attention mechanism formula is: wherein, respectively are the query matrix, key matrix, value matrix, is the feature dimension, is the feature importance weight matrix, is a matrix transpose symbol, indicates that the key matrix is transposed, represents the attention mechanism formula; the decoder output layer adopts linear transformation and softmax function to obtain the final prediction value: wherein, is an output weight matrix, is an output bias, is a decoder output; in the model training process, an improved loss function is adopted, combining mean square error and prediction bias penalty term: wherein is a prediction value, is an actual value, is a penalty coefficient, denotes the mean square error, and the model parameters are adjusted by a back propagation algorithm.

[0112] S3. Load prediction: inputting the current load data of the solid state disk collected in real time into the trained load prediction model to obtain a load prediction result in a future preset time period;

[0113] S4. Power consumption optimization: according to the load prediction result, a multi-objective optimization dynamic power consumption adjustment algorithm is used to optimize the power consumption of the solid state disk, and the dynamic power consumption adjustment algorithm includes a voltage adjustment sub-algorithm, a frequency adjustment sub-algorithm, a storage unit activation quantity adjustment sub-algorithm and a cache strategy adjustment sub-algorithm based on the load prediction result;

[0114] The specific steps of the voltage adjustment sub-algorithm based on the load prediction result in S4 are as follows:

[0115] S411. Based on the load prediction value and the performance constraint, a target voltage value is calculated by a voltage optimization formula, and the voltage optimization formula is: wherein, is the load prediction value, is the performance constraint, is a reference voltage, is a voltage adjustment coefficient, is a performance compensation coefficient, is a current performance.

[0116] S412. The working voltage of the solid state disk is adjusted to the target voltage value according to a preset voltage adjustment step, so that the voltage change rate does not exceed a voltage change rate threshold, and the adjustment formula is: wherein, is a preset voltage adjustment step, is a voltage change rate threshold, and are the voltages at and moments, respectively, is a sign function.

[0117] The specific steps of the frequency adjustment sub-algorithm based on the load prediction result in S4 are as follows:

[0118] S421. According to the load prediction result, a target working frequency is calculated by a frequency optimization formula, and the frequency optimization formula is: wherein, is the load prediction result, is the target working frequency, is a reference frequency, is a reference load, is a frequency adjustment coefficient, Energy consumption coefficient, Current power consumption, Optimal power consumption;

[0119] S422. The current working frequency of the solid state disk is adjusted to the target working frequency in a smooth transition manner, and the frequency change formula is: , wherein, is a smooth transition coefficient, and are the frequencies at and moments, respectively.

[0120] The specific steps of the storage unit activation quantity adjustment sub-algorithm in S4 based on the load prediction result are as follows: , wherein, is the number of storage units, is the preset time, is the processing capacity of a single storage unit, is a redundancy coefficient;

[0121] S432. Activate the corresponding number of storage units and close the redundant storage units.

[0122] The specific steps of the cache strategy adjustment sub-algorithm in S4 are as follows:

[0123] S441. Determine the data access hotspots according to the load prediction result;

[0124] S442. Based on the data access hotspots and the cache capacity, calculate the cache allocation ratio of different data through the cache allocation formula, which is: , wherein: is the cache allocation ratio of the th data, is the access probability of the th data, is the sum of all data access probabilities, is the cache capacity;

[0125] S443. Adjust the cache strategy according to the calculated cache allocation ratio, and preferentially cache high access probability data;

[0126] S5. Feedback adjustment: Real-time monitoring of the actual power consumption and performance parameters of the optimized solid state disk, and adjusting the load prediction model and dynamic power consumption adjustment algorithm in combination with the reinforcement learning algorithm.

[0127] The specific steps of the feedback adjustment in S5 are as follows:

[0128] S51. Construct a real-time monitoring parameter system;

[0129] The real-time monitoring parameter system adopts a three-layer monitoring architecture, collecting core parameters every 5ms, specifically including:

[0130] Power consumption parameters: core chip power supply voltage, operating current, and flash memory array power consumption percentage;

[0131] Performance parameters: random read / write IOPS, sequential read / write bandwidth, average response time, read / write latency jitter;

[0132] Prediction deviation parameter: the instantaneous error between the predicted and actual load values, i.e. ; the mean error of the sliding window, i.e. ,in, express The instantaneous error between the predicted and actual load values ​​at any given time. To show Forecast load value at any time for Actual load value at all times Represents the mean error of the sliding window, and the window size. That is, for the current time Instantaneous error of the first 20 moments, including Take the average value;

[0133] All parameters are transmitted to the monitoring database via an encrypted communication channel and stored in a time-series database, retaining high-frequency data from the most recent 72 hours and minute-level summary data for 90 days.

