Runtime PE array dynamic power-down task migration and wake-up prediction method

By constructing a phase space state vector sequence and using the maximum Lyapunov exponent to evaluate prediction reliability, the problem that linear regression models in the existing technology cannot capture nonlinear dynamic characteristics is solved, achieving high-precision task migration and wake-up prediction and improving the real-time performance of task migration for processing element arrays.

CN123068077APending Publication Date: 2026-11-10CHANGSHA SEMICONDUCTOR TECHNOLOGY & APPLICATION INNOVATION RESEARCH INSTITUTE +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611169974.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-04
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing predictive early wake-up techniques, which use linear regression models, cannot capture the nonlinear dynamic characteristics in the load sequence of processing components. This leads to a bottleneck in prediction accuracy and an inability to quantify prediction reliability, thus affecting the real-time performance of task migration.

Method used

By collecting and processing time-series data from components, a phase space state vector sequence is constructed through delayed coordinate embedding. The maximum Lyapunov exponent is used to determine the reliable prediction time range, thereby enabling hierarchical pre-wake-up decision-making.

Benefits of technology

It breaks through the prediction accuracy ceiling of linear regression models, realizes the reliability assessment of prediction results, avoids false wake-ups or missed wake-ups caused by unreliable predictions, and improves the real-time performance and accuracy of task migration.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068077A_ABST
    Figure CN123068077A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of power management technology for high-performance computing chips, specifically a method for predicting task migration and wake-up during dynamic power-down of a runtime PE array. It involves collecting runtime time-series data of each processing element in the processing element array; embedding the time-series data of each processing element into delay coordinates according to preset delay times and embedding dimensions to construct a state vector sequence of each processing element in phase space; predicting the load state of each processing element in future control cycles based on the state vector sequence; and calculating the maximum Lyapunov exponent of the load sequence of each processing element to quantify the reliable time range of the prediction results; and performing graded pre-wake-up on processing elements in the power-down state based on the predicted load state and the reliable time range. This invention realizes a paradigm shift in processing element load prediction from linear fitting to nonlinear dynamic system state reconstruction, significantly improving prediction accuracy and the reliability of pre-wake-up decisions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power management technology for high-performance computing chips, and in particular to a method for predicting task migration and wake-up when the PE array is dynamically powered down during runtime. Background Technology

[0002] In dynamic power management of processing element arrays, task migration is widely used to move tasks from high-load or high-temperature processing elements to idle or low-temperature processing elements to achieve load balancing and thermal management. However, there is a significant wake-up delay when a powered-off processing element is re-waked, which seriously affects the real-time performance of task migration.

[0003] To overcome this problem, existing technologies have proposed predictive early wake-up methods that wake up or partially wake up powered-down processing elements in advance by predicting potential future task migration needs. For example, existing technologies use an online sliding window ridge regression model, taking the temperature, power consumption, and processing element state within a historical window as input, to predict the future core temperature. If a task migration is predicted to occur in the next control cycle, the dark silicon processing elements are woken up in advance.

[0004] However, existing predictive early wake-up technologies have the following shortcomings: First, the predictive model capacity is insufficient. Existing technologies mainly use linear regression models to fit and predict temperature or load sequences, which implicitly assume that the system behavior follows a linear relationship. However, the task load sequence of processing component arrays is essentially the output of a nonlinear dynamic system driven by multiple coupled factors. Linear models struggle to capture the implicit nonlinear dynamic characteristics, resulting in a fundamental bottleneck in prediction accuracy.

[0005] Second, the reliability of predictions cannot be quantified. Existing technologies assign the same reliability to all predictions, failing to distinguish between reliable and unreliable predictions. In nonlinear dynamic systems, prediction errors grow rapidly over time, and predictions exceeding a certain time frame are inherently unreliable. The lack of quantitative assessment methods for this inherent unreliability of predictions in existing technologies leads to false or missed wake-ups in pre-wake decisions, potentially guided by unreliable predictions. Summary of the Invention

[0006] This invention proposes a method for predicting task migration and wake-up when the PE array is dynamically powered down during runtime. It aims to solve the technical problem that existing predictive early wake-up technologies use linear regression models to predict the load of processing elements, which cannot capture the nonlinear dynamic characteristics hidden in the load sequence, resulting in a fundamental bottleneck in prediction accuracy.

