Resource allocation methods, devices, equipment, and storage media based on time constants

By deploying timestamp probes at the processing level of embodied intelligent control devices, calculating time constant spectrum vectors and performing sliding window updates, the problems of end-to-end latency measurement and bottleneck identification in embodied intelligent systems are solved, enabling dynamic resource allocation and improved system performance.

CN123068104APending Publication Date: 2026-11-10YOUDI ROBOT (WUXI) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611299442.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In existing embodied intelligent control systems, end-to-end latency cannot be measured accurately in real time, and bottlenecks cannot be automatically identified, resulting in a disconnect between computing resource scheduling and actual bottlenecks, which affects system performance.

Method used

Timestamp probes are deployed at the input and output boundaries of each processing level of the embodied intelligent control device. The time constant spectrum vector is obtained by calculating the timestamp difference between adjacent probes. The statistical data is updated using a sliding window to determine bottlenecks and perform resource tilting and allocation to dynamically adjust resource allocation in order to identify and alleviate bottlenecks.

Benefits of technology

It has improved the granularity of end-to-end latency measurement to the stage level, enabling real-time identification of bottleneck stages and dynamic adjustment of resource allocation, thereby improving resource utilization and system performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068104A_ABST
    Figure CN123068104A_ABST
Patent Text Reader

Abstract

This application provides a resource allocation method, apparatus, device, and storage medium based on time constants. The method includes: deploying timestamp probes at the input-output boundaries of each processing level of an embodied intelligent control device; calculating the time constant of each processing level based on the timestamp difference between adjacent probes to obtain a time constant spectrum vector, and updating the statistical data of the spectrum vector based on a sliding window; performing bottleneck determination processing on each processing level based on the statistical data to identify the bottleneck level currently limiting performance; performing resource skew allocation on the bottleneck level, and restoring resource balance after the bottleneck level is transferred. This application can achieve real-time observability of the time constants of each link in the entire chain and automatic identification of bottleneck levels, ensuring that computing resources are always allocated to the slowest link currently limiting system performance, thereby improving the resource utilization rate of the embodied intelligent control device.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of embodied intelligence and robot control technology, and in particular to a resource allocation method, apparatus, device and storage medium based on time constant. Background Technology

[0002] Embossed intelligent control devices are intelligent agents deployed in a physical environment and capable of interacting with that environment in real time, such as industrial robots, humanoid robots, and unmanned vehicles. These devices can run an embodied intelligent control system internally, encompassing functional levels such as perception, cognition, decision-making, and execution, to achieve understanding of the environment, estimation of their own state, and planning and execution of actions. The sum of the processing times of each stage of an embodied intelligent system constitutes the system's end-to-end latency, which directly affects the device's real-time performance and task execution quality.

[0003] Currently, there are two main methods for measuring system latency. One method involves setting up recording points on both the system's input and output sides, calculating the end-to-end total latency by measuring the time difference between the output and input. However, this method only provides the overall latency value and cannot pinpoint the specific location of the bottleneck. The other method involves conducting independent benchmark tests on each functional module during the development phase, recording the processing time of each module under standard input as a reference. However, the latency data obtained through this method is static, while the system faces dynamically changing workloads and input data during actual operation. Static data cannot reflect the real-time changes in the processing time of each stage.

[0004] Therefore, how to acquire the delay data of each link in real time during the actual operation of the system, and identify the bottleneck link that limits the overall performance, is a technical problem that urgently needs to be solved in the field of embodied intelligent control. Summary of the Invention

[0005] This application provides a resource allocation method, apparatus, device, and storage medium based on time constant, aiming to solve the technical problems of the inability to accurately measure the time constant of each link in the entire chain during the operation of an embodied intelligent system in real time, the inability to automatically identify bottleneck links, and the disconnect between computational resource scheduling and actual bottlenecks.

[0006] In a first aspect, a resource allocation method based on time constants is provided, applied to an embodied intelligent control device. The method includes: deploying timestamp probes at the input-output boundaries of each processing level of the embodied intelligent control device; calculating the time constant of each processing level based on the timestamp difference between adjacent probes to obtain a time constant spectrum vector, and updating the statistical data of the spectrum vector based on a sliding window; performing bottleneck determination processing on each processing level based on the statistical data to identify the bottleneck level that is currently limiting performance; performing resource skew allocation on the bottleneck level, and restoring the resource balance state after the bottleneck level is transferred.

[0007] By deploying timestamp probes at the input and output boundaries of each processing level and calculating the time constant of each processing level based on the timestamp difference between adjacent probes, the time constant of each link in the entire chain can become an independently observable quantity, improving the information granularity of delay measurement from the system level to the link level. By identifying bottlenecks based on statistical data updated by a sliding window and performing resource tilting at the bottleneck level, computing resources can always be allocated to the slowest link that is currently limiting system performance. By restoring the resource balance state after the bottleneck level is transferred, excessive concentration of resources at non-bottleneck levels can be avoided, thereby improving the resource utilization rate of the embodied intelligent control device.

[0008] In conjunction with the first aspect, in one possible implementation, the processing layers include: a perception layer, a cognition layer, a decision-making layer, and an execution layer; the deployment of timestamp probes at the input-output boundaries of each processing layer of the embodied intelligent control device includes: deploying a first timestamp probe at the input end of the perception layer to record a first timestamp when raw data is input; deploying a second timestamp probe between the output end of the perception layer and the input end of the cognition layer to record a second timestamp when a perception result is generated; deploying a third timestamp probe between the output end of the cognition layer and the input end of the decision-making layer to record a third timestamp when a cognition result is generated; deploying a fourth timestamp probe between the output end of the decision-making layer and the input end of the execution layer to record a fourth timestamp when a decision result is generated; and deploying a fifth timestamp probe at the output end of the execution layer to record a fifth timestamp when an execution result is generated; wherein each timestamp probe reads the current time from a global synchronization clock, attaches the timestamp to the data packet, and passes it forward.

[0009] By deploying timestamp probes at the five input / output boundaries of the four processing levels and having each probe read the current time from a global synchronization clock, it can be ensured that the timestamps of each probe are based on the same clock reference. The real-time delay of the corresponding stage can be obtained by subtracting the timestamps of adjacent probes, and the measurement results are accurate and reliable.

[0010] In conjunction with the first aspect, in one possible implementation, the time constant spectrum vector includes the perception layer time constant, the cross-layer communication time constant from the perception layer to the cognition layer, the cognition layer time constant, the cross-layer communication time constant from the cognition layer to the decision layer, the decision layer time constant, the cross-layer communication time constant from the decision layer to the execution layer, and the execution layer time constant; the sliding window is used to retain the time constant spectrum vector for a preset number of operation cycles; the statistical data includes the mean, standard deviation, and peak value.