[0134] S52. Deployment of reinforcement learning algorithms;

[0135] Deep Q-Network (DQN) is used as the core framework for reinforcement learning, specifically as follows:

[0136] State space definition: ,in, express The state vector at any given time integrates multiple key parameters related to the power consumption and operating status of the solid-state drive. The power consumption at the current moment. The reference power consumption is a preset baseline value used to measure the power consumption at the current moment. Normalization is performed. This is the historical average load. This is the frequency adjustment amount. For voltage regulation, For current performance, Due to performance constraints, As the reference frequency, For the reference voltage;

[0137] Action space design: discretize the action set The strategy corresponding to each action is:

[0138] Load prediction model parameter adjustment: dynamically adjust the learning rate, support ±10% or ±20% gradient update amplitude adjustment, optimize model training efficiency;

[0139] Dynamic power consumption algorithm parameter calibration: adjust key coefficients, support ±5% or ±10% parameter floating, realize dynamic balance of power consumption and performance;

[0140] Reward function optimization: , , For the weight coefficient;

[0141] Experience replay mechanism: maintain an experience pool with a capacity of 10000, randomly extract 32 samples for training each iteration, use - Greedy strategy, and From 0.9 linearly decay to 0.1;

[0142] S53. Dynamic adjustment of load prediction model;

[0143] When the continuous 5 monitoring periods, i.e. 25ms Trigger model fine-tuning:

[0144] Attention weight matrix The update formula is: , The learning rate of reinforcement learning output, Indicates the loss function The gradient of the attention weight matrix ;

[0145] LSTM gating parameter correction: increase the forgetting gate bias To: ;

[0146] S54. Adaptive adjustment of dynamic power consumption algorithm;

[0147] Voltage adjustment sub-algorithm correction:

[0148] Real-time calculation of voltage adjustment error , The actual voltage is The target voltage is when , Where, The voltage adjustment coefficient is The corrected voltage adjustment coefficient is

[0149] Frequency adjustment coefficient optimization:

[0150] Based on energy-performance ratio , dynamically adjust: Wherein is the preset optimal energy-performance ratio, is the input / output operation rate of the solid state disk at moment, is the actual power consumption of the solid state disk at moment, is the corrected frequency adjustment coefficient, .

[0151] Embodiment 2

[0152] Please refer to Figure 2 , the solid state disk power consumption optimization system based on load prediction, the system comprises:

[0153] Data acquisition module 1: for collecting historical load data and corresponding power consumption data of solid state disk, and collecting current load data, optimized actual power consumption and performance parameters of solid state disk in real time;

[0154] For historical load data and power consumption data, collect them once every 10ms, and store them in local cache, and every hour, the cache data is summarized to the database for long-term preservation. In the collection process, real-time preprocessing is carried out on the data, including outlier detection and correction, 3σ criterion is used to identify outliers, and for data exceeding the mean value ± 3 times standard deviation, linear interpolation is used for correction. For real-time load data, actual power consumption and performance parameters, the collection frequency is increased to once every 5ms to ensure the timeliness of the data, and accurate data support is provided for subsequent load prediction and feedback adjustment;

[0155] Load prediction module 2: for constructing a bidirectional LSTM-Transformer load prediction model with fusion attention mechanism based on the historical load data collected by the data acquisition module 1, and training, also for inputting the real-time collected current load data into the trained model to get the load prediction result, wherein the bidirectional LSTM-Transformer load prediction model, first, the input historical load data is processed by feature standardization, and the data is mapped to the [0, 1] interval;

[0156] Power consumption optimization module 3: for optimizing the power consumption of the solid state disk by using the multi-objective optimization dynamic power consumption adjustment algorithm according to the load prediction result obtained by the load prediction module 2;

[0157] Feedback adjustment module 4: used to adjust the load prediction model in load prediction module 2 and the dynamic power consumption adjustment algorithm in power consumption optimization module 3 according to the optimized actual power consumption and performance parameters collected by data collection module 1, combined with the reinforcement learning algorithm.

[0158] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0159] The above description is only the preferred embodiment of the present application, and it should be pointed out that for those skilled in the art, without departing from the technical principles of the present application, a number of improvements and refinements can be made, which should be considered as the protection scope of the present application.