[0007] This invention provides a method for predicting task migration and wake-up when a PE array is dynamically powered down during runtime, including: Collect time-series data of each PE in the PE array during runtime; The time series data of each PE are embedded with delayed coordinates according to a preset delay time and embedding dimension to construct a state vector sequence of each PE in phase space; Based on the state vector sequence of each PE in the phase space, predict the load state of each PE in future control cycles. Based on the state vector sequence of each PE in the phase space, calculate the maximum Lyapunov exponent of the load sequence of each PE, and determine the prediction reliability time range of each PE based on the maximum Lyapunov exponent. Based on the predicted load state and the predicted reliable time range, a graded pre-wake-up is performed on the PE in the power-down state.

[0008] Furthermore, the time series data is the load rate sequence of each PE.

[0009] Furthermore, the delay time is determined by calculating the mutual information of the time series data at different delay times, and selecting the delay time when the mutual information first reaches a local minimum as the delay time.

[0010] Furthermore, the embedding dimension is determined by the following method: starting from the embedding dimension equal to 1 and incrementing, the proportion of pseudo nearest neighbors in the phase space under each embedding dimension is calculated, and the embedding dimension when the proportion of pseudo nearest neighbors first drops below a preset threshold is selected as the embedding dimension.

[0011] Further, predicting the load state of each PE in future control cycles based on the state vector sequence of each PE in the phase space includes: For the current state vector, search for K nearest neighbor state vectors in the historical state vector sequence of the phase space, where K is a preset positive integer; Obtain the state vector of each of the K nearest neighbor state vectors at the next time step; Based on the state vectors of the K nearest neighbors in the next time step, the state vector of the current state vector in the next time step is predicted by weighted average, wherein the weight of each nearest neighbor is inversely proportional to its distance from the current state vector. By iterating the above process, the predicted load state values ​​for multiple future control cycles are obtained.

[0012] Further, the step of calculating the maximum Lyapunov exponent of the load sequence of each PE based on the state vector sequence of each PE in the phase space specifically includes: In the phase space, for each state vector, find its nearest neighbor state vector and calculate the initial distance; Track the evolution of each state vector and its nearest neighbor state vector at subsequent time steps, and calculate the distance after k steps; Calculate the average logarithmic divergence of all nearest neighbor pairs after k steps, and use the slope of the linear portion of the curve of the average logarithmic divergence as a function of k as the maximum Lyapunov exponent.

[0013] Furthermore, the determination of the prediction reliability time range for each PE based on the maximum Lyapunov exponent specifically includes: When the maximum Lyapunov exponent is greater than 0, the reliable prediction time range is equal to a preset constant divided by the maximum Lyapunov exponent. When the maximum Lyapunov exponent is less than or equal to 0, the prediction reliability time range is set to the preset maximum value.

[0014] Furthermore, the determination of the prediction reliability time range for each PE based on the maximum Lyapunov exponent specifically includes: When the predicted load state is greater than the preset high load threshold and the prediction time is less than or equal to the prediction reliability time range of the corresponding PE, a full pre-wake-up is performed on the PE. The full pre-wake-up refers to switching the power state of the PE from a completely power-off state to a full voltage and full clock working state. When the predicted load state is greater than the preset high load threshold but the prediction time is greater than the prediction reliability time range of the corresponding PE, a partial pre-wake-up is performed on the PE. The partial pre-wake-up refers to switching the power state of the PE from a completely power-off state to a data holding voltage state and clock supply restored. When the predicted load state is less than or equal to the preset high load threshold, the PE will not be pre-wake-up and will remain in its current power-down state.

[0015] Furthermore, it also includes the following steps: After each control cycle ends, the actual observed time series data is added to the historical series of each PE, and the historical window is updated in a sliding window manner. The delay time and embedding dimension of each PE are redefined every preset number of control cycles, and the maximum Lyapunov exponent of each PE is recalculated. The state vector sequence of each PE is updated incrementally in each control cycle.