[0011] By organizing the time constants of each link and the time constants of each cross-layer communication into a spectrum vector, and calculating the mean, standard deviation and peak value based on the sliding window, the statistical data can always reflect the most recent actual operating state of the system. The mean reflects the typical delay level, the standard deviation reflects the degree of delay fluctuation, and the peak value reflects the worst-case delay, providing a multi-dimensional basis for bottleneck identification.

[0012] In conjunction with the first aspect, in one possible implementation, the bottleneck determination process for each processing level based on the statistical data to identify the current bottleneck level limiting performance includes: comparing the current average time constant of each level with a preset latency tolerance threshold, and marking levels exceeding the latency tolerance threshold as absolute latency bottleneck candidates; calculating the proportion of the time constant of each level in the total latency of the entire link, and marking levels with a proportion exceeding a preset proportion threshold as relative proportion bottleneck candidates; comparing the trend direction of the time constant of each level within the current sliding window with the previous window, and marking levels showing an increasing trend for several consecutive periods as trend drift bottleneck candidates; taking the union of the absolute latency bottleneck candidates, the relative proportion bottleneck candidates, and the trend drift bottleneck candidates, sorting them from highest to lowest severity, and determining the top-ranked bottleneck level as the current performance-limiting bottleneck level.

[0013] By combining the union of three levels of parallel judgments—absolute delay judgment, relative proportion judgment, and trend drift judgment—it can capture explicit bottlenecks that have exceeded the tolerance, implicit bottlenecks that have not reached the tolerance but have consumed most of the delay budget, and potential bottlenecks that are forming. This makes the comprehensiveness and foresight of bottleneck detection significantly better than the single threshold judgment method.

[0014] In conjunction with the first aspect, in one possible implementation, the step of performing resource skew allocation on the bottleneck level and restoring the resource balance state after the bottleneck level is transferred includes: allocating computing resources from non-bottleneck levels to complete the resource skew allocation on the bottleneck level; the resource skew allocation includes any one or more of the following: increasing the available video memory quota of the bottleneck level, increasing the scheduling priority of the computing threads of the bottleneck level, and delaying the execution of prediction tasks of non-bottleneck levels; when it is determined that the bottleneck level has been transferred to other levels, the skewed computing resources are restored to a balanced state.

[0015] By allocating computing resources from non-bottleneck levels to bottleneck levels, and by increasing memory quotas, raising scheduling priority, and delaying prediction tasks in non-bottleneck levels to achieve rapid tilting, resource pressure at the bottleneck level can be quickly alleviated within a single operation cycle. By restoring resource balance after the bottleneck level is transferred, long-term excessive concentration of resources at non-bottleneck levels can be avoided, forming a dynamic following mechanism where resources are allocated to the bottleneck.

[0016] In conjunction with the first aspect, in one possible implementation, the method further includes: when multiple consecutive operation cycles at the same level are determined to be the bottleneck level, triggering a deep optimization configuration process, wherein the deep optimization configuration process includes: reducing the inference accuracy level of the bottleneck level, reducing the batch size, and migrating the computational tasks of the bottleneck level to a higher-performance computing unit.

[0017] By triggering deep optimization configuration processing when multiple consecutive operation cycles at the same level are identified as bottlenecks, in cases where rapid skew allocation fails to eliminate bottlenecks, deeper configuration adjustments can be made, such as reducing inference accuracy levels to improve inference speed, reducing batch size to reduce waiting latency, and migrating to higher-performance computing units. This fundamentally alleviates stubborn bottlenecks and prevents the system from being in a state of unbalanced resource skew for a long time.

[0018] In conjunction with the first aspect, in one possible implementation, the method further includes: performing correlation analysis on historical time constant spectrum data to monitor the coupling degree between time constants at each level; when at least two levels of time constants are detected to exhibit a pattern of mutual increase or decrease, determining that the at least two levels have a resource coupling relationship, and marking the levels with the coupling relationship as a joint scheduling unit; when any level in the joint scheduling unit is determined to be a bottleneck, simultaneously performing resource allocation on at least two levels in the joint scheduling unit.

[0019] By performing correlation analysis on historical time constant spectrum data, resource coupling relationships with the time constants of each link can be automatically discovered. The levels with coupling relationships are marked as joint scheduling units and simultaneously allocated when any level becomes a bottleneck. This can avoid the seesaw effect where optimizing one level leads to the deterioration of the performance of another level. Moreover, the identification of coupling relationships is completely based on the operation data and is completed automatically without manual predefinition. The coupling relationship map can be automatically updated after the system configuration changes.

[0020] Secondly, a time constant-based resource allocation device is provided for use in an embodied intelligent control device. The device includes: a deployment module for deploying timestamp probes at the input-output boundaries of each processing level of the embodied intelligent control device; a calculation module for calculating the time constant of each processing level based on the timestamp difference between adjacent probes, obtaining a time constant spectrum vector, and updating the statistical data of the spectrum vector based on a sliding window; an identification module for performing bottleneck determination processing on each processing level based on the statistical data, and identifying the bottleneck level that is currently limiting performance; and an allocation module for performing resource tilt allocation on the bottleneck level and restoring a resource balance state after the bottleneck level is transferred.

[0021] Thirdly, a personal intelligent control device is provided, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any of the first aspects.

[0022] Fourthly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the first aspects.

[0023] Implementing this application can achieve the following beneficial effects: By deploying timestamp probes at the input and output boundaries of each processing level and calculating the time constant of each processing level based on the timestamp difference between adjacent probes, the time constant of each link in the entire chain can become an independently observable quantity, improving the information granularity of delay measurement from the system level to the link level; By determining bottlenecks based on statistical data updated by a sliding window and performing resource tilting allocation at the bottleneck level, computing resources can always be allocated to the slowest link that currently limits system performance; By restoring the resource balance state after the bottleneck level is transferred, excessive concentration of resources at non-bottleneck levels can be avoided, thereby improving the resource utilization rate of the embodied intelligent control device. Attached Figure Description

[0024] Figure 1 A schematic diagram of the structure of a embodied intelligent control device provided in an embodiment of this application; Figure 2 A flowchart illustrating a resource allocation method based on a time constant provided in an embodiment of this application; Figure 3 A schematic diagram illustrating the deployment of processing levels and timestamp probes, provided for an embodiment of this application; Figure 4 A flowchart illustrating another resource allocation method based on time constant provided in this application embodiment; Figure 5 This is a schematic diagram of a resource allocation device based on a time constant, provided in an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0026] It should be noted that, unless there is a conflict, the various features in the embodiments of this application can be combined with each other, all of which are within the protection scope of this application. Furthermore, although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed differently from the module division in the device or the order in the flowchart. Moreover, the terms "first," "second," and "third" used in this application do not limit the data or execution order, but only distinguish identical or similar items with essentially the same function and effect.

