SSD (Solid State Disk) heat dissipation optimization method and system based on dynamic temperature control

Abstract

The invention relates to an SSD (Solid State Disk) heat dissipation optimization method and system based on dynamic temperature control. The method comprises the following steps: acquiring a thermal state vector of an SSD through inner and outer ring temperature monitoring; predicting a future 5S temperature track through a long short-term memory network model, and generating a temperature prediction sequence; generating a computing power distribution proportion based on the temperature prediction sequence, and determining a capacity migration proportion at the same time; according to the real-time temperature value of each logic block in the block temperature mapping table and the computing power distribution proportion and capacity migration proportion, heat-sensitive and heat-tolerant classification is carried out on the I / O queue, and an optimized scheduling sequence is generated; calculating a power supply phase parameter according to the access frequency of the heat-sensitive block of the optimized scheduling sequence, and generating a pulse type heat dissipation control signal; the SSD power supply phase is dynamically adjusted according to the pulse type heat dissipation control signal, heat dissipation optimization is conducted through the pulse width modulation technology, updated heat state data is fed back to inner and outer ring temperature monitoring, and therefore the purpose of dynamic intelligent temperature control of the SSD is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of intersection of computer storage and heat dissipation control, and in particular to a method and system for optimizing heat dissipation of an SSD solid state drive based on dynamic temperature control. Background Art

[0002] The explosive growth in data storage demand in data centers and edge computing is placing higher demands on the performance of storage devices. Storage controller firmware design, at the heart of SSD performance and reliability, directly determines the stability and lifespan of data storage systems under high loads. Storage devices must not only meet high throughput and low latency performance targets but also address complex thermal management challenges. Effective thermal management strategies not only ensure stable device performance but are also key to extending hardware life and reducing energy consumption, significantly improving the overall performance of data storage systems.

[0003] However, existing solutions for SSD thermal management suffer from significant shortcomings. Traditional thermal management methods mostly rely on temperature monitoring mechanisms with a single time granularity. This approach cannot simultaneously address both the transient thermal shocks generated by flash memory operations and the dynamic demands of the device's overall thermal balance, significantly limiting the flexibility and adaptability of cooling strategies. Furthermore, existing scheduling mechanisms and power adjustment strategies lack the ability to precisely perceive temperature distribution, making it difficult to accurately adapt to diverse workloads. This results in chronically low cooling efficiency. These technical issues make storage controllers highly susceptible to thermal bottlenecks under high load, severely impacting device performance and reliability. Summary of the Invention

[0004] The main purpose of the present invention is to provide an SSD solid state drive heat dissipation optimization method and system based on dynamic temperature control. Through real-time prediction, precise scheduling and closed-loop feedback, the heat dissipation efficiency and equipment reliability are improved while ensuring performance, so as to achieve the purpose of dynamic intelligent temperature control of the SSD solid state drive.

[0005] To achieve the above objectives, the present invention provides a method for optimizing heat dissipation of an SSD solid state drive based on dynamic temperature control, comprising the following steps: Obtain the thermal state vector of the SSD hard disk by monitoring the internal and external ring temperatures. The thermal state vector includes the thermal shock peak value and the overall average temperature of the SSD; Input the thermal state vector into a pre-trained long short-term memory network model to predict the temperature trajectory in the next 5 seconds and generate a temperature prediction sequence; Based on the temperature prediction sequence, a computing power allocation ratio is generated, and a capacity migration ratio is determined; Based on the real-time temperature values of each logical block in the block temperature mapping table, as well as the computing power allocation ratio and capacity migration ratio, I / O queues are classified into heat-sensitive and heat-tolerant types, and an optimized scheduling sequence is generated. Calculating power supply phase parameters based on the access frequency of the heat-sensitive blocks in the optimized scheduling sequence and generating a pulsed heat dissipation control signal; The SSD power supply phase is dynamically adjusted according to the pulsed heat dissipation control signal, heat dissipation is optimized through pulse width modulation technology, and updated thermal status data is fed back to the inner and outer ring temperature monitoring.

[0006] Furthermore, the step of obtaining the thermal state vector of the SSD hard disk by monitoring the internal and external ring temperatures includes: Through the inner loop monitoring mechanism, 10ms granularity thermal shock data is obtained from the SSD flash memory controller, and the temperature change values of flash memory page programming and erasing operations are collected to generate a real-time thermal shock sequence; extracting thermal shock peak values and frequency component characteristic values from the real-time thermal shock sequence; The outer loop monitoring mechanism obtains SSD overall temperature data with 1-second granularity and calculates the overall average temperature of the SSD. The thermal shock peak value and the overall average temperature are weightedly fused and normalized to generate a standardized thermal state vector.

[0007] Furthermore, before inputting the thermal state vector into the pre-trained long short-term memory network model, a model triggering condition judgment step is also included, including: Calculating a ratio K of a thermal shock peak value to an overall average temperature in the thermal state vector; When the composite trigger condition is met, K ≥ thermal shock coefficient threshold and the overall average temperature ≥ basic temperature threshold, the long short-term memory network model is activated, and the thermal shock coefficient threshold and the basic temperature threshold are pre-set according to the SSD model; If the composite trigger condition is not met, the temperature prediction sequence generated in the previous monitoring cycle is used for heat dissipation control.

[0008] Furthermore, the step of inputting the thermal state vector into a pre-trained long short-term memory network model to predict the temperature trajectory in the next 5 seconds and generate a temperature prediction sequence includes: Inputting the thermal state vector into a pre-trained long short-term memory network model; Based on the temporal correlation between the thermal shock peak value and the overall average temperature, predict the temperature change trajectory in the next 5 seconds; Outputs a temperature forecast sequence containing timestamps and temperature values.

