Multi-type task scheduling system and method based on hierarchical self intelligence

Abstract

The invention discloses a hierarchical intelligent multi-type task scheduling system and method, and the system comprises a runtime situation perception analysis module which is used for collecting runtime situation information in a cloud edge collaborative streaming data processing system in real time, and analyzing and extracting runtime situation features; the hierarchical intelligent scheduling decision module is used for generating a system multi-dimensional configuration knob scheduling decision by adopting macro scheduling and micro scheduling according to the situation characteristics during operation; and the stream data processing control module is used for updating system configuration according to the scheduling decision and controlling the analysis processing flow of the stream data on the cloud edge distributed equipment. According to the method, flow data task scheduling of sensing, calculation and control integration is realized based on hierarchical body intelligence, through cooperation of a deep reinforcement learning network and a plurality of independent negative feedback adjustment algorithms, an exponential decision space in a multi-system knob scheduling scene is decomposed into two stages of coarse-grained decision and fine-grained decision, the decision complexity is reduced, and the decision efficiency is improved. And the model convergence difficulty is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of cloud-edge distributed collaborative stream data scheduling and processing, and specifically relates to a multi-type task scheduling system and method based on hierarchical embodied intelligence. Background Art

[0002] In recent years, with the widespread adoption and deployment of various sensor devices, such as cameras, in daily life, the demand for stream data content analysis based on artificial intelligence algorithms has continued to grow. Among these, analysis of real-time video streams is crucial for a variety of critical applications, such as road monitoring, virtual reality, and driver assistance. Typical video analytics applications involve multi-stage processing. To ensure a high-quality user experience, these applications require low latency and high accuracy. However, in real-world open environments, the dynamic instability of scenes poses significant challenges to real-time stream data analysis. Network fluctuations and instabilities in the spatial dimension, as well as mismatches in processing speed between previous and subsequent stages in the temporal dimension, can lead to task backlogs, severely degrading the performance of real-time stream data analysis. Therefore, there is an urgent need to alleviate task backlogs and improve stream data analysis performance by dynamically adjusting key adjustable system parameters (i.e., "knobs" such as data configuration and offload locations).

[0003] Existing research attempts to solve the above problems, mainly including the following three types of stream data scheduling methods:

[0004] 1) Configuration Optimization-Based Approach: This approach extracts the performance of different system knob combinations from offline stream data analysis history, stores them as configuration files, and then uses optimization algorithms to search for the optimal knob configuration combination in the current state during online stream data analysis to form stream data scheduling decisions. However, this approach relies heavily on offline historical data and struggles to effectively adapt to the dynamic fluctuations of online operating scenarios. Furthermore, as the number of adjustable knobs in the system increases, the optimization complexity of the configuration file search increases exponentially, making it difficult to maintain good scheduling results in real-time stream data processing systems.

[0005] 2) End-to-end reinforcement learning-based approach: This approach utilizes an end-to-end deep reinforcement learning network to learn the optimal knob adjustment strategy through interaction with the environment. This implicit experience is encoded in the neural network parameters, and pre-trained model parameters are loaded during inference to output scheduling decisions. However, this approach relies on a single network to determine the adjustment of all knobs. As the number of knobs increases, the model convergence becomes significantly more difficult, making it difficult to reach the global optimal state.

[0006] 3) Negative feedback-based approach: This method adjusts knobs inversely based on the performance deviation of the current streaming data analysis task (e.g., the difference between actual and expected latency), achieving low-overhead, real-time response. However, this approach only works for independent decisions made by a single knob and struggles to effectively handle the complex problem of coordinated adjustments of multiple knobs. Even if a separate negative feedback mechanism is assigned to each knob, the lack of a global collaborative perspective leads to low overall performance.