[0027] This application applies to the fields of embodied intelligence and robot control technology, and is particularly applicable to embodied intelligence control systems that include multi-level pipelines with perception, cognition, decision-making and execution layers, especially industrial robot systems based on a large vision-language-action model and a world model.

[0028] An embodied intelligent control device refers to a computing device deployed in a physical environment that is capable of sensing the environment and performing physical actions. An embodied intelligent control device can be an industrial robot, a robotic arm, an autonomous vehicle, a drone, a service robot, or other intelligent device configured with a multi-stage processing pipeline, and is equipped with an embodied intelligent control system. This application does not limit the specific form of the embodied intelligent control device.

[0029] In some feasible implementations, two main methods are used to measure end-to-end latency. Method one is the end-to-end latency measurement method: A timestamp is recorded at both the system input and output sides. The total end-to-end latency is obtained by subtracting the input timestamp from the output timestamp. This method is simple to implement, requiring only two timing points at the beginning and end of the system, but it only provides a general total latency and cannot pinpoint the specific stage where the latency is consumed. Method two is the offline segmented benchmarking method: During system development, independent latency benchmarking tests are performed on each functional module, recording the inference time of each module on a standard test set. This method provides static reference data and cannot reflect the dynamic changes in latency at each stage due to input data distribution and workload variations during actual system operation.

[0030] Furthermore, existing load balancing methods typically employ pre-defined static resource allocation strategies, resulting in scheduling decisions that are disconnected from the actual bottleneck locations in the system. For example, the system allocates computing resources in a fixed manner based on the time consumption ratio of each module measured during the development phase. When changes in runtime workload cause the bottleneck to shift from the cognitive layer to the decision-making layer, the static allocation strategy still favors the cognitive layer, leading to a coexistence of insufficient resources at the bottleneck level and idle resources at non-bottleneck levels. Consequently, the overall system performance is constrained by the slowest link and cannot be improved.

[0031] In view of this, embodiments of this application propose a resource allocation method, apparatus, device, and storage medium based on time constants. The core inventive concept is as follows: deploy timestamp probes at each processing level of the embodied intelligent control device, making the time constant of each link an independently observable quantity; obtain the delay of each link by the difference between adjacent probes and form a spectrum vector, and continuously update the statistical characteristics with a sliding window; then use a three-layer mechanism of absolute delay determination, relative proportion determination, and trend drift determination to comprehensively identify the bottleneck level that currently limits system performance; finally, drive dynamic resource allocation with the bottleneck determination result to form a closed loop of "measurement-determination-scheduling", so that system resources are always allocated to the current bottleneck link.

[0032] The architecture of a embodied intelligent control device according to an embodiment of this application is described below. Please refer to... Figure 1 , Figure 1 This is a schematic diagram of the structure of a embodied intelligent control device provided in an embodiment of this application. Figure 1 As shown, the embodied intelligent control device 10 includes: a processor 101, a memory 102, a communication interface 103, a timestamp probe network 104, a bottleneck detector 105, and a resource scheduler 106.

[0033] Processor 101 may be a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Neural Processing Unit (NPU), Field-Programmable Gate Array (FPGA), or other processors with data processing and control capabilities. Processor 101 is used to execute the control logic in the time-constant-based resource allocation method provided in this application embodiment, such as scheduling computational tasks at each processing level, triggering bottleneck determination, and resource allocation.

[0034] The memory 102 is used to store computer programs, model parameters, timestamp records, and historical data of time constant spectral lines. The memory 102 may include volatile memory and non-volatile memory.

[0035] The communication interface 103 is used to realize data communication between the embodied intelligent control device 10 and sensors, actuators and external devices, such as receiving raw data collected by sensors and issuing action commands to actuators.

[0036] The timestamp probe network 104 consists of multiple timestamp probes deployed at the input and output boundaries of each processing layer, used to record timestamps as data packets pass through. The bottleneck detector 105 is used to perform three-level bottleneck determination based on the time constant spectrum statistics within a sliding window.

[0037] Resource scheduler 106 is used to perform resource tilting, deep optimization configuration and joint scheduling at the bottleneck level based on the bottleneck detection results.

[0038] It should be noted that the bottleneck detector 105 and resource scheduler 106 described above can be implemented by the processor 101 executing the computer program stored in the memory 102, or they can be implemented by dedicated hardware circuits, firmware logic, or a combination thereof. This application embodiment does not limit the physical implementation of each unit.

[0039] The technical solution of this application will now be described in detail with reference to the accompanying drawings. Please refer to the attached drawings. Figure 2 , Figure 2 This is a flowchart illustrating a resource allocation method based on a time constant, provided as an embodiment of this application. This method can be applied to, for example... Figure 1 On the embodied intelligent control device shown, Figure 2 The sequence of steps in the code describes a feasible processing path. Where there are no conflicts, some steps can be executed in parallel, interleaved, or repeated in different rounds. For example... Figure 2 As shown, the method includes the following steps: S101, The embodied intelligent control device deploys timestamp probes at the input and output boundaries of each processing level.

[0040] It should be noted that the processing level refers to the serial processing stages in an embodied intelligent control device, divided according to the data processing flow. The processing level in this embodiment may include a perception layer, a cognition layer, a decision-making layer, and an execution layer. Specifically, the perception layer is used to collect raw data such as images, point clouds, and force sensing; the cognition layer is used to perform scene understanding and future state prediction on the raw data; the decision-making layer is used to plan and execute actions based on the prediction results; and the execution layer is used to issue action commands to the actuators to complete the physical operations.

[0041] A timestamp probe is a lightweight timing component deployed at the input / output boundaries of the processing layer. As a data packet passes through, the timestamp probe reads the current time from a global synchronization clock, attaches a timestamp to the packet, and continues propagating it forward. The timestamp probe itself does not participate in computation, intercept data streams, or modify data content. The added latency of a single probe is on the nanosecond level, having no perceptible impact on normal system operation.

[0042] In one embodiment, the processing layer may include a perception layer, a cognition layer, a decision-making layer, and an execution layer. Step S101 includes: deploying a first timestamp probe at the input of the perception layer to record a first timestamp when raw data is input; deploying a second timestamp probe between the output of the perception layer and the input of the cognition layer to record a second timestamp when a perception result is generated; deploying a third timestamp probe between the output of the cognition layer and the input of the decision-making layer to record a third timestamp when a cognition result is generated; deploying a fourth timestamp probe between the output of the decision-making layer and the input of the execution layer to record a fourth timestamp when a decision result is generated; and deploying a fifth timestamp probe at the output of the execution layer to record a fifth timestamp when an execution result is generated; wherein each timestamp probe reads the current time from a global synchronization clock, attaches the timestamp to the data packet, and passes it forward.