[0009] Furthermore, the steps for constructing the long short-term memory network model include: Building a training data set based on SSD historical operation data, the data set includes thermal shock peak sequence data, average temperature sequence data and temperature change trajectory data; Design an LSTM network with a bidirectional gating structure. The number of nodes in its input layer is consistent with the dimension of the thermal state vector, and the output layer is the temperature prediction value per second in the next 5 seconds. The sliding window mechanism is used to segment the training data into time series, with a window length of 10 monitoring cycles and a sliding step of 1 cycle; The loss function is set as the root mean square error between the predicted temperature trajectory and the actual temperature value, and the weight penalty term of the thermal shock peak is superimposed.

[0010] Furthermore, based on the temperature prediction sequence, the step of generating a computing power allocation ratio and determining a capacity migration ratio includes: According to the block temperature mapping table stored in the SSD, the temperature prediction sequence is trend-fitted, and the computing power distribution ratio between the host CPU thread and the SSD master control is calculated using a dynamic programming algorithm; According to the peak distribution of the temperature prediction sequence, a greedy algorithm is used to determine the capacity migration ratio of the non-volatile write cache and the DRAM cache.

[0011] Furthermore, the steps of classifying I / O queues into heat-sensitive and heat-tolerant types and generating an optimized scheduling sequence include: Based on the real-time temperature values of each logical block in the block temperature mapping table, I / O requests of logical blocks with associated temperature values ≥ a preset temperature threshold are classified as heat-sensitive queues; Based on the capacity migration ratio, the non-volatile write cache allocation weight corresponding to the heat-sensitive queue is reduced, and the DRAM cache allocation weight is increased; Re-ordering the priorities of the I / O requests in the heat-sensitive queue from low to high according to the temperature values of their associated logical blocks; Allocate additional SSD master computing resources to heat-sensitive queues based on the computing power allocation ratio; Generate an optimized scheduling sequence including reordering results and resource allocation parameters.

[0012] Furthermore, the step of calculating power supply phase parameters and generating a pulsed heat dissipation control signal based on the access frequency of the heat-sensitive blocks in the optimized scheduling sequence includes: Counting the access frequency of heat-sensitive blocks in the optimized scheduling sequence; The thermal resistance model is called by the UEFI runtime service to calculate the power phase adjustment parameter that is positively correlated with the access frequency; A pulse width modulation signal is generated according to the power supply phase adjustment parameter.

[0013] Furthermore, the steps of dynamically adjusting the SSD power supply phase according to the pulsed heat dissipation control signal, optimizing heat dissipation through pulse width modulation technology, and feeding back updated thermal status data to the inner and outer ring temperature monitoring include: Converting the pulsed heat dissipation control signal into a duty cycle parameter of the SSD power supply phase through a pulse width modulator; Collect temperature sensor data from the SSD chipset in real time and update the block temperature mapping table; The updated thermal state vector is re-input into the inner and outer loop temperature monitoring to complete the dynamic heat dissipation optimization closed loop.

[0014] The present invention also provides an SSD solid state drive heat dissipation optimization system based on dynamic temperature control, comprising: A thermal status monitoring unit is used to obtain a thermal status vector of the SSD hard disk by monitoring the internal and external ring temperatures. The thermal status vector includes a thermal shock peak value and an average temperature of the entire SSD; A temperature prediction unit is used to input the thermal state vector into a pre-trained long short-term memory network model, predict the temperature trajectory in the next 5 seconds, and generate a temperature prediction sequence; a resource allocation unit, configured to generate a computing power allocation ratio based on the temperature prediction sequence and determine a capacity migration ratio; The scheduling optimization unit is used to classify I / O queues into heat-sensitive and heat-tolerant types based on the real-time temperature values of each logical block in the block temperature mapping table and the computing power allocation ratio and capacity migration ratio, and generate an optimized scheduling sequence; A signal generating unit, configured to calculate a power supply phase parameter and generate a pulsed heat dissipation control signal according to the access frequency of the heat-sensitive block in the optimized scheduling sequence; The heat dissipation unit is used to dynamically adjust the SSD power supply phase according to the pulse heat dissipation control signal, optimize heat dissipation through pulse width modulation technology, and feed back updated thermal status data to the inner and outer ring temperature monitoring.

[0015] The SSD solid-state hard disk heat dissipation optimization method and system based on dynamic temperature control provided by the present invention have the following beneficial effects: the present invention predicts the temperature trajectory through the LSTM model, combines the identification of heat-sensitive blocks, and actively adjusts the power supply strategy and I / O scheduling before the temperature exceeds the standard, converting traditional passive heat dissipation into pre-intervention, significantly reducing the risk of thermal runaway. In addition, based on real-time temperature data, computing power, cache and power supply resources are dynamically allocated to adaptively match heat dissipation control with performance requirements, which not only avoids energy waste caused by excessive heat dissipation, but also alleviates the inherent contradiction between performance and temperature. A complete closed loop is formed from temperature perception, prediction decision-making to execution feedback, which can realize self-iterative optimization of heat dissipation strategy, and is particularly suitable for long-term stable operation in complex environments such as data centers. The technical solution of the present invention breaks through the rigid constraints of traditional static heat dissipation solutions, gives SSDs more intelligent thermal management capabilities, and achieves systematic improvements in dimensions such as reliability, energy efficiency and lifespan. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a method for optimizing heat dissipation of an SSD solid-state hard disk based on dynamic temperature control according to an embodiment of the present invention; Figure 2 This is a structural block diagram of an SSD solid state drive heat dissipation optimization system based on dynamic temperature control in one embodiment of the present invention.