Claims

1. A method for optimizing power consumption of a solid state drive based on load prediction, characterized in that, The method comprises the following steps: S1. Data collection: collecting historical load data and corresponding power consumption data of the solid state disk, the historical load data comprising I / O request quantity, data transmission quantity, request response time, request type; S2. Load prediction model construction and training: based on the collected historical load data, a bidirectional LSTM-Transformer load prediction model fusing an attention mechanism is constructed, and the model is trained using the historical load data to obtain a trained load prediction model; S3. Load prediction: inputting real-time collected current load data of the solid state disk into the trained load prediction model to obtain a load prediction result in a future preset time period; S4. Power consumption optimization: based on the load prediction result, a multi-objective optimization dynamic power consumption adjustment algorithm is adopted to optimize the power consumption of the solid state disk, the dynamic power consumption adjustment algorithm comprising a voltage adjustment sub-algorithm, a frequency adjustment sub-algorithm, a storage unit activation quantity adjustment sub-algorithm and a cache strategy adjustment sub-algorithm based on the load prediction result; S5. Feedback adjustment: real-time monitoring of actual power consumption and performance parameters of the optimized solid state disk, and adjustment of the load prediction model and the dynamic power consumption adjustment algorithm in combination with a reinforcement learning algorithm; The specific steps of the feedback adjustment in S5 are as follows: S51. Construction of a real-time monitoring parameter system; The real-time monitoring parameter system adopts a three-layer monitoring framework, and core parameters are collected every 5 ms, specifically comprising: Power consumption parameters: core chip power supply voltage, working current, flash array power consumption proportion; Performance parameters: random read / write IOPS, continuous read / write bandwidth, average response time, read / write delay jitter value; Prediction deviation parameter: the instantaneous error between the predicted and actual load values, i.e. ; the mean error of the sliding window, i.e. ,in, express The instantaneous error between the predicted and actual load values ​​at any given time. To show Forecast load value at any time for Actual load value at all times Represents the mean error of the sliding window, and the window size. That is, for the current time Instantaneous error of the first 20 moments, including Take the average value; All parameters are transmitted to a monitoring database through an encrypted channel and stored in a time series database, retaining high-frequency data for the last 72 hours and minute-level summary data for 90 days; S52. Deployment of a reinforcement learning algorithm; A deep Q network (DQN) is used as the core framework of the reinforcement learning, specifically comprising: State space definition: ,in, express The state vector at any given time integrates multiple key parameters related to the power consumption and operating status of the solid-state drive. The power consumption at the current moment. The reference power consumption is a preset baseline value used to measure the power consumption at the current moment. Normalization is performed. This is the historical average load. This is the frequency adjustment amount. For voltage regulation, For current performance, Due to performance constraints, As the reference frequency, The reference voltage; Action space design: Discretized action set The strategy corresponding to each action is: Load prediction model parameter adjustment: dynamically adjusting a learning rate, supporting a gradient update amplitude adjustment of ±10% or ±20% to optimize model training efficiency; Dynamic power consumption algorithm parameter calibration: adjusting key coefficients, supporting a parameter floating of ±5% or ±10% to achieve dynamic balance between power consumption and performance; Reward function optimization: , , are weight coefficients; Experience replay mechanism: maintain an experience pool with capacity of 10000, randomly sample 32 samples for training in each iteration, use - a greedy policy, and linearly decay from 0.9 to 0.1; S53. Dynamic adjustment of the load prediction model; When 5 consecutive monitoring periods, i.e. 25 ms Model fine-tuning is triggered: attention weight matrix The update formula is: , is the learning rate of reinforcement learning output, represents the loss function The gradient of the attention weight matrix ; LSTM gating parameter correction: forget gate bias increase to: ; S54. Adaptive adjustment of the dynamic power consumption algorithm; Voltage adjustment sub-algorithm correction: Real-time calculation of voltage regulation error , is the actual voltage, is the target voltage, when , wherein, is the voltage regulation coefficient, is the corrected voltage regulation coefficient; Frequency adjustment coefficient optimization: Based on energy consumption-performance ratio Dynamic adjustment: ,in To preset the optimal energy consumption-performance ratio, In order to be in The input / output operation speed of a solid-state drive at all times. In order to be in The actual power consumption of the solid-state drive at all times. This is the corrected frequency adjustment coefficient. .

2. The method of claim 1, wherein the method further comprises: The specific process of constructing the bidirectional LSTM-Transformer load prediction model fusing an attention mechanism in S2 is as follows: S21. Feature extraction of historical load data is performed using a bidirectional LSTM network to capture bidirectional time sequence dependency of the data; S22. The features output by the bidirectional LSTM network are input into an encoder of a Transformer model, and the features are weighted processed through an attention mechanism to highlight important features; S23. The features output by the encoder are input into a decoder of the Transformer model to obtain a load prediction result.