[0043] In one feasible implementation, please refer to Figure 3 , Figure 3 This is a schematic diagram of an architecture for processing hierarchy and timestamp probe deployment provided in an embodiment of this application. Figure 3 As shown, the embodied intelligent control device may include a perception layer 110, a cognition layer 120, a decision layer 130 and an execution layer 140 connected in sequence, and deploy five timestamp probes at the input and output boundaries of each processing layer.

[0044] The first timestamp probe 151 is deployed at the input end of the perception layer 110 to record the first timestamp when the raw data is input; the second timestamp probe 152 is deployed between the output end of the perception layer 110 and the input end of the cognition layer 120 to record the second timestamp when the perception result is generated; the third timestamp probe 153 is deployed between the output end of the cognition layer 120 and the input end of the decision layer 130 to record the third timestamp when the cognition result is generated; the fourth timestamp probe 154 is deployed between the output end of the decision layer 130 and the input end of the execution layer 140 to record the fourth timestamp when the decision result is generated; and the fifth timestamp probe 155 is deployed at the output end of the execution layer 140 to record the fifth timestamp when the execution result is generated.

[0045] It should be noted that each timestamp probe reads the current time from the global synchronization clock, attaches the timestamp to the data packet, and transmits it forward. Since all probes are based on the same global synchronization clock, the difference between the timestamps of any two probes has a clear temporal meaning: the difference between the second and first timestamps is the perception layer time constant; the difference between the third and second timestamps is the sum of the cognition layer time constant and the cross-layer communication time constant from the perception layer to the cognition layer, and so on. The cross-layer communication time constants can be obtained by subtracting the timestamps of adjacent probe pairs and separating the processing time within each layer. Specifically, for cross-layer transmission from the perception layer to the cognition layer, the cross-layer communication time constant equals the third probe timestamp minus the second probe timestamp, and then minus the processing time of the cognition link itself; the other cross-layer communication time constants follow the same principle. Since the link processing time itself is also a statistical quantity obtained from the probe differences, the cross-layer communication time constant is a statistically expected value rather than a single precise value; estimation using the average of multiple data packets within a sliding window is more stable.

[0046] In step S101, by deploying timestamp probes at the input and output boundaries of each processing level, the data packets can automatically carry the entry and exit times of each link during the entire link flow process, providing a measurement basis for subsequent link-by-link time constant calculation, and the measurement process does not affect the normal operation of the system.

[0047] S102. The embodied intelligent control device calculates the time constant of each processing level based on the timestamp difference between adjacent probes, obtains the time constant spectrum vector, and updates the statistical data of the spectrum vector based on the sliding window.

[0048] It should be noted that the time constant refers to the time consumed by a processing stage from receiving input to generating output, and can be represented by the symbol τ. The sum of the time constants of all stages constitutes the total system delay.

[0049] It should also be noted that the time constant spectrum vector is a vector that organizes the time constants of each link in the entire link and the time constants of each cross-layer communication in a fixed order, and is used to completely depict the distribution of system delay within an operation cycle.

[0050] In one embodiment, the time constant spectrum vector may include seven components: the perception layer time constant, the cross-layer communication time constant from the perception layer to the cognition layer, the cognition layer time constant, the cross-layer communication time constant from the cognition layer to the decision layer, the decision layer time constant, the cross-layer communication time constant from the decision layer to the execution layer, and the execution layer time constant. A sliding window is used to retain the time constant spectrum vector for a preset number of operation cycles. Statistical data includes the mean, standard deviation, and peak value.

[0051] For example, the time constant of the perception layer can be obtained by subtracting the first time stamp T1 from the second time stamp T2; the time constant of the cognition layer can be obtained by subtracting the second time stamp T2 from the third time stamp T3, and then subtracting the cross-layer communication overhead from the perception layer to the cognition layer; the cross-layer communication time constant from the perception layer to the cognition layer can be obtained by subtracting the start and end times of the cross-layer transmission; and the time constants of the other stages can be deduced in the same way.

[0052] It should be noted that the mean reflects the typical delay level of each stage, the standard deviation reflects the degree of delay fluctuation, and the peak reflects the worst-case delay. These three statistical features characterize the operational status of each stage from different dimensions, providing multi-dimensional criteria for subsequent bottleneck identification. Among them, the cross-layer communication time constant is included as an independent component of the spectrum vector in the sliding window statistics, making the fluctuation of cross-layer transmission delay an independent signal source for bottleneck identification.

[0053] In one specific implementation, whenever a new data packet completes end-to-end processing, its time constant spectrum vector is pushed into a sliding window, while the oldest data in the window is removed. The sliding window mechanism ensures that statistical characteristics always reflect the system's most recent operational state, avoiding interference from outdated historical data in current bottleneck identification. For example, the sliding window can retain the time constant spectrum vectors of the most recent twenty operating cycles.

[0054] In one specific implementation, the size of the sliding window can be configured according to the system operating frequency and the required sensitivity of the judgment. A larger window results in more stable statistical characteristics and can resist the interference of occasional delay fluctuations on the judgment results, but the response to bottleneck shifts will be slower. A smaller window is more sensitive to performance changes, but it is prone to misjudging instantaneous jitter as trend changes. For scenarios with short control cycles, a larger window (e.g., 100-200 cycles) can be configured to obtain stable statistical characteristics; for scenarios with long control cycles, a smaller window (e.g., 20-50 cycles) can be configured to ensure that bottleneck judgment can respond promptly to performance changes.

[0055] In one specific implementation, the size of the sliding window can be dynamically adjusted according to the task stage or operating status. When the system is in a transitional phase of task switching or rapid load changes, the window size can be temporarily reduced to allow statistical features to converge quickly to the new operating point; once the system enters steady-state operation, the window size is restored to the normal size to maintain the stability of the judgment. Furthermore, different processing levels can be configured with different window sizes; for example, a larger window can be configured for a perception layer with stable fluctuations, while a smaller window can be configured for a decision layer with drastic fluctuations, to further improve the accuracy and response speed of bottleneck identification.

[0056] S103. The embodied intelligent control device performs bottleneck judgment processing on each processing level based on statistical data, and identifies the bottleneck level that is currently limiting performance.

[0057] It should be noted that the bottleneck level refers to the slowest processing level that currently limits the overall performance of the system.

[0058] In one embodiment, step S103 includes: comparing the current average time constant of each level with a preset latency tolerance threshold, and marking levels that exceed the latency tolerance threshold as absolute latency bottleneck candidates; calculating the proportion of the time constant of each level in the total latency of the entire link, and marking levels whose proportion exceeds a preset proportion threshold as relative proportion bottleneck candidates; comparing the trend direction of the time constant of each level in the current sliding window with the previous window, and marking levels that show an increasing trend for several consecutive periods as trend drift bottleneck candidates; taking the union of the absolute latency bottleneck candidates, relative proportion bottleneck candidates, and trend drift bottleneck candidates, sorting them from high to low according to severity, and determining the top-ranked bottleneck level as the current performance-limiting bottleneck level.