[0017] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0019] Reference Figure 1 , which is a flow chart of a method for optimizing SSD heat dissipation based on dynamic temperature control proposed by the present invention, comprising the following steps: S1, obtain the thermal state vector of the SSD hard disk through internal and external ring temperature monitoring, the thermal state vector includes the thermal shock peak value and the overall average temperature of the SSD; S2, inputting the thermal state vector into a pre-trained long short-term memory network model to predict the temperature trajectory in the next 5 seconds and generate a temperature prediction sequence; S3, generating a computing power allocation ratio based on the temperature prediction sequence and determining a capacity migration ratio; S4: Based on the real-time temperature values of each logical block in the block temperature mapping table and the computing power allocation ratio and capacity migration ratio, the I / O queues are classified into heat-sensitive and heat-tolerant types, and an optimized scheduling sequence is generated; S5, calculating power supply phase parameters according to the access frequency of the heat-sensitive blocks in the optimized scheduling sequence, and generating a pulsed heat dissipation control signal; S6, dynamically adjusting the SSD power supply phase according to the pulsed heat dissipation control signal, optimizing heat dissipation through pulse width modulation technology, and feeding back updated thermal status data to the inner and outer ring temperature monitoring.

[0020] In one embodiment, for step S1, The steps for obtaining the SSD's thermal status vector by monitoring the internal and external ring temperatures include: Through the inner loop monitoring mechanism, 10ms granularity thermal shock data is obtained from the SSD flash memory controller, and the temperature change values of flash memory page programming and erasing operations are collected to generate a real-time thermal shock sequence; extracting thermal shock peak values and frequency component characteristic values from the real-time thermal shock sequence; The outer loop monitoring mechanism obtains SSD overall temperature data with 1-second granularity and calculates the overall average temperature of the SSD. The thermal shock peak value and the overall average temperature are weightedly fused and normalized to generate a standardized thermal state vector.

[0021] In the specific implementation process, the thermal state vector is acquired through the collaborative monitoring architecture of the inner and outer loops. At the hardware level, the inner loop monitoring embeds a high-precision digital temperature sensor (such as TS3001B) in the NAND flash memory chip to collect the instantaneous temperature during programming / erase operations with a period of 10ms. The data is transmitted to the main control via the low-latency I3C bus; the outer loop monitoring integrates the built-in sensor of the main control chip (such as MAX1617) and the 4-point thermocouple array distributed on the outer shell, and calculates the weighted average temperature with a period of 1s (the main control weight is 60%, the NAND array is 30%, and the interface circuit is 10%). The inner loop data is processed in real time by the FPGA coprocessor, and a dual-threshold peak detection algorithm (absolute threshold ≥85℃, temperature rise rate ≥5℃ / ms) is used to identify thermal shock events. The optimized 16-point FFT is used to extract the 0.1-10Hz frequency band features (such as the 2-5Hz component caused by burst writes). The outer loop data and the inner loop features are fused through dynamic weights to generate the thermal state vector: ,in is the overall average temperature of the outer ring, is the peak value of the inner loop thermal shock. The weight coefficients (α, β) are dynamically adjusted according to the temperature change rate (α = 0.7, β = 0.3 when ΔT / Δt < 2°C / s in steady-state mode; α = 0.4, β = 0.6 when ΔT / Δt ≥ 2°C / s in transient mode). To eliminate differences in SSD models, dynamic range normalization is further used: in Indicates the actual measured temperature value (unit: °C). is the lower limit temperature (fixed at 25°C, corresponding to the SSD idle state), Set by product type (consumer-grade 125℃, enterprise-grade 105℃). For example, an enterprise-grade SSD detects hour, , quantitatively characterizing the degree to which the current temperature is approaching the upper limit. Data is aligned using hardware timestamps (inner loop 10ms accuracy) and software timestamps (outer loop 1ms accuracy), and sliding window correlation analysis is used to compensate for a maximum 5ms delay for timing synchronization. Data reliability is ensured through three levels of verification: CRC verification of raw data, temperature rise rate rationality check (ΔT ≤ 20°C / s), and verification of the temperature difference between adjacent sensors (≤ 15°C). To reduce power consumption, an adaptive sampling rate control module is designed. When the temperature is stable (ΔT < 0.5°C for 10 consecutive cycles), the inner loop sampling rate is reduced to 50ms. Energy efficiency optimization is achieved in conjunction with a 256KB DRAM circular buffer.

[0022] In one embodiment, before inputting the thermal state vector into the pre-trained long short-term memory network model, a model trigger condition determination step is further included, including: Calculating a ratio K of a thermal shock peak value to an overall average temperature in the thermal state vector; When the composite trigger condition is met, K ≥ thermal shock coefficient threshold and the overall average temperature ≥ basic temperature threshold, the long short-term memory network model is activated, and the thermal shock coefficient threshold and the basic temperature threshold are pre-set according to the SSD model; If the composite trigger condition is not met, the temperature prediction sequence generated in the previous monitoring cycle is used for heat dissipation control.