[0007] Based on the above considerations, existing methods are insufficient in coping with the high dynamics of open environments and the complexity of multi-knob collaborative optimization. Therefore, it is urgent to propose a streaming data analysis system scheduling solution with high adaptability to dynamic environments and low decision complexity to significantly improve the system's latency compliance rate and stability when processing streaming data tasks in unstable scenarios. Summary of the Invention

[0008] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a multi-type task scheduling system and method based on hierarchical embodied intelligence to solve the problems of high decision-making complexity and poor dynamic adaptability of existing streaming data scheduling methods.

[0009] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0010] The present invention provides a multi-type task scheduling system based on hierarchical embodied intelligence, comprising: a runtime context awareness analysis module, a hierarchical embodied intelligence scheduling decision module, and a stream data processing control module;

[0011] The runtime context-aware analysis module is used to collect runtime context information in the cloud-edge collaborative stream data processing system in real time, analyze and extract runtime context features;

[0012] A hierarchical embodied intelligent scheduling decision module is used to generate system multi-dimensional configuration knob scheduling decisions based on runtime context characteristics using macro-scheduling and micro-scheduling;

[0013] The stream data processing control module is used to update the system configuration according to the scheduling decision and control the processing flow of stream data in the cloud-edge collaborative stream data processing system.

[0014] Furthermore, the runtime context awareness analysis module includes:

[0015] Runtime context awareness module, used to collect resource context, task context, and decision context information in the cloud-edge collaborative stream data processing system;

[0016] The runtime context analysis module is used to analyze the resource context, task context, and decision context information collected by the runtime context awareness module and extract runtime context features.

[0017] Furthermore, the runtime context awareness module includes: a resource context monitor, a task context extractor, and a decision context memory;

[0018] Resource context monitor, used to collect resource context information of cross-cloud-edge network communication bandwidth in cloud-edge collaborative stream data processing system in real time R,t ;

[0019] The task context extractor is used to extract task context information C including task execution delay, number of targets in the task, and target size from the cloud-edge collaborative stream data processing system in real time. T,t ;

[0020] Decision context storage, used to collect decision context information C in the cloud-edge collaborative stream data processing system D,t .

[0021] Furthermore, the analysis process of the runtime context analysis module is specifically as follows:

[0022] Save resource context information, task context information, and decision context information to the corresponding context information queue;

[0023] Each context information queue is updated using a sliding window method, and the length of all context information queues is maintained at 2l. When the number of context information in the context information queue reaches the upper limit and new context information arrives, the oldest context information stored in the context information queue is removed and the newly arrived context information is added to the context information queue;

[0024] Each context information queue is sampled in an evenly spaced manner, and the length of each context information sequence obtained by sampling is l;

[0025] The sampled context information sequences are divided into three groups according to resource context, task context, and decision context. Each group uses an independent one-dimensional convolutional neural network to extract features and generate feature embedding vectors representing resources, tasks, and decisions.

[0026] The three feature embedding vectors are cross-fused using a fully connected neural network layer to extract a low-dimensional feature embedding vector s t , which represents the analysis results of the runtime scenario of the current cloud-edge collaborative stream data processing system.

[0027] Furthermore, the hierarchical embodied intelligent scheduling decision module includes: a macro scheduling module and a micro scheduling module;

[0028] The macro scheduling module is used to extract the runtime context features s from the runtime context awareness analysis module. t , using a deep reinforcement learning network to output the coarse-grained decision ξ of the system's multidimensional configuration knob t ;

[0029] The micro-scheduling module is used to make coarse-grained decisions based on the output of the macro-scheduling module. t , for each scheduling knob (i.e., an adjustable variable containing several values), an independent negative feedback adjustment algorithm is used to calculate the fine-grained decision τ of the output configuration knob t , and serves as the decision output of the hierarchical embodied intelligent scheduling decision module.