[0016] Furthermore, the PE array is a multi-core processor, a reconfigurable computing array, an artificial intelligence acceleration chip, or a graphics processor; the control cycle ranges from 10 microseconds to 100 microseconds.

[0017] The technical effects of the technical solution disclosed in this invention are: First, by embedding delayed coordinates into the time series data of the processing elements to construct a phase space state vector sequence, load prediction is transformed from linear time series fitting to nonlinear dynamic system state space analysis, fundamentally breaking through the prediction accuracy ceiling of the linear regression model.

[0018] Second, by calculating the maximum Lyapunov exponent of the load sequence of each processing element, a quantitative assessment of the reliable time range of the prediction is achieved, enabling the pre-wake decision to distinguish the reliability differences of different prediction results on the time scale and avoid performing pre-wake operations under the guidance of unreliable predictions. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the task migration and wake-up prediction method for dynamic power-down of the PE array proposed in this embodiment of the invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The predictive early wake-up technology mentioned in the background section uses a linear regression model to predict the load of processing elements, which fails to capture the nonlinear dynamic characteristics implicit in the load sequence, resulting in a fundamental bottleneck in prediction accuracy. This embodiment uses an artificial intelligence acceleration chip containing 16 processing elements arranged in a four-row, four-column grid topology as an example for illustration, but the scope of application of this invention is not limited to this.

[0022] I. System Initialization and Parameter Configuration: During system startup, configure the following core parameters: Control cycle: 50 microseconds. Historical window length: 300 sampling points, corresponding to 15 milliseconds of historical data. Delay time search range: 1 to 30 control cycles. Embedding dimension search range: 1 to 10. Nearest neighbor count K: 20. High load threshold: 0.80. Lyapunov exponent constant C: 1.

[0023] The above parameters can be optimized offline based on the specific chip architecture and workload characteristics, or they can be dynamically adjusted during runtime.

[0024] II. Data Acquisition: At the beginning of each control cycle, i.e., at the beginning of every 50 microseconds, the chip's built-in monitoring circuit synchronously acquires the runtime time-series data of the 16 processing elements. In this embodiment, the acquired time-series data is the load rate of each processing element.

[0025] The load factor is calculated by dividing the number of active clock cycles of each processing element's execution unit in the past control cycle by the total number of clock cycles. For example, if a processing element has 2500 clock cycles in 50 microseconds, assuming a clock frequency of 50 MHz, and 1800 of those cycles are active, then the load factor of that processing element is 1800 divided by 2500, which equals 0.72.

[0026] The collected load rate data is stored in the historical buffer of each processing element in chronological order, with each buffer retaining the data from the most recent 300 sampling points. When new data arrives, the oldest data point is removed, implementing a sliding window update to always maintain a historical window length of 300 sampling points.

[0027] Taking the processing element located in the second row and second column of the array as an example, its history buffer stores the load rate sequence of 300 sampling points.

[0028] III. Phase Space Reconstruction.

[0029] (a) Determine the delay time. Taking the load rate sequence of the processing element located in the second row and second column of the array as an example, determine the delay time.

[0030] The delay time is determined by calculating the mutual information of the load rate sequence at different delay times, and selecting the delay time when the mutual information first reaches a local minimum as the delay time.

[0031] The formula for calculating mutual information is: ; in, Let be the probability that the load rate sequence takes the value 'a'. Let b be the probability that the delayed load rate sequence takes the value b. Let τ be the joint probability that the load rate sequence takes the value a and the delayed load rate sequence takes the value b, where τ is the delay time.

[0032] The curve of mutual information change with delay time was calculated. Calculations showed that the mutual information first reaches a local minimum when the delay time is equal to 4 control cycles; therefore, a delay time of 4 was chosen as the delay time.

[0033] Similarly, the delay time is calculated for each of the other 15 processing elements. Since the load patterns of each processing element may differ, the delay time values ​​may vary slightly, typically ranging from 3 to 6 control cycles.

[0034] (ii) Determine the embedding dimension. Taking the processing element located in the second row and second column of the array as an example, determine the embedding dimension.