[0023] Specifically, in the implementation of the model triggering conditions, the system dynamically controls the activation of the LSTM prediction model through a dual-threshold mechanism to balance computational overhead and heat dissipation accuracy. In the specific implementation, the thermal shock characteristic ratio of the thermal state vector is calculated: ,in Thermal shock peak temperature extracted for inner ring monitoring, The overall average temperature calculated for the outer loop. This ratio quantifies the relative relationship between the transient thermal shock intensity and the steady-state temperature (for example, K = 1.2 means that the peak temperature is 20% higher than the average temperature). The trigger condition is a logical AND combination: ,in is the thermal shock coefficient threshold, The temperature threshold is the basic temperature threshold. The two are pre-set according to the thermal design specifications of the SSD model. For consumer-grade SSDs (low reliability scenarios), , enterprise-level SSD (high reliability requirements) Double constraints are used to avoid false triggering of a single indicator. For example, when a sudden write to the SSD causes A short-term rise but When the threshold is not reached (e.g. ), only the K condition is met but the overall temperature does not exceed the standard. At this time, the prediction model is not activated to prevent excessive response to transient fluctuations. On the contrary, if the overall temperature is close to the critical value (such as ), even if K does not reach the threshold (such as K=1.1), other protection mechanisms (such as frequency reduction) may still be triggered, but the LSTM model remains dormant to reduce the computing load. When any condition is not met, the system reuses the temperature prediction sequence generated in the previous monitoring cycle (usually 1 second) and performs trend extrapolation through the sliding window algorithm. For example, the prediction sequence of the previous cycle is , when the model is not currently triggered, update the sequence to , where ΔT is calculated based on the historical temperature change rate (usually the average change rate in the last 3 seconds). The threshold parameters are designed as configurable items in the firmware and can be dynamically adjusted through the UEFI interface or manufacturer tools. For example, in a high temperature environment, the user can Lower the temperature by 5℃ to trigger the prediction in advance; while in the cold storage nodes of the data center, the temperature can be raised This mechanism reduces the theoretical maximum response delay from 200ms for a fixed model to 80ms (trigger condition determination takes only 0.5μs), while ensuring that the CPU utilization of the SSD controller does not increase by more than 3%.

[0024] In the specific implementation, an enterprise-level SSD is monitored under continuous load. but Due to the default , while satisfying , immediately activate the LSTM model to predict the temperature for 5 seconds. If the next cycle Drop to 58℃, even if K=1.25, because it does not meet If the conditions are met, the system will use historical forecast data and extrapolate updates to avoid unnecessary model inference consumption.

[0025] In one embodiment, for step S2, The steps of inputting the thermal state vector into a pre-trained long short-term memory network model, predicting the temperature trajectory in the next 5 seconds, and generating a temperature prediction sequence include: Inputting the thermal state vector into a pre-trained long short-term memory network model; Based on the temporal correlation between the thermal shock peak value and the overall average temperature, predict the temperature change trajectory in the next 5 seconds; Outputs a temperature forecast sequence containing timestamps and temperature values.

[0026] In the specific implementation process, an LSTM architecture with coupled spatiotemporal features is used to achieve accurate short-term temperature trajectory prediction. The thermal state vector is expanded into a time series sequence. Each input sample contains standardized data for 10 consecutive monitoring cycles (10×100ms=1s), with the following structure: ,in is the normalized temperature value, The thermal shock peak flag (0 / 1). A time decay factor λ = 0.95 is introduced in the input layer to give more weight to recent data: A variant of the Seq2Seq structure is used to directly output the temperature value every second for the next 5 seconds to avoid the error accumulation of iterative predictions. The real-time ambient temperature is introduced as a bias term in the output layer: , where γ=0.15 is the environmental sensitivity coefficient, 25℃ is the laboratory reference temperature, is the real-time ambient temperature. The 5 predicted values output by the model ( ) synchronized with the hardware clock to generate an absolute timestamp: , calculate the fluctuation range of the predicted value: , where N is the number of historical prediction samples in the sliding window (usually N=10). If the jump between adjacent prediction values exceeds 10°C, the smoothing process based on Kalman filtering is started. In addition, the real-time optimization technology improves the system response speed through multi-dimensional strategies: in terms of model lightweighting, the channel pruning technology is used to greatly compress the model parameters from the original 3.2M to 780K, and the inference time is sharply reduced from 15ms to 3.8ms; in hardware acceleration, the NPU unit of the main control chip is used to perform INT8 precision quantization inference, effectively reducing memory usage by 62%; at the cache prediction level, the latest 3 sets of prediction sequences are pre-stored in the DDR cache, and interpolation reuse is realized when the model is not triggered, so as to achieve efficient real-time data processing and response. Set a fail-safe mechanism to compare the prediction sequence with the real-time monitoring data. When the error of 3 consecutive points exceeds 5°C, the model hot update is triggered; set the gradient threshold in back propagation , to prevent sudden noise from causing the model to diverge; if any value in the prediction sequence exceeds (such as 100°C for enterprise-class SSDs), the forced cooling strategy is immediately activated.

[0027] In one embodiment, the steps of constructing a long short-term memory network model include: Building a training data set based on SSD historical operation data, the data set includes thermal shock peak sequence data, average temperature sequence data and temperature change trajectory data; Design an LSTM network with a bidirectional gating structure. The number of nodes in its input layer is consistent with the dimension of the thermal state vector, and the output layer is the temperature prediction value per second in the next 5 seconds. The sliding window mechanism is used to segment the training data into time series, with a window length of 10 monitoring cycles and a sliding step of 1 cycle; The loss function is set as the root mean square error between the predicted temperature trajectory and the actual temperature value, and the weight penalty term of the thermal shock peak is superimposed.