[0059] It should be noted that the latency tolerance threshold is set according to the real-time requirements of the operation task. For example, precision assembly tasks require an end-to-end latency of no more than 100 milliseconds, while non-precision tasks require no more than 500 milliseconds. The corresponding latency tolerance thresholds for each level are set separately according to the task type.

[0060] It should also be noted that the relative proportion determination is used to capture hidden bottlenecks that are not within the tolerance limit but consume most of the latency budget: for links with an excessively large proportion, any small fluctuation will have an amplified effect on overall performance. Therefore, even if its absolute latency does not exceed the tolerance threshold, it should be marked as a bottleneck candidate. The trend drift determination is used to capture potential bottlenecks that are forming: links that show an increasing trend for several consecutive periods, even if they have not yet reached their maximum value in terms of absolute latency or relative proportion, are marked as early warning signals, enabling the system to initiate preventive intervention before the bottleneck actually forms.

[0061] It should also be noted that in the above three-level analysis, the cross-layer communication time constant is determined in parallel as one of the process time constants. The first level is absolute delay determination: the current average value of each cross-layer communication time constant is compared with a preset delay tolerance threshold. For example, the allowable upper limit for cross-layer communication in precision assembly tasks is 20 milliseconds, and for non-precision tasks, it is 100 milliseconds. Cross-layer communication processes exceeding the tolerance are marked as bottleneck candidates. The second level is relative proportion determination: the proportion of each cross-layer communication time constant to the total delay of the entire link is calculated. For example, if the combined delay of three cross-layer communication segments exceeds 40% of the total delay, even if their absolute values ​​do not exceed the tolerance, they are still marked as bottleneck candidates. The third level is trend drift determination: the increasing trend of the cross-layer communication time constant within the sliding window is monitored. A continuous upward trend for multiple periods is marked as a warning signal. The time constants of cross-layer communication and internal processing within the process are sent together to the three-level determination, and the results are combined and sorted by severity.

[0062] For example, suppose the average time constants of the perception layer, cognition layer, decision layer, and execution layer are 15 milliseconds, 42 milliseconds, 18 milliseconds, and 10 milliseconds respectively within a certain operation cycle, and the total latency of the entire link is 85 milliseconds. The latency tolerance threshold for the cognition layer is 40 milliseconds, and the preset proportion threshold is 40%. Then, the average time constant of the cognition layer, 42 milliseconds, exceeds the tolerance threshold of 40 milliseconds and is marked as a candidate for absolute latency bottleneck; the cognition layer's proportion, 42 / 85, is approximately 49.4%, exceeding the preset proportion threshold of 40%, and is thus marked as a candidate for relative proportion bottleneck; if the time constant of the perception layer gradually increases from 12 milliseconds to 15 milliseconds over five consecutive cycles, showing an increasing trend, it is marked as a candidate for trend drift bottleneck. The union of these three factors is then used to rank them by severity, and the top-ranked factor is determined as the current bottleneck level.

[0063] In one specific implementation, the severity ranking can comprehensively consider three factors: absolute deviation, percentage, and trend slope. Absolute deviation refers to the difference between the current average time constant and the latency tolerance threshold; a larger deviation indicates a more urgent stage. Percentage refers to the proportion of the time constant to the total latency of the entire link; a higher percentage indicates a greater impact of the stage on overall performance. Trend slope refers to the trend of time constant change within the current window; a steeper slope indicates a faster rate of performance degradation in the stage. This embodied intelligent device can obtain a comprehensive score by weighted summation of the three indicators for each stage, and the highest score is determined as the current bottleneck level. The weighting coefficients can be dynamically adjusted according to the current system state. For example, when the system is close to the task deadline, the weight of absolute deviation is increased; when the system is in steady-state operation, the weight of trend slope is increased to identify potential risks in advance.

[0064] For example, assuming the current total end-to-end latency is 500 milliseconds, the average latency of the perception layer is 120 milliseconds (tolerance 100 milliseconds, deviation 20 milliseconds, accounting for 24%, with a gentle trend), the average latency of the cognition layer is 110 milliseconds (tolerance 150 milliseconds, not exceeding the tolerance, accounting for 22%, but rising from 90 milliseconds to 110 milliseconds for five consecutive cycles, with a steep trend slope), and the average latency of the decision layer is 200 milliseconds (tolerance 150 milliseconds, deviation 50 milliseconds, accounting for 40%, with a stable trend). The perception layer exceeds the tolerance and remains significantly higher, but the trend is stable; the decision layer exceeds the tolerance most severely and accounts for the highest percentage; the cognition layer, although not exceeding the tolerance, shows a clear upward trend. If the system is in a precision assembly task with strict latency tolerance, the decision layer is likely to have the highest overall score because its absolute deviation and percentage are the largest, thus identifying it as the current bottleneck. If the system has just undergone a task switch and the load is still changing, the cognition layer may also be marked as high-risk due to its significant upward trend. This mechanism can adaptively determine bottlenecks based on the optimization goals of different systems, making resource scheduling more targeted.

[0065] In one specific implementation, the ranking of candidate levels can also be based on different weighting strategies according to the current system's optimization objectives. For example, if the total latency across the entire link is close to the task deadline, the system will prioritize addressing the link with the largest absolute latency. In this case, increasing the weight of the absolute deviation can quickly allocate resources to the most pressing bottleneck, maximizing latency reduction within a limited time. If the system is operating normally and none of the links are exceeding their limits, the system focuses more on resource utilization efficiency. In this case, increasing the weight of the percentage can allocate resources to the link that consumes the most latency budget, achieving the greatest overall performance gain by optimizing the link with the largest percentage. If the system has just completed a task switch or load change, the latency of each link may be drifting. In this case, increasing the weight of the trend slope can allow the system to intervene early, initiating preventative scheduling before the bottleneck actually forms, avoiding a passive response after the bottleneck has truly formed. Through the above methods, the system can adaptively adjust its judgment strategy according to the current operating status, taking into account the scheduling needs of different scenarios.

[0066] It should be noted that by combining the union of three levels of parallel judgment—absolute delay judgment, relative proportion judgment, and trend drift judgment—explicit bottlenecks that have exceeded the tolerance limit, implicit bottlenecks that have not reached the tolerance limit but have consumed most of the delay budget, and potential bottlenecks that are forming, can be captured respectively, making bottleneck identification both accurate and forward-looking. When multiple levels of severity are equal, a level is selected as the current bottleneck level according to the preset priority.

[0067] S104. The embodied intelligent control device performs resource allocation at the bottleneck level and restores the resource balance state after the bottleneck level is transferred.