[0035] The embedding dimension is determined as follows: starting from an embedding dimension of 1, it is incremented to embed the one-dimensional load rate time series into a multi-dimensional phase space, resulting in a state vector. For each state vector, its nearest neighbor state vector is found in the phase space. Then, it is determined whether the nearest neighbor is still a true nearest neighbor when the dimension increases by 1. If the growth rate of the distance between two points after the dimension increase exceeds a preset threshold (usually 10), then the nearest neighbor is a pseudo-nearest neighbor. The proportion of pseudo-nearest neighbors to all nearest neighbors under each embedding dimension is calculated. When this proportion first drops below the preset threshold, the embedding dimension is the true embedding dimension. In this embodiment, the preset threshold is set to 5%.

[0036] Calculate the proportion of pseudo-nearest neighbors under different embedding dimensions. The calculation shows that when the embedding dimension is 1, the proportion of pseudo-nearest neighbors is 45%; when it is 2, it is 28%; when it is 3, it is 12%; when it is 4, it is 4%; and when it is 5, it is 3%. When the embedding dimension is 4, the proportion of pseudo-nearest neighbors drops below 5% for the first time. Therefore, embedding dimension 4 is chosen as the embedding dimension.

[0037] (iii) Delay coordinate embedding to construct state vector sequence: After determining the delay time and embedding dimension, delay coordinate embedding is performed on the load rate time series of each processing element to construct its state vector sequence in phase space.

[0038] Taking the processing element located in the second row and second column of the array as an example, its load rate sequence is L(t), the delay time τ is 4 control cycles, and the embedding dimension m is 4. The state vector of this processing element at time t is: ; Substituting specific values, in the 150th control cycle: =[0.72, 0.55, 0.68, 0.62]; Following the above method, state vectors are constructed point by point for the 300 sampling points in the historical window, resulting in the state vector sequence: X(1), X(2), ..., X(300).

[0039] The theoretical basis of phase space reconstruction lies in the fact that the complete evolutionary information of a nonlinear dynamic system can be recovered in a high-dimensional phase space through the time series of any one of its state variables. After reconstructing the load rate sequence of the processing element into the phase space, the future evolution trend of the load can be analyzed and predicted through the geometric properties of the state trajectory in the phase space, without the need to pre-assuming that the system follows a linear relationship. This is a fundamental improvement of the present invention over existing linear regression prediction methods.

[0040] IV. Load Status Prediction.

[0041] In the phase space of each processing element, the load state of each processing element in future control cycles is predicted based on the state vector sequence.

[0042] The prediction method is the local nearest neighbor prediction method. Its basic idea is that adjacent state vectors in the phase space remain adjacent in subsequent evolution. Therefore, the future evolution of the current state can be predicted by using the subsequent evolution of the nearest state vectors of the current state vector.

[0043] The specific prediction process is as follows.

[0044] For the state vector X(t) of the processing element located in the second row and second column of the array at time t, search for its K nearest neighbor state vectors in the historical state vector sequence, where K is 20. The nearest neighbor metric is Euclidean distance, which is the square root of the sum of the squares of the differences between the corresponding dimensions of the two state vectors.

[0045] Find the K nearest neighbor state vectors X( ), X( ), ..., X( After that, obtain the state vector X( of each of these K state vectors at the next time step.) ), X( ), ..., X( ).

[0046] The weighted average of these K successor state vectors is calculated as the predicted value of X(t+1). The weights are inversely proportional to the Euclidean distance from each nearest neighbor to X(t), meaning that the closer the neighbor is, the greater its weight. Specifically, the weight of the j-th nearest neighbor is: ; in Let X(t) be the Euclidean distance from the j-th nearest neighbor to X(t).

[0047] The predicted value of X(t+1) is: ; The first component of the predicted value is the predicted load rate for the next control cycle.

[0048] Repeat the above process, using the predicted state vector as the known state to continue iterating, and you can obtain the predicted load state value for the next H control cycles, where H is the preset number of prediction steps, which is 5 in this embodiment.

[0049] For example, it is predicted that the load rate of the processing element located in the second row and second column will be: 0.78 in the next control cycle, 0.82 in the second control cycle, 0.85 in the third control cycle, 0.79 in the fourth control cycle, and 0.72 in the fifth control cycle.