[0028] Specifically, multi-dimensional time series data is extracted from the SSD historical log, including thermal shock peak sequences (sudden temperature spikes recorded by the inner-loop sensor, with a sampling rate of 10ms); average temperature sequences (weighted average temperatures calculated by the outer-loop monitoring, with a sampling rate of 1s); and temperature change trajectories (actually measured temperature values over the next 5 seconds, used as supervisory labels). Data preprocessing operations include unifying data of different sampling rates to 10ms granularity through cubic spline interpolation; removing temperature data exceeding physical limits (such as noise points with ΔT / Δt > 20°C / ms); and calculating the thermal shock duration. , add temperature change acceleration The preprocessed dataset was split into training, validation, and test sets in a 7:2:1 ratio, ensuring that each set encompassed the complete thermal shock event cycle. When constructing the long-short-term memory network model, the core design employed a bidirectional gated-residual hybrid architecture to capture the spatiotemporal characteristics of temperature changes. The number of input layer nodes strictly matched the dimensions of the thermal state vector (typically 6 dimensions), receiving features including normalized temperature values, thermal shock peak flags, shock duration, and temperature acceleration. The formula is: ; The bidirectional LSTM layer processes temporal dependencies through forward and reverse dual paths (the forward propagation formula is , the reverse is , the coupled input forget gate structure is used to reduce the number of parameters by 30%, and finally the temperature prediction value for the next 5 seconds is directly output through the fully connected layer. In the data preprocessing stage, the sliding window mechanism is used to split the time series data. The window length is set to 10 monitoring cycles (10 seconds), and the sliding step is 1 second to achieve 90% overlapping sampling. When a thermal shock event is detected ( ) to focus on transient features. This strategy increases the amount of training data by ninefold while improving the model's sensitivity to unexpected events. The loss function is designed as a hybrid of root mean square error (RMS) and a thermal shock penalty. Training is performed on the NVIDIA Jetson AGX Orin embedded platform using the Nesterov momentum Adam optimizer (with an initial learning rate of 3e-4, decaying by 15% every 20 epochs), batch normalization, and a dropout rate of 0.3 to prevent overfitting. The final model achieved an RMSE of 1.31°C on the test set and a peak prediction accuracy of 86.5%. After quantization, the model's size is only 2.7MB, meeting the storage limitations of the SSD controller. During the implementation, the prediction results for the input sequence [72, 75, 80, 85 (peak), 83, 78, 76, 74, 73, 72]°C were [73.2, 74.5, 76.1, 78.3, 79.8]°C, with a standard deviation of 1.06°C from the actual values [73.0, 74.8, 76.5, 77.9, 80.1]°C. The prediction error for the thermal shock point (the 3rd second) was only 0.4°C, verifying the model's ability to capture key temperature events.

[0029] In one embodiment, for step S3, The step of generating a computing power allocation ratio and determining a capacity migration ratio based on the temperature prediction sequence includes: According to the block temperature mapping table stored in the SSD, the temperature prediction sequence is trend-fitted, and the computing power distribution ratio between the host CPU thread and the SSD master control is calculated using a dynamic programming algorithm; According to the peak distribution of the temperature prediction sequence, a greedy algorithm is used to determine the capacity migration ratio of the non-volatile write cache and the DRAM cache.

[0030] Specifically, the precise allocation of computing power and cache resources is achieved through the coordinated optimization of dynamic programming and greedy algorithms. In the dynamic programming decision of computing power allocation ratio, the next 5 seconds are divided into 5 decision stages, and the state variables are defined as ,in The remaining computing power of the SSD controller (unit: TOPS) To predict the temperature value, the action space is the computing power allocation ratio , represents the proportion of I / O processing load borne by the host CPU. The goal is to minimize the cumulative temperature cost and computing power switching penalty. Through the reverse recursive method, the optimal decision of each state is calculated from the 5th second, and finally the global optimal allocation sequence is generated. The capacity migration ratio is achieved through a greedy algorithm. First, the peak temperature is extracted from the temperature prediction sequence. and its appearance time , temperature rise slope , high temperature duration The system considers key features such as scalability, performance, and cost; then determines the migration ratio ρ based on the priorities of emergency peak avoidance, slope priority, and continuous control; finally, dynamically divides the DRAM and NVM cache areas based on ρ, reserving 30% of the DRAM cache as a safety margin for heat-sensitive data. To ensure real-time performance, a layered computing architecture is adopted. The dynamic programming algorithm runs on the host CPU to leverage multi-core parallelism, while greedy decision-making is implemented in the hardware accelerator of the SSD controller, achieving a latency of < 100μs. A decision template library of typical temperature patterns is simultaneously established, and pattern matching is used to quickly generate near-optimal solutions, reducing computational latency from 15ms to 2ms. Step S3 combines the long-term optimization of dynamic programming with the rapid response of the greedy algorithm to achieve global temperature suppression in computing power allocation and eliminate local hotspots in cache migration, forming a coordinated temperature control mechanism across multiple timescales.

[0031] In one embodiment, for step S4, The steps for classifying I / O queues into heat-sensitive and heat-tolerant types and generating an optimized scheduling sequence include: Based on the real-time temperature values of each logical block in the block temperature mapping table, I / O requests of logical blocks with associated temperature values ≥ a preset temperature threshold are classified as heat-sensitive queues; Based on the capacity migration ratio, the non-volatile write cache allocation weight corresponding to the heat-sensitive queue is reduced, and the DRAM cache allocation weight is increased; Re-ordering the priorities of the I / O requests in the heat-sensitive queue from low to high according to the temperature values of their associated logical blocks; Allocate additional SSD master computing resources to heat-sensitive queues based on the computing power allocation ratio; Generate an optimized scheduling sequence including reordering results and resource allocation parameters.