[0068] It's important to note that resource skewing refers to increasing the resource allocation ratio at the bottleneck level within the current cycle or several adjacent operation cycles, while simultaneously decreasing the resource allocation ratio at one or more non-bottleneck levels. The allocation ratio can refer to the available memory quota, the scheduling priority of computing threads, or the batch size quota for each level. When resources are skewed towards a certain level, its processing speed increases, and the time constant decreases accordingly. Non-bottleneck levels may experience a slight increase in their time constant due to reduced resources, but since they are not the limiting factor for overall performance, this slight increase does not affect the overall system throughput. When the original bottleneck level is no longer the slowest point, the system reclaims the previously skewed resources to a balanced state to avoid excessive concentration of resources on non-bottleneck levels.

[0069] For example, suppose the four levels of a certain intelligent control device have equal initial resource quotas. At a certain moment, the bottleneck detector determines that the decision-making layer is the bottleneck level, and the resource scheduler allocates some calculated resources from the perception layer to the decision-making layer. After the allocation, the decision-making layer receives additional resources, and its time constant gradually decreases; the perception layer, due to reduced resources, experiences a slight increase in its time constant, but it remains lower than other stages. After several cycles, the decision-making layer is no longer the slowest stage, while the cognition layer becomes the new bottleneck due to increased workload. The bottleneck detector determines that the bottleneck has shifted to the cognition layer, and the resource scheduler immediately reclaims the resources previously allocated to the decision-making layer and reallocates them to the cognition layer. Through this method, system resources always dynamically shift with the bottleneck, and the resource allocation of each stage matches the actual performance requirements.

[0070] In one embodiment, step S104 includes: allocating computing resources from non-bottleneck levels to complete resource tilting allocation to bottleneck levels; wherein, resource tilting allocation includes any one or more of the following: increasing the available video memory quota of bottleneck levels, increasing the scheduling priority of computing threads in bottleneck levels, and delaying the execution of prediction tasks in non-bottleneck levels; when it is determined that the bottleneck level has moved to other levels, the tilted computing resources are restored to a balanced state.

[0071] It should be noted that increasing the available video memory quota at the bottleneck level can alleviate computational waiting caused by insufficient video memory; increasing the scheduling priority of bottleneck-level computation threads can ensure that bottleneck-level computational tasks receive processor time slices first; and delaying the execution of prediction tasks at non-bottleneck levels can free up computational resources occupied by non-bottleneck levels for use by the bottleneck level. These rapid adjustments take effect within a single operation cycle, offering high response speed but limited adjustment range.

[0072] For example, assuming the cognitive layer is identified as the bottleneck, the resource scheduler increases the available GPU memory quota for the cognitive layer from 4GB to 6GB, raises the scheduling priority of the cognitive layer's computing threads by one level, and postpones some non-critical prediction tasks in the perception layer by two operation cycles. After the above adjustments, if the bottleneck detector determines in a subsequent cycle that the bottleneck has shifted to the decision layer, the resource scheduler immediately restores the GPU memory quota and scheduling priority previously tilted to the cognitive layer to a balanced state, and performs a new round of resource tilting and allocation for the decision layer.

[0073] As can be seen, through the method shown in the embodiments of this application, the embodied intelligent device deploys timestamp probes at the input and output boundaries of each processing layer, making the time constant of each stage an independently observable quantity, thus improving the granularity of latency information from the system level to the stage level. By continuously updating the statistical characteristics of the spectrum vector through a sliding window, the latency measurement always reflects the most recent actual operating state of the system, enabling real-time capture of explicit bottlenecks that have exceeded tolerance limits, implicit bottlenecks with excessive latency, and emerging potential bottlenecks, significantly improving the comprehensiveness and foresight of bottleneck detection. Finally, the bottleneck detection results drive dynamic resource allocation, forming a closed loop of "measurement-judgment-scheduling," ensuring that system resources are always allocated to the slowest stage, eliminating the waste of resources existing simultaneously in bottleneck stages and idle resources in non-bottleneck stages.

[0074] Please see Figure 4 , Figure 4 This is a flowchart illustrating another resource allocation method based on a time constant provided in an embodiment of this application. This method can be executed by an embodied intelligent control device. Figure 4 The method shown, based on the above process, further performs coupling relationship discovery and joint scheduling based on historical time constant spectrum data, and triggers deep optimization configuration when bottlenecks persist. It may include the following steps: S201, the embodied intelligent control device performs correlation analysis on historical time constant spectrum data and monitors the degree of coupling between time constants at each level.

[0075] It's important to note that the different processing layers do not operate completely independently; they may share certain hardware resources. For example, the perception layer and the decision-making layer may share the same GPU's video memory and computing cores, while the cognition layer and the decision-making layer may share CPU computing power or memory bandwidth. When one stage consumes a large amount of shared resources, the available resources for another stage decrease, leading to an increase in its time constant. This phenomenon of one stage accelerating while another slows down is called resource coupling. If this coupling is ignored, simply allocating resources to the bottleneck stage may result in optimizing the latency of one stage but worsening the latency of another, failing to effectively improve the overall system performance.

[0076] It should be noted that correlation analysis refers to performing correlation analysis on the historical time constant spectrum data stored in a sliding window. In one feasible implementation, correlation analysis can be based on Pearson correlation coefficient, Spearman rank correlation coefficient, or other correlation measures to calculate the pairwise correlation coefficients between the time constants of each stage. The correlation coefficient ranges from -1 to 1. A negative value indicates that the time constants of the two stages have an inverse relationship (i.e., when one increases, the other decreases); a positive value indicates a unidirectional relationship (i.e., when one increases, the other also increases); and a value of zero indicates that there is no linear correlation.

[0077] In one specific implementation, coupling determination can be performed in real time based on data within a sliding window, or periodically after each operation cycle. When the hardware configuration of the embodied intelligent control device changes (such as replacing computing units, adding or removing accelerator cards), the model version is upgraded (such as replacing the perception model or decision model with a larger-scale network), or the task type is switched (such as switching from precision assembly to handling and palletizing), old coupling relationships may become invalid, and new coupling relationships may be generated. The embodied intelligent control device automatically updates the coupling relationship graph through continuous correlation analysis, without the need for manual reconfiguration. For example, when the embodied intelligent control device is upgraded from a single GPU to a dual GPU, the perception layer and decision layer, which previously shared the same GPU, each obtain independent computing resources, and the previous coupling relationship automatically disappears. The embodied intelligent control device detects through correlation analysis in subsequent cycles that the correlation coefficient no longer meets the coupling threshold and automatically removes the coupling relationship from the coupling graph.

[0078] S202, The embodied intelligent control device determines whether there is a pattern in which the time constants of at least two levels exhibit a zero-sum pattern.