[0050] The computational complexity of the aforementioned local nearest neighbor prediction method is O(N·K·m), where N is the historical window length of 300, K is the number of nearest neighbors of 20, and m is the embedding dimension of 4. It can be completed in real time on the microcontroller built into the chip without causing any additional burden on the normal operation of the chip.

[0051] V. Determination of the reliable time range for the calculation and prediction of the maximum Lyapunov index.

[0052] The maximum Lyapunov exponent is a quantitative indicator that measures the sensitivity of a nonlinear dynamic system to initial conditions. A maximum Lyapunov exponent greater than 0 indicates that the system exhibits chaotic characteristics, with prediction errors increasing exponentially with time. Its reciprocal gives the maximum predictable time scale of the system.

[0053] (a) Calculate the maximum Lyapunov exponent. The maximum Lyapunov exponent of each processing element's load sequence is calculated using the small data volume method. The specific calculation process is as follows: In the phase space of this processing element, for each state vector Find its nearest neighbor state vector Calculate the initial distance: ; Then, the evolution of these two states in subsequent time steps is tracked, and the time progression is calculated. Distance after step: ; For all initial nearest neighbor pairs, calculate the average logarithmic divergence rate: ; in, For evolutionary step size, The number of nearest neighbor pairs, Let be the initial distance of the j-th nearest neighbor pair. Let be the distance between the j-th nearest neighbor pair after k steps.

[0054] draw The slope of the linear portion of the curve representing the change of k is the maximum Lyapunov exponent. .

[0055] The maximum Lyapunov exponent of the load sequence of the processing element located in the second row and second column was calculated. The value is 0.082, measured per control cycle. Since this value is greater than 0, it indicates that the load sequence of this processing element exhibits chaotic characteristics, meaning it is sensitive to initial conditions and the prediction error increases exponentially with time.

[0056] (ii) Determine the prediction reliability time range. The prediction reliability time range of each processing element is determined based on the maximum Lyapunov index.

[0057] when When greater than 0, the reliable time range of the prediction is... The calculation formula is: ; Where C is a preset constant, and in this embodiment it is taken as 1.

[0058] when When the value is less than or equal to 0, it indicates that the load sequence of the processing element does not have chaotic characteristics, and the reliable time range will be predicted. The preset maximum value is set; in this embodiment, it is set to 100 control cycles.

[0059] The predicted reliable time range for the processing element located in the second row and second column is calculated. =1 / 0.082≈12.2 control cycles, rounded down to 12 control cycles.

[0060] This means that for this processing element, the prediction results within 12 control cycles, or 600 microseconds, have physical reliability; however, the reliability of predictions beyond 12 control cycles cannot be guaranteed due to the inherent unpredictability of chaotic systems.

[0061] This quantification boundary is something that existing linear regression prediction methods simply cannot provide. Existing techniques treat all predictions the same, while this invention provides a definitive answer, through the maximum Lyapunov exponent, to how far in advance the load of each processing element can be predicted.

[0062] The maximum Lyapunov exponent and predicted reliability time range were calculated for the other 15 processing elements. The calculations showed that the chaotic nature of different processing elements varied: the maximum Lyapunov exponent for elements at the array edges was larger, typically greater than 0.10, and the predicted reliability time range was shorter, typically less than 10 control cycles; while the maximum Lyapunov exponent for elements at the array center was smaller, typically less than 0.07, and the predicted reliability time range was longer, typically greater than 15 control cycles.

[0063] VI. Graded pre-wake-up decision: Based on the load state prediction results in step four and the predicted reliability time range in step five, graded pre-wake-up is performed on the processing elements that are in the power-off state.

[0064] (a) Determine the pre-wake level. The graded pre-wake decision shall be executed in accordance with the following rules.

[0065] The first type occurs when the predicted load rate of the processing element is greater than a preset high load threshold, and the predicted time of occurrence of this high load state is... Less than or equal to the predicted reliable time range of the processing element At this time, a full pre-wake-up is performed on the processing element. In this embodiment, the high load threshold is set to 0.80.