[0032] In the specific implementation process, the dynamic thermal-aware scheduling algorithm is used to achieve intelligent classification of I / O queues and optimal resource allocation. Specifically, the real-time classification of heat-sensitive queues is first performed, and the preset temperature threshold is adaptively adjusted according to the SSD model and historical temperature rise pattern. , the formula is (in is the maximum allowable temperature of the SSD, is the average temperature over the past 5 minutes, The DMA engine of the main control chip is used to refresh the temperature data of the logical block in a 10μs cycle, and the spatial interpolation algorithm is used to compensate for the blind area of the sensor to update the block temperature mapping table; a thermal perception filter is inserted in the submission queue layer of the NVMe protocol stack. When the temperature corresponding to the logical block address LBA is The request is queued into the hot queue when it is busy, otherwise it is queued into the cold queue. The classification delay is controlled within 150ns. The cache weight is dynamically adjusted. The capacity migration ratio output by the prediction model is (range 0 - 1), exponential relationship Assign cache weights; modify the address mapping table in FTL to prioritize DRAM cache pages for write requests from heat-sensitive queues, delay program operations submitted to NAND to merge write times, and the measured NAND write times are reduced by 38% and the temperature rise is reduced by 2.1°C. Re-prioritize the heat-sensitive queues. Use the temperature gradient sorting algorithm to sort the (in is the logic block temperature, (where is the original queue depth) sorts I / O requests in the heat-sensitive queue in ascending order by the temperature of the associated block. A hardware-accelerated radix sorter is used to implement parallel sorting on the FPGA, supporting 200,000 priority recalculations per second, with insertion latency < 2μs and resource usage of 15% LUTs + 8% Block RAM. Then, dynamic allocation of computing resources is implemented. Let the total computing resources be ,pass (where n is the number of requests in the heat-sensitive queue) Calculate the allocation ratio of the heat-sensitive queue; reserve 3 CPU cores (assuming an 8-core master) for the heat-sensitive queue through Linux cgroups or the hardware QoS engine, allocate NVMe SQ dedicated DMA channels, and set the NAND flash channel priority weight (hot queue: cold queue = 3:1). Generate an optimized scheduling sequence, and the generated structured scheduling instructions use a specific JSON format to improve the encoding compression rate and reduce the transmission bandwidth requirement; ensure real-time performance, synchronize the scheduling cycle with the temperature prediction (1 second benchmark), and trigger preemptive scheduling when an emergency occurs. Transmission is via a PCIe Gen4 x4 link with an end-to-end delay of < 50μs. In addition, a fault tolerance and recovery mechanism is also provided. When the block temperature is lower than When the hot-sensitive queue accounts for more than 70%, the load balancing is triggered to migrate some requests to the cold queue; Mirroring+CRC check is implemented in the DRAM cache to ensure the integrity of the migrated data.

[0033] In one embodiment, for step S5, The step of calculating power supply phase parameters and generating a pulsed heat dissipation control signal according to the access frequency of the heat-sensitive block in the optimized scheduling sequence includes: Counting the access frequency of heat-sensitive blocks in the optimized scheduling sequence; The thermal resistance model is called by the UEFI runtime service to calculate the power phase adjustment parameter that is positively correlated with the access frequency; A pulse width modulation signal is generated according to the power supply phase adjustment parameter.

[0034] Specifically, precise pulsed cooling is achieved through access frequency-power phase coupling control. To measure access frequency for heat-sensitive blocks, a hardware-level access counter is integrated into the SSD controller's DMA engine. Periodic statistics are collected for heat-sensitive blocks with a priority ≥ 0.7 in the optimized scheduling sequence, and the results are normalized to dimensionless values between 0 and 1. Temperature weights are also applied to frequently accessed blocks. A thermal resistance model and phase parameter calculations are performed. A pre-set thermal resistance model containing parameters such as thermal resistance and thermal capacitance is called via the EFI_THERMAL_PROTOCOL interface. Phase adjustment parameters are dynamically solved based on the thermal resistance network equation, and the results are constrained to the range of 0° to 90°. A phase parameter lookup table is pre-generated, mapping phase adjustments and PWM duty cycles to different access frequency ranges. Pulse-width modulation signal generation is performed, and the PWM generator of the SSD power supply PMIC is configured and relevant parameters are set. Asymmetric PWM waveforms are generated based on the phase parameters. Multi-phase interleaving control is implemented for the SSD's four power supply phases to reduce input current ripple. In addition, safety protection mechanisms are set up, including automatically increasing the phase margin when the temperature changes too quickly; real-time monitoring of bus voltage fluctuations, and resetting the phase parameters if they exceed the range; and quickly switching to the default power supply mode when an abnormality is triggered.

[0035] In a specific implementation, a 3.2TB enterprise-class SSD under 70% load saw an increase in access frequency and a larger phase adjustment angle, resulting in a higher actual temperature drop and improved energy efficiency. When the phase adjustment reached 90°, the NAND array temperature rise rate significantly decreased, the input current peak was reduced, and the impact on I / O performance was minimal. In one embodiment, for step S6, The steps of dynamically adjusting the SSD power supply phase according to the pulsed heat dissipation control signal, optimizing heat dissipation through pulse width modulation technology, and feeding back updated thermal status data to the inner and outer ring temperature monitoring include: Converting the pulsed heat dissipation control signal into a duty cycle parameter of the SSD power supply phase through a pulse width modulator; Collect temperature sensor data from the SSD chipset in real time and update the block temperature mapping table; The updated thermal state vector is re-input into the inner and outer loop temperature monitoring to complete the dynamic heat dissipation optimization closed loop.

[0036] Specifically, real-time thermal management of SSD is achieved through dynamic adjustment of power supply phase and closed-loop feedback control.