3. The method of claim 2, wherein: The forward propagation and backward propagation formulas of the bidirectional LSTM network are as follows: Forward propagation formula: ， ， ， ， ； The back propagation formula is: ， ， ， ， ； in, ... Forward propagation Input gate, forget gate, and output gate at any time. For positive propagation Cellular state at any given moment This represents the hidden state from the previous moment in the forward propagation. The hyperbolic tangent activation function is used. , , , The inputs during forward propagation are respectively The weight matrices for the input gate, forget gate, cell state update, and output gate; , , , These are the hidden states from the previous time step during forward propagation. The weight matrices for the input gate, forget gate, cell state update, and output gate; , , , These are the bias terms for the input gate, forget gate, cell state update, and output gate in forward propagation, respectively. , , In reverse propagation Input gate, forget gate, and output gate at any time. For backpropagation Cellular state at any given moment This is the hidden state for the next time step in backpropagation. , , , The inputs in the backpropagation are respectively The weight matrices for the input gate, forget gate, cell state update, and output gate; , , , These are the hidden states in the next time step during backpropagation. The weight matrices for the input gate, forget gate, cell state update, and output gate. , , , These are the bias terms for the input gate, forget gate, cell state update, and output gate in backpropagation, respectively, indicated by superscript. Indicates positive, superscript Indicates the opposite direction. For the sigmoid function, For element-wise multiplication, For a moment Input data; bidirectional LSTM output features ; The attention mechanism formula is: Wherein, Query matrix, key matrix, value matrix, Feature dimension, Feature importance weight matrix, Matrix transpose symbol, Indicates that the key matrix is transposed, Represent the attention mechanism formula; the decoder output layer adopts linear transformation and softmax function to obtain the final prediction value: Wherein, Output weight matrix, Output bias, Decoder output; in the model training process, an improved loss function is used, combining mean square error and prediction bias penalty term: Wherein Prediction value, Actual value, Penalty coefficient, Indicates the mean square error, and the model parameters are adjusted by the back propagation algorithm.

4. The method of claim 3, wherein: The specific steps of the voltage adjustment sub-algorithm based on the load prediction result in S4 are: S411. Calculate the target voltage value based on the load prediction value and the performance constraint through a voltage optimization formula, wherein the voltage optimization formula is: wherein, is the load prediction value, is the performance constraint, is the reference voltage, is the voltage adjustment coefficient, is the performance compensation coefficient, is the current performance; S412. Adjust the working voltage of the solid state disk to the target voltage value according to the preset voltage adjustment step, so that the voltage change rate does not exceed the voltage change rate threshold, and the adjustment formula is: wherein, is the preset voltage adjustment step, is the voltage change rate threshold, and are the voltages at the time points of and respectively, is a sign function.

5. The method of claim 4, wherein: The specific steps of the frequency adjustment sub-algorithm based on the load prediction result in S4 are: S421. According to the load prediction result, the target working frequency is calculated by a frequency optimization formula, the frequency optimization formula is: wherein, is the load prediction result, is the target working frequency, is the reference frequency, is the reference load, is the frequency adjustment coefficient, is the energy consumption coefficient, is the current power consumption, is the optimal power consumption; S422. Adopting a smooth transition mode to adjust the current working frequency of the solid state disk to the target working frequency, and the frequency change formula is: wherein, is a smooth transition coefficient, and are the frequencies at the time of and respectively.

6. The method of claim 5, wherein: The specific steps of the storage unit activation quantity adjustment sub-algorithm based on the load prediction result in S4 are: S431. According to the load prediction result and the performance parameter of the storage unit, the required activated storage unit quantity is calculated through a storage unit quantity calculation formula, and the storage unit quantity calculation formula is: wherein, is the storage unit quantity, is the preset time, is the processing capacity of a single storage unit, is the redundancy coefficient; S432. Activate the corresponding number of storage units, and close the redundant storage units.

7. The method of claim 6, wherein: The specific steps of the cache strategy adjustment sub-algorithm in S4 are: S441. Determine the data access hot spot according to the load prediction result; S442. Based on data access hotspots and cache capacity, calculate the cache allocation ratio for different types of data using the cache allocation formula: ,in: For the first The caching allocation ratio of this type of data For the first The probability of accessing this type of data. The sum of the probabilities of accessing all data. This refers to the cache capacity. S443. Adjust the cache strategy according to the calculated cache allocation ratio, and preferentially cache high access probability data.

8. The system for solid state drive power consumption optimization based on load prediction of any of claims 1-7, wherein, The system comprises: A data acquisition module (1) for acquiring historical load data and corresponding power consumption data of a solid state disk, and acquiring current load data, optimized actual power consumption and performance parameters of the solid state disk in real time; a load prediction module (2) for constructing a bidirectional LSTM-Transformer load prediction model fused with an attention mechanism based on the historical load data acquired by the data acquisition module (1), and training the model, and for inputting the current load data acquired in real time into the trained model to obtain a load prediction result; a power consumption optimization module (3) for optimizing the power consumption of the solid state disk by using a multi-objective optimization dynamic power consumption adjustment algorithm according to the load prediction result obtained by the load prediction module (2); and a feedback adjustment module (4) for adjusting the load prediction model in the load prediction module (2) and the dynamic power consumption adjustment algorithm in the power consumption optimization module according to the optimized actual power consumption and performance parameters acquired by the data acquisition module (1) and in combination with a reinforcement learning algorithm. ​ ​

Images