[0066] Full wake-up refers to switching the power state of the processing element from a completely power-off state to a full-voltage, full-clock operating state, enabling the processing element to execute tasks immediately.

[0067] The second category involves situations where the predicted load rate of the processing element exceeds a preset high load threshold, but the predicted time of occurrence of this high load state is not specified. Greater than the predicted reliable time range of the processing element At that time, a partial pre-wake-up is performed on the processing element.

[0068] Partial pre-wake-up refers to switching the power state of the processing element from a completely powered-off state to a data holding voltage state with the clock restored. At this time, the processing element is in a shallow sleep state, and the wake-up latency is reduced from 50 microseconds in the completely powered-off state to approximately 5 microseconds, but it is not yet fully activated. This processing method reduces wake-up latency and avoids unnecessary power consumption caused by unreliable predictions leading to full activation of the processing element.

[0069] The third type is when the predicted load rate of the processing element is less than or equal to the preset high load threshold, the processing element is not pre-wake-up and the current power-down state is maintained.

[0070] (II) The decision-making process is illustrated using a specific processing element as an example, taking the processing element located in the second row and second column as an example. This processing element is currently in a power-down state. The prediction results in step four show that its load rates for the next five control cycles are 0.78, 0.82, 0.85, 0.79, and 0.72, respectively, all greater than 0.80. The calculation results in step five show that its predicted reliable time range is 12 control cycles. High load states occur in the first to fifth control cycles, all less than or equal to 12, meaning that all prediction results are within the reliable time range. Therefore, a full pre-wake-up is performed on this processing element.

[0071] Take the processing element located in the third row and fourth column as an example. Its maximum Lyapunov exponent is 0.118, and the predicted reliable time range is 8 control cycles. The prediction results show that it may experience high load in the 15th control cycle, but 15 is greater than 8, which is outside the reliable prediction range. Therefore, only a partial pre-wake-up is performed on this processing element, raising it from deep sleep to shallow sleep, rather than fully activating it.

[0072] Taking the processing element located in the first row and fourth column as an example, its predicted load rate is below 0.60 for the next 5 control cycles, which is less than the high load threshold of 0.80. Therefore, no pre-wake-up operation is performed on this processing element, and it remains in a power-down state.

[0073] (iii) Execute the pre-wake-up operation. After receiving the pre-wake-up command, the power management unit of the chip performs a power state switching operation on the corresponding processing element.

[0074] The full pre-wake-up operation includes the following sub-steps: first, restoring the power supply voltage of the processing element to its normal operating voltage; then, restoring the clock signal supply; and finally, initializing the state of the internal registers and caches of the processing element. The entire process takes approximately 10 microseconds.

[0075] The partial wake-up operation includes the following sub-steps: first, restoring the power supply voltage of the processing element to the data holding voltage, then restoring the clock signal supply, but keeping most of the functional units inside the processing element in an inactive state. The entire process takes approximately 5 microseconds.

[0076] 7. Online incremental update: After each control cycle, the actual observed load rate data is added to the historical buffer of each processing element, the oldest data point is removed, and the window length of 300 is kept unchanged to realize sliding window update.

[0077] Every 10 control cycles, or 500 microseconds, the following calculations are re-executed based on the updated historical data: the latency and embedding dimension of each processing element are re-determined, and the maximum Lyapunov exponent of each processing element is recalculated.

[0078] In each control cycle, the state vector sequence of each processing element is incrementally updated. That is, new state vectors are constructed using newly acquired data according to the determined delay time and embedding dimension, added to the state vector sequence, and the oldest state vector is removed.

[0079] Based on the updated prediction results and the reliable prediction time range, the next round of hierarchical pre-wake-up decision-making is executed.

[0080] The online incremental update mechanism enables the system to adaptively track changes in workload patterns. When the chip switches from performing image recognition tasks to performing natural language processing tasks, the load patterns of each processing element will change significantly, and the maximum Lyapunov exponent may increase or decrease. The pre-wake strategy will then automatically adjust without manual intervention.