[0037] The dynamic adjustment mechanism for power supply phases uses a digital multi-phase Buck controller (such as the Infineon IR35217) to convert thermal control signals into four interleaved power supply phases. These phases include the core voltage (VCC), NAND array voltage (VCCP), I / O voltage (VCCQ), and high-voltage programming voltage (VPP), each with a different duty cycle adjustment range and operating mode. When the PWM duty cycle varies by more than 20%, voltage feed-forward compensation is enabled, the switching frequency is adjusted, and a soft-start interval is inserted. An energy efficiency balancing algorithm dynamically selects the operating mode based on temperature and performance requirements, as follows:

[0038] For real-time temperature data collection and mapping, a sensor network monitors die-level temperature using TSMC 40nm process temperature sensors (±1°C accuracy) integrated into each NAND die. An inverse distance weighted (IDW) algorithm is used to generate a full-disk temperature map based on 16 sensor data points. The system refreshes the logical block-to-physical die temperature mapping every 50ms, implementing differentiated temperature compensation for SLC, MLC, and TLC blocks. The master controller mapping table is directly updated through the PCIe memory window with a latency of <5μs. Closed-loop feedback control recalculates the thermal state vector, including updated thermal shock peak values, corrected average temperatures, and temperature gradient indicators. The feedback path is optimized, with temperature critical events directly triggering hardware resets via sideband signals. Regular updates are written to a ring buffer via DMA, and PTP is used to align sensor data timestamps with power phase timestamps (with an error of <100ns). When conditions such as single-phase current exceeding 120% of the rated value, PWM controller temperature ≥95°C, or input voltage drop >10% are detected, it automatically switches to full-phase 0° mode; data integrity is ensured by enabling ECC error correction code enhancement mode, forcibly flushing DRAM cache data, and implementing triple-module redundant storage of key FTL metadata.

[0039] Reference Figure 2 , is a structural block diagram of an SSD solid state drive heat dissipation optimization system based on dynamic temperature control in one embodiment of the present invention, including: A thermal status monitoring unit is used to obtain a thermal status vector of the SSD hard disk by monitoring the internal and external ring temperatures. The thermal status vector includes a thermal shock peak value and an average temperature of the entire SSD; A temperature prediction unit is used to input the thermal state vector into a pre-trained long short-term memory network model, predict the temperature trajectory in the next 5 seconds, and generate a temperature prediction sequence; a resource allocation unit, configured to generate a computing power allocation ratio based on the temperature prediction sequence and determine a capacity migration ratio; The scheduling optimization unit is used to classify I / O queues into heat-sensitive and heat-tolerant types based on the real-time temperature values of each logical block in the block temperature mapping table and the computing power allocation ratio and capacity migration ratio, and generate an optimized scheduling sequence; A signal generating unit, configured to calculate a power supply phase parameter and generate a pulsed heat dissipation control signal according to the access frequency of the heat-sensitive block in the optimized scheduling sequence; The heat dissipation unit is used to dynamically adjust the SSD power supply phase according to the pulse heat dissipation control signal, optimize heat dissipation through pulse width modulation technology, and feed back updated thermal status data to the inner and outer ring temperature monitoring.

[0040] For the specific implementation of each unit in the above device example, please refer to the above method embodiment, which will not be repeated here.

[0041] In summary, the present invention obtains the thermal state vector of the SSD hard disk through internal and external ring temperature monitoring, and the thermal state vector includes the thermal shock peak and the overall average temperature of the SSD; the thermal state vector is input into the pre-trained long short-term memory network model to predict the temperature trajectory of the next 5S and generate a temperature prediction sequence; based on the temperature prediction sequence, the computing power allocation ratio is generated, and the capacity migration ratio is determined at the same time; according to the real-time temperature value of each logical block in the block temperature mapping table and the computing power allocation ratio and capacity migration ratio, the I / O queue is classified into heat-sensitive and heat-tolerant types, and an optimized scheduling sequence is generated; according to the access frequency of the heat-sensitive blocks of the optimized scheduling sequence, the power supply phase parameters are calculated to generate a pulsed heat dissipation control signal; the SSD power supply phase is dynamically adjusted according to the pulsed heat dissipation control signal, heat dissipation is optimized through pulse width modulation technology, and the updated thermal state data is fed back to the internal and external ring temperature monitoring to achieve the purpose of dynamic intelligent temperature control of the SSD solid state hard disk.

[0042] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware using a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media provided herein and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM.

[0043] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0044] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for optimizing heat dissipation of an SSD solid state drive based on dynamic temperature control, characterized in that: The following steps are involved: Obtain the thermal state vector of the SSD hard disk by monitoring the internal and external ring temperatures. The thermal state vector includes the thermal shock peak value and the overall average temperature of the SSD; Input the thermal state vector into a pre-trained long short-term memory network model to predict the temperature trajectory in the next 5 seconds and generate a temperature prediction sequence; Based on the temperature prediction sequence, a computing power allocation ratio is generated, and a capacity migration ratio is determined; Based on the real-time temperature values of each logical block in the block temperature mapping table, as well as the computing power allocation ratio and capacity migration ratio, I / O queues are classified into heat-sensitive and heat-tolerant types, and an optimized scheduling sequence is generated. Calculating power supply phase parameters based on the access frequency of the heat-sensitive blocks in the optimized scheduling sequence and generating a pulsed heat dissipation control signal; The SSD power supply phase is dynamically adjusted according to the pulsed heat dissipation control signal, heat dissipation is optimized through pulse width modulation technology, and updated thermal status data is fed back to the inner and outer ring temperature monitoring.

2. The SSD solid state drive heat dissipation optimization method based on dynamic temperature control according to claim 1, characterized in that: The step of obtaining the thermal state vector of the SSD hard disk by monitoring the internal and external ring temperatures includes: Through the inner loop monitoring mechanism, 10ms granularity thermal shock data is obtained from the SSD flash memory controller, and the temperature change values of flash memory page programming and erasing operations are collected to generate a real-time thermal shock sequence; extracting thermal shock peak values and frequency component characteristic values from the real-time thermal shock sequence; The outer loop monitoring mechanism obtains SSD overall temperature data with 1-second granularity and calculates the overall average temperature of the SSD. The thermal shock peak value and the overall average temperature are weightedly fused and normalized to generate a standardized thermal state vector.