[0030] Furthermore, the decision-making process of the macro-scheduling module is as follows:

[0031] Load the pre-trained parameters W into the deep reinforcement learning network θ , and outputs a coarse-grained decision ξ for the system's multi-dimensional configuration knob based on the input runtime context characteristics and network parameters. t , which contains the decisions of monotonic knobs and non-monotonic knobs, and outputs a value of -1, 0, or 1 for each knob;

[0032] For the monotonicity knob (Knob performance changes linearly with knob value, such as task end-to-end latency increases with video frame resolution). Deep reinforcement learning network output monotonicity adjustment direction Output Represents a decrease, Represents unchanged, Represents elevation;

[0033] For non-monotonic knobs (Knob performance varies nonlinearly with knob value. For example, the end-to-end latency of a task has no linear relationship with the offload location and depends on multiple factors such as network conditions and node computing power.) The deep reinforcement learning network outputs preliminary coarse-grained adjustments. For collaborative split point decision in task offloading, Indicates the offloading decision of multi-stage processing tasks on the edge and cloud. It corresponds to three split point values ​​0, 1, and 2, which represent the subscript of the array composed of multi-stage processing tasks. The processing tasks on the left side of the split point are offloaded to the edge for execution, and the processing tasks on the right side of the split point are offloaded to the cloud for execution. When the streaming data task has two processing stages (detection stage and classification stage), (Split point is 0) means that both the detection stage and the classification stage are performed in the cloud. (Split point 1) means the detection phase is performed on the edge and the classification phase is performed on the cloud. (The split point is 2) means that both the detection stage and the classification stage are performed on the edge;

[0034] The decisions of all monotonic knobs and non-monotonic knobs in the output results of the macro scheduling module constitute the coarse-grained decision, that is,

[0035] Furthermore, the decision-making process of the micro-scheduling module is as follows:

[0036] According to the coarse-grained decision ξ output by the macro scheduling module t , assign a separate negative feedback adjustment algorithm to each knob, and calculate the final adjustment value of each knob as the fine-grained decision τ t , which contains the decisions of monotonic knob and non-monotonic knob;

[0037] For the monotonicity knob Negative feedback adjustment algorithm calculates fine-grained decision values ​​based on the AIMD (aggregate growth, multiplicative reduction) principle The calculation formula is as follows:

[0038]

[0039] in, Knobs arranged in ascending order according to linear influence An array of values, For knob In the current task T t The value subscript on , For knob In the previous task T t-1 The value subscript on ;

[0040] For non-monotonic knobs Negative feedback adjustment algorithm calculates fine-grained decision values ​​based on the performance of the previous task The calculation formula is as follows:

[0041]

[0042] in, Display knob The coarse-grained decision value, n t Indicates the current task T t The number of regions of interest detected in represents the number of interest regions offloaded to the cloud server when the classification stage is offloaded to the cloud for execution (i.e., all interest regions are offloaded to the cloud server for processing), μ represents the edge node number, and U represents the maximum edge node number. represents the number of interest regions offloaded to the edge node μ when the classification stage is offloaded to the edge execution, Indicates the previous task T t-1 Analysis time at edge node μ;

[0043] The decisions of all monotonic knobs and non-monotonic knobs in the output results of the micro-scheduling module constitute fine-grained decisions as the decision results of the hierarchical embodied intelligent scheduling decision module, namely

[0044] Furthermore, the pre-trained parameter W in the macro scheduling module θ The learning process is:

[0045] (11) Based on the runtime context information collected by the runtime context awareness analysis module, extract the stream data task T t The runtime context characteristics at that moment t ;

[0046] (12) According to the runtime context characteristics s t , through the hierarchical embodied intelligent scheduling decision module, the coarse-grained decision ξ and fine-grained decision τ of the system's multi-dimensional configuration knob are output t ;

[0047] (13) According to the decision result τ t Adjust the system knob to schedule the current streaming data task T t Execute in the stream data processing control module and get task T t Execution results, and based on the runtime context awareness analysis module to collect runtime context information, extract the stream data task T t+1 The runtime context at that moment t+1 ;

[0048] (14) According to task T t Execution result calculation scheduling decision reward value r t , the reward value calculation formula is as follows:

[0049]

[0050] Among them, Δ L,t Represents task T t The deviation between the delay constraint and the actual end-to-end execution delay, f t Indicates the current data source frame rate used to calculate the task delay constraint, L t Represents the end-to-end delay result of the task, A t Represents the accuracy results of the task, α, β, θ and γ are preset hyperparameters that control the dimension and range;

[0051] (15) According to the quadruple of Markov decision process MDP t ,ξ t ,s t+1 ,r t >, update the deep reinforcement learning network parameters W θ ;

[0052] (16) Repeat steps (11) to (15) until the deep reinforcement learning network parameter W​θ convergence.