[0081] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A method for predicting task migration and wake-up during runtime when the PE array is dynamically powered down, characterized in that: include: Collect time-series data of each PE in the PE array during runtime; The time series data of each PE are embedded with delayed coordinates according to a preset delay time and embedding dimension to construct a state vector sequence of each PE in phase space; Based on the state vector sequence of each PE in the phase space, predict the load state of each PE in future control cycles. Based on the state vector sequence of each PE in the phase space, calculate the maximum Lyapunov exponent of the load sequence of each PE, and determine the prediction reliability time range of each PE based on the maximum Lyapunov exponent. Based on the predicted load state and the predicted reliable time range, a graded pre-wake-up is performed on the PE in the power-down state.

2. The method according to claim 1, characterized in that, The time series data is the load rate sequence of each PE.

3. The method according to claim 1, characterized in that, The delay time is determined by calculating the mutual information of the time series data at different delay times, and selecting the delay time when the mutual information first reaches a local minimum as the delay time.

4. The method according to claim 1, characterized in that, The embedding dimension is determined by incrementing the embedding dimension from 1, calculating the proportion of pseudo nearest neighbors in the phase space under each embedding dimension, and selecting the embedding dimension when the proportion of pseudo nearest neighbors first drops below a preset threshold as the embedding dimension.

5. The method according to claim 1, characterized in that, The step of predicting the load state of each PE in future control cycles based on the state vector sequence of each PE in the phase space includes: For the current state vector, search for K nearest neighbor state vectors in the historical state vector sequence of the phase space, where K is a preset positive integer; Obtain the state vector of each of the K nearest neighbor state vectors at the next time step; Based on the state vectors of the K nearest neighbors in the next time step, the state vector of the current state vector in the next time step is predicted by weighted average, wherein the weight of each nearest neighbor is inversely proportional to its distance from the current state vector. By iterating the above process, the predicted load state values ​​for multiple future control cycles are obtained.

6. The method according to claim 1, characterized in that, The step of calculating the maximum Lyapunov exponent of the load sequence of each PE based on the state vector sequence of each PE in the phase space specifically includes: In the phase space, for each state vector, find its nearest neighbor state vector and calculate the initial distance; Track the evolution of each state vector and its nearest neighbor state vector at subsequent time steps, and calculate the distance after k steps; Calculate the average logarithmic divergence of all nearest neighbor pairs after k steps, and use the slope of the linear portion of the curve of the average logarithmic divergence as a function of k as the maximum Lyapunov exponent.

7. The method according to claim 1, characterized in that, The determination of the reliable prediction time range for each PE based on the maximum Lyapunov exponent specifically includes: When the maximum Lyapunov exponent is greater than 0, the reliable prediction time range is equal to a preset constant divided by the maximum Lyapunov exponent. When the maximum Lyapunov exponent is less than or equal to 0, the prediction reliability time range is set to the preset maximum value.

8. The method according to claim 1, characterized in that, The determination of the reliable prediction time range for each PE based on the maximum Lyapunov exponent specifically includes: When the predicted load state is greater than the preset high load threshold and the prediction time is less than or equal to the prediction reliability time range of the corresponding PE, a full pre-wake-up is performed on the PE. The full pre-wake-up refers to switching the power state of the PE from a completely power-off state to a full voltage and full clock working state. When the predicted load state is greater than the preset high load threshold but the prediction time is greater than the prediction reliability time range of the corresponding PE, a partial pre-wake-up is performed on the PE. The partial pre-wake-up refers to switching the power state of the PE from a completely power-off state to a data holding voltage state and clock supply restored. When the predicted load state is less than or equal to the preset high load threshold, the PE will not be pre-wake-up and will remain in its current power-down state.

9. The method according to claim 1, characterized in that, It also includes the following steps: After each control cycle ends, the actual observed time series data is added to the historical series of each PE, and the historical window is updated in a sliding window manner. The delay time and embedding dimension of each PE are redefined every preset number of control cycles, and the maximum Lyapunov exponent of each PE is recalculated. The state vector sequence of each PE is updated incrementally in each control cycle.

10. The method according to claim 1, characterized in that, The PE array is a multi-core processor, a reconfigurable computing array, an artificial intelligence acceleration chip, or a graphics processor; the control cycle ranges from 10 microseconds to 100 microseconds.