[0079] It should be noted that the "give-and-take" pattern refers to a negative correlation where the time constant of one layer decreases while the time constant of another layer increases, and this phenomenon recurs. For example, if the perception layer speeds up due to access to more computing resources, while the cognition layer slows down due to the competition for shared resources, and this pattern repeats across multiple operating cycles, it indicates that the perception layer and the cognition layer are competing for a certain shared resource.

[0080] For example, suppose the system performs correlation analysis on historical data and finds that when the time constant of the perception layer decreases from 80 milliseconds to 50 milliseconds, the time constant of the decision layer increases from 100 milliseconds to 140 milliseconds; when the time constant of the perception layer rises back to 70 milliseconds, the time constant of the decision layer falls back to 110 milliseconds. This pattern repeats over multiple cycles, and the correlation analysis yields a correlation coefficient of -0.85 between the two stages. The system determines that there is a trade-off between the perception layer and the decision layer. A possible reason is that both share the same GPU's memory and computing resources. When the perception layer accelerates, it consumes more GPU resources, resulting in fewer available resources and slower processing for the decision layer, and vice versa.

[0081] S203. The embodied intelligent control device determines that there is a resource coupling relationship between at least two levels and marks the levels with the coupling relationship as joint scheduling units.

[0082] It should be noted that resource coupling refers to a competitive relationship between two or more levels regarding shared computing resources (such as video memory, processor time slices, and storage bandwidth). A joint scheduling unit refers to a set of levels that need to be considered as a whole during resource scheduling.

[0083] In one specific implementation, the embodied intelligent control device can construct a coupling relationship graph, recording the currently identified joint scheduling units. This graph uses each processing level as a node and coupling relationships as edges. Edges can carry coupling strength values ​​(e.g., the absolute value of the correlation coefficient) to represent the strength of the coupling. When coupling relationships change, the embodied intelligent control device automatically updates this graph, ensuring that the coupling information used for scheduling decisions remains consistent with the actual state of the embodied intelligent control device.

[0084] S204. When any level in the joint scheduling unit is determined to be a bottleneck, the intelligent control device simultaneously performs resource allocation on at least two levels in the joint scheduling unit.

[0085] It should be noted that if resources are allocated only to the bottleneck level within the joint scheduling unit, the latency of the bottleneck level may increase the latency of the coupled levels due to resource contention between levels, resulting in a decrease in overall performance instead of an improvement. By simultaneously allocating resources to at least two levels within the joint scheduling unit, the latency of one level can be reduced without increasing the latency of another.

[0086] For example, suppose the correlation analysis determines that the perception layer and the cognition layer have a resource coupling relationship and are marked as a joint scheduling unit. When the cognition layer is identified as the bottleneck layer, the embodied intelligent control device does not allocate resources solely to the cognition layer, but adjusts the resource allocation ratio between the perception layer and the cognition layer simultaneously. For example, while increasing the video memory quota for the cognition layer, it reserves the minimum video memory quota to meet the basic processing needs of the perception layer, preventing the perception layer from becoming a new bottleneck due to excessive resource consumption.

[0087] In one feasible implementation, the resource allocation method may include one or more of the following: increasing the available memory quota of the joint scheduling unit, increasing the scheduling priority of the joint scheduling unit's computation threads, and postponing the execution of prediction tasks of non-joint scheduling units. When it is determined that the joint scheduling unit will be transferred to another level, the skewed computing resources can be restored to a balanced state.

[0088] In one feasible implementation, the specific steps of this joint scheduling are as follows: S2041. Identify coupling pairs. Based on accumulated phylogenetic historical data, perform correlation analysis to find that two links repeatedly show a pattern of one increasing and the other decreasing, and mark them as resource coupling pairs.

[0089] S2042. Treat at least two coupled levels as a whole and include them in the joint scheduling unit table. Then treat them as an indivisible scheduling object.

[0090] S2043. Simultaneously adjust the resource quotas of at least two coupled levels.

[0091] When one of the components is identified as a bottleneck and needs to be accelerated, the memory quota of the bottleneck component is increased, the priority of the computing thread is raised, and the corresponding resources of the coupled components are reclaimed in equal amounts to prevent the coupled components from deteriorating due to the crowding out of shared resources.

[0092] S2044. Verify that there is no deterioration on the other side.

[0093] In the next sliding window after adjustment, observe the time constants of at least two levels to confirm that the bottleneck has indeed shortened and the coupling has not lengthened; otherwise, revert to the equilibrium state.

[0094] S2045. Maintain joint scheduling until the bottleneck detector determines that at least two levels are no longer bottlenecks.

[0095] S205. The embodied intelligent control device determines whether multiple consecutive operation cycles at the same level are identified as bottleneck levels.

[0096] It should be noted that if the same level is identified as a bottleneck level in multiple consecutive operating cycles, it means that rapid tilting and allocation cannot eliminate the bottleneck. This bottleneck is a stubborn bottleneck and requires deeper configuration adjustments.

[0097] If multiple consecutive operation cycles at the same level are identified as bottleneck levels, execute S206; otherwise, return to S201 and proceed to the next round of monitoring.

[0098] S206, The embodied intelligent control device triggers deep optimization configuration processing and returns to S201 to enter the next round of monitoring.

[0099] In one embodiment, deep optimization configuration processing may include: reducing the inference accuracy level of the bottleneck level, reducing the batch size, and migrating the computational tasks of the bottleneck level to higher-performance computing units.

[0100] It's important to note that reducing the inference precision level at the bottleneck level refers to adjusting the data precision of the model computation from a higher bit width to a lower bit width (e.g., from FP32 to FP16 or INT8), sacrificing precision for increased inference speed. Reducing batch size means decreasing the data batch size for each inference operation to reduce data waiting time and inter-batch synchronization latency. Migrating bottleneck-level computational tasks to higher-performance computing units means moving bottleneck-level computational tasks from the current computing unit (e.g., CPU) to a higher-performance computing unit (e.g., GPU or NPU). Deep optimization configurations may take several operation cycles to take effect, but the adjustments are more significant and can fundamentally alleviate persistent bottlenecks.

[0101] The three methods described above can be used individually or in combination. For example, migrating computing units and reducing batch size simultaneously, within the limits of power consumption, can achieve a greater reduction in latency. Alternatively, in scenarios where the task does not require high accuracy, the inference accuracy level can be reduced first; if this still does not meet real-time requirements, the computing task can be further migrated. Deep optimization may take several operation cycles to take effect, but the adjustment range is larger, effectively alleviating persistent bottlenecks. Once the bottleneck level returns to normal, the system can gradually revert the above configuration to its original level to reduce power consumption or restore model accuracy.