[0053] Furthermore, the stream data processing control module includes: a task generator, a control forwarder, an application processor and a storage distributor;

[0054] The task generator is deployed on the edge node and is used to obtain real-time streaming data from the streaming data source address. It preprocesses and packages the streaming data according to the scheduling decision output by the hierarchical embodied intelligent scheduling decision module. Specifically, it adjusts the quality, frequency, and encoding configuration of the streaming data, packages it into a streaming data file, uses the data configuration decision and task offloading decision information in the scheduling decision as metadata information, and packages the streaming data file and metadata information into a streaming data task package.

[0055] The control forwarder is deployed on all edge nodes and cloud servers and is used to forward the stream data task package according to the metadata information in the stream data task package. The forwarding logic is executed based on the current stage of the stream data task package as follows: if the unloading node of the current stage is the same as the node where the control forwarder is located, the stream data task package is forwarded to the corresponding stage application processor on the current node; if the unloading node of the current stage of the stream data task package is different from the node where the control forwarder is located, the stream data task package is forwarded to the control forwarder of the unloading node; if all current stages of the stream data task package have been processed, the stream data task package is forwarded to the storage distributor on the cloud server;

[0056] Application processors, deployed on all edge nodes and cloud servers, are used to analyze streaming data task packets and store the task processing results in the streaming data task packets. Each application processor handles a single stage of streaming data analysis and performs analysis on the streaming data task packets based on an encapsulated neural network algorithm.

[0057] The storage distributor is deployed on the cloud server and is used to collect streaming data tasks that have been processed in all stages, store the task processing results in the streaming data task package into the database, and distribute them to the runtime context-aware analysis module and the hierarchical embodied intelligent scheduling decision module.

[0058] Furthermore, the processing process of the stream data processing control module is specifically as follows:

[0059] (21) The task generator binds the stream data source address and obtains real-time stream data from the stream data source address;

[0060] (22) The task generator requests scheduling decision information from the hierarchical embodied intelligent scheduling decision module (the scheduling decision information includes data configuration decision and task offloading decision), and pre-processes and packages the stream data according to the data configuration decision in the obtained scheduling decision information to form a stream data task package, and stores the scheduling decision information in the metadata information of the stream data task package;

[0061] (23) According to the task offloading decision of the first processing stage in the metadata information in the stream data task packet, the task generator forwards the current stream data task packet to the control forwarder on the corresponding device for processing;

[0062] (24) After the control forwarder obtains the stream data task packet, it forwards the stream data task packet according to the unloading decision in the metadata information in the stream data task packet;

[0063] If the unloading node of the current stage of the stream data task packet is the same as the node where the control forwarder is located, the stream data task packet is forwarded to the corresponding stage application processor on the current node, the stream data analysis of this stage is completed, and the process goes to step (25);

[0064] If the unloading node of the current stage of the stream data task packet is different from the node where the control forwarder is located, the stream data task packet is forwarded to the control forwarder of the unloading node, and the process returns to step (24);

[0065] If all current stages of the stream data task package have been processed, the stream data task package is forwarded to the storage distributor on the cloud server and the process goes to step (26);

[0066] (25) The application processor processes the corresponding stream data analysis application according to the obtained stream data task packet, stores the processing result in the stream data task packet, forwards it to the control forwarder on the current node, and returns to step (24);

[0067] (26) The storage distributor stores the obtained stream data task package in the database and distributes the task processing result information to the runtime context awareness analysis module and the hierarchical embodied intelligent scheduling decision module;

[0068] (27) Repeat steps (21) to (26) until the stream data analysis is completed.

[0069] The present invention