3. The SSD solid state drive heat dissipation optimization method based on dynamic temperature control according to claim 1, characterized in that: Before inputting the thermal state vector into the pre-trained long short-term memory network model, the method further includes a model triggering condition judgment step, including: Calculating a ratio K of a thermal shock peak value to an overall average temperature in the thermal state vector; When the composite trigger condition is met, K ≥ thermal shock coefficient threshold and the overall average temperature ≥ basic temperature threshold, the long short-term memory network model is activated, and the thermal shock coefficient threshold and the basic temperature threshold are pre-set according to the SSD model; If the composite trigger condition is not met, the temperature prediction sequence generated in the previous monitoring cycle is used for heat dissipation control.

4. The SSD solid state drive heat dissipation optimization method based on dynamic temperature control according to claim 1, characterized in that: The step of inputting the thermal state vector into a pre-trained long short-term memory network model, predicting the temperature trajectory in the next 5 seconds, and generating a temperature prediction sequence includes: Inputting the thermal state vector into a pre-trained long short-term memory network model; Based on the temporal correlation between the thermal shock peak value and the overall average temperature, predict the temperature change trajectory in the next 5 seconds; Outputs a temperature forecast sequence containing timestamps and temperature values.

5. The SSD solid state drive heat dissipation optimization method based on dynamic temperature control according to claim 1, characterized in that: The steps of constructing the long short-term memory network model include: Building a training data set based on SSD historical operation data, the data set includes thermal shock peak sequence data, average temperature sequence data and temperature change trajectory data; Design an LSTM network with a bidirectional gating structure. The number of nodes in its input layer is consistent with the dimension of the thermal state vector, and the output layer is the temperature prediction value per second in the next 5 seconds. The sliding window mechanism is used to segment the training data into time series, with a window length of 10 monitoring cycles and a sliding step of 1 cycle; The loss function is set as the root mean square error between the predicted temperature trajectory and the actual temperature value, and the weight penalty term of the thermal shock peak is superimposed.

6. The SSD solid state drive heat dissipation optimization method based on dynamic temperature control according to claim 1, characterized in that: The step of generating a computing power allocation ratio based on the temperature prediction sequence and determining a capacity migration ratio includes: According to the block temperature mapping table stored in the SSD, the temperature prediction sequence is trend-fitted, and the computing power distribution ratio between the host CPU thread and the SSD master control is calculated using a dynamic programming algorithm; According to the peak distribution of the temperature prediction sequence, a greedy algorithm is used to determine the capacity migration ratio of the non-volatile write cache and the DRAM cache.

7. The SSD solid state drive heat dissipation optimization method based on dynamic temperature control according to claim 1, characterized in that: The step of classifying the I / O queues into heat-sensitive and heat-tolerant types and generating an optimized scheduling sequence includes: Based on the real-time temperature values of each logical block in the block temperature mapping table, I / O requests of logical blocks with associated temperature values ≥ a preset temperature threshold are classified as heat-sensitive queues; Based on the capacity migration ratio, the non-volatile write cache allocation weight corresponding to the heat-sensitive queue is reduced, and the DRAM cache allocation weight is increased; Re-ordering the priorities of the I / O requests in the heat-sensitive queue from low to high according to the temperature values of their associated logical blocks; Allocate additional SSD master computing resources to heat-sensitive queues based on the computing power allocation ratio; Generate an optimized scheduling sequence including reordering results and resource allocation parameters.

8. The SSD solid state drive heat dissipation optimization method based on dynamic temperature control according to claim 1, characterized in that: The step of calculating the power supply phase parameter and generating the pulsed heat dissipation control signal according to the access frequency of the heat-sensitive block in the optimized scheduling sequence includes: Counting the access frequency of heat-sensitive blocks in the optimized scheduling sequence; The thermal resistance model is called by the UEFI runtime service to calculate the power phase adjustment parameter that is positively correlated with the access frequency; A pulse width modulation signal is generated according to the power supply phase adjustment parameter.

9. The SSD solid state drive heat dissipation optimization method based on dynamic temperature control according to claim 1, characterized in that: The steps of dynamically adjusting the SSD power supply phase according to the pulsed heat dissipation control signal, optimizing heat dissipation through pulse width modulation technology, and feeding back updated thermal status data to the inner and outer ring temperature monitoring include: Converting the pulsed heat dissipation control signal into a duty cycle parameter of the SSD power supply phase through a pulse width modulator; Collect temperature sensor data from the SSD chipset in real time and update the block temperature mapping table; The updated thermal state vector is re-input into the inner and outer loop temperature monitoring to complete the dynamic heat dissipation optimization closed loop.

10. An SSD solid state drive heat dissipation optimization system based on dynamic temperature control, characterized in that: include: A thermal status monitoring unit is used to obtain a thermal status vector of the SSD hard disk by monitoring the internal and external ring temperatures. The thermal status vector includes a thermal shock peak value and an average temperature of the entire SSD; A temperature prediction unit is used to input the thermal state vector into a pre-trained long short-term memory network model, predict the temperature trajectory in the next 5 seconds, and generate a temperature prediction sequence; a resource allocation unit, configured to generate a computing power allocation ratio based on the temperature prediction sequence and determine a capacity migration ratio; The scheduling optimization unit is used to classify I / O queues into heat-sensitive and heat-tolerant types based on the real-time temperature values of each logical block in the block temperature mapping table and the computing power allocation ratio and capacity migration ratio, and generate an optimized scheduling sequence; A signal generating unit, configured to calculate a power supply phase parameter and generate a pulsed heat dissipation control signal according to the access frequency of the heat-sensitive block in the optimized scheduling sequence; The heat dissipation unit is used to dynamically adjust the SSD power supply phase according to the pulse heat dissipation control signal, optimize heat dissipation through pulse width modulation technology, and feed back updated thermal status data to the inner and outer ring temperature monitoring.

Images