[0102] As can be seen, by implementing another resource allocation method based on time constant as shown in the embodiments of this application, the intelligent control device, based on real-time measurement of time constant at the link level and closed-loop scheduling of bottlenecks, further automatically discovers the resource coupling relationship between levels through correlation analysis and performs joint scheduling, avoiding the seesaw effect where the overall performance decreases instead of increasing after optimizing one level; at the same time, by triggering deep optimization configuration processing when bottlenecks exist continuously, the stubborn bottlenecks are fundamentally alleviated, and the system resource utilization and task real-time performance are continuously improved.

[0103] The method of this application has been described above; the apparatus of this application will be described below.

[0104] Please see Figure 5 , Figure 5 This is a schematic diagram of a resource allocation device based on a time constant, provided as an embodiment of this application. Figure 5 As shown, the time constant-based resource allocation device 50 is applied to an embodied intelligent control device, including: Deployment module 501 is used to deploy timestamp probes at the input / output boundaries of each processing level of the embodied intelligent control device.

[0105] The calculation module 502 is used to calculate the time constant of each processing level based on the timestamp difference between adjacent probes, obtain the time constant spectrum vector, and update the statistical data of the spectrum vector based on the sliding window.

[0106] The identification module 503 is used to perform bottleneck determination processing on each processing level based on statistical data, and to identify the bottleneck level that is currently limiting performance.

[0107] The allocation module 504 is used to perform resource allocation biasing at the bottleneck level and restore the resource balance state after the bottleneck level is transferred.

[0108] It should be noted that, Figure 5 For any content not mentioned in the corresponding embodiments, please refer to the description of the foregoing method embodiments, which will not be repeated here.

[0109] As can be seen, the device shown in this application can achieve real-time observability of the time constant at each processing level by deploying timestamp probes at the input and output boundaries of each processing level, and automatically identify bottleneck levels based on statistical data, drive resource tilting and allocation in a closed loop, so that computing resources are always allocated to the slowest link that currently limits the system performance, thereby improving the resource utilization of the embodied intelligent control device.

[0110] This application also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the method as described in the foregoing embodiments.

[0111] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0112] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A resource allocation method based on time constant, characterized in that, Applied to embodied intelligent control devices, the method includes: Deploy timestamp probes at the input / output boundaries of each processing level of the embodied intelligent control device; The time constant of each processing level is calculated based on the timestamp difference between adjacent probes to obtain the time constant spectrum vector, and the statistical data of the spectrum vector is updated based on the sliding window. Based on the statistical data, bottleneck determination is performed on each processing level to identify the current bottleneck level that is limiting performance; Resource allocation is tilted to the bottleneck level, and resource balance is restored after the bottleneck level is transferred.

2. The method according to claim 1, characterized in that, The processing layers include: perception layer, cognition layer, decision-making layer, and execution layer; The deployment of timestamp probes at the input / output boundaries of each processing level of the embodied intelligent control device includes: A first timestamp probe is deployed at the input end of the perception layer to record the first timestamp when the raw data is input; A second timestamp probe is deployed between the output of the perception layer and the input of the cognition layer to record the second timestamp when the perception result is generated; A third timestamp probe is deployed between the output of the cognitive layer and the input of the decision layer to record a third timestamp when the cognitive result is generated; A fourth timestamp probe is deployed between the output of the decision layer and the input of the execution layer to record the fourth timestamp when the decision result is generated; A fifth timestamp probe is deployed at the output of the execution layer to record the fifth timestamp when the execution result is generated; Each timestamp probe reads the current time from the global synchronization clock, attaches the timestamp to the data packet, and passes it on.

3. The method according to claim 2, characterized in that, The time constant spectrum vector includes the perception layer time constant, the cross-layer communication time constant from the perception layer to the cognition layer, the cognition layer time constant, the cross-layer communication time constant from the cognition layer to the decision layer, the decision layer time constant, the cross-layer communication time constant from the decision layer to the execution layer, and the execution layer time constant. The sliding window is used to retain the time constant spectrum vector of a preset number of operation cycles; The statistical data includes the mean, standard deviation, and peak value.

4. The method according to claim 1, characterized in that, The step of performing bottleneck determination processing on each processing level based on the statistical data to identify the current bottleneck level limiting performance includes: The current average time constant of each level is compared with a preset delay tolerance threshold, and levels that exceed the delay tolerance threshold are marked as absolute delay bottleneck candidates. Calculate the proportion of the time constant of each level in the total latency of the entire link, and mark the levels whose proportion exceeds the preset proportion threshold as relative proportion bottleneck candidates; The trend direction of the time constant of each level in the current sliding window is compared with that of the previous window, and the level that shows an increasing trend for several consecutive periods is marked as a candidate for trend drift bottleneck. The absolute latency bottleneck candidates, the relative proportion bottleneck candidates, and the trend drift bottleneck candidates are combined and sorted from highest to lowest severity. The one ranked first is determined as the current bottleneck level that limits performance.

5. The method according to claim 4, characterized in that, The step of performing resource skew allocation on the bottleneck level and restoring resource balance after the bottleneck level shifts includes: Computational resources are allocated from non-bottleneck levels to achieve resource allocation bias towards the bottleneck level; The resource allocation includes one or more of the following: increasing the available video memory quota of the bottleneck level, increasing the scheduling priority of the computing threads of the bottleneck level, and delaying the execution of prediction tasks in non-bottleneck levels. When it is determined that the bottleneck level has shifted to another level, the skewed computing resources will be restored to a balanced state.

6. The method as described in claim 5, characterized in that, Also includes: When multiple consecutive operation cycles at the same level are identified as the bottleneck level, a deep optimization configuration process is triggered. The deep optimization configuration process includes: reducing the inference accuracy level of the bottleneck level, reducing the batch size, and migrating the computing tasks of the bottleneck level to a higher-performance computing unit.

7. The method according to claim 1, characterized in that, Also includes: Perform correlation analysis on historical time constant spectral data to monitor the degree of coupling between time constants at different levels; When it is detected that the time constants of at least two levels exhibit a pattern of mutual increase or decrease, it is determined that there is a resource coupling relationship between the at least two levels, and the levels with the coupling relationship are marked as joint scheduling units; When any level in the joint scheduling unit is identified as a bottleneck, resource allocation is performed simultaneously on at least two levels in the joint scheduling unit.

8. A resource allocation device based on a time constant, characterized in that, Applied to embodied intelligent control devices, the device includes: The deployment module is used to deploy timestamp probes at the input / output boundaries of each processing level of the embodied intelligent control device; The calculation module is used to calculate the time constant of each processing level based on the timestamp difference between adjacent probes, obtain the time constant spectrum vector, and update the statistical data of the spectrum vector based on a sliding window; The identification module is used to perform bottleneck determination processing on each processing level based on the statistical data, and to identify the bottleneck level that is currently limiting performance. The allocation module performs resource allocation bias to the bottleneck level and restores the resource balance state after the bottleneck level is transferred.

9. A personal intelligent control device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which, when executed by a processor, implements the method as described in any one of claims 1 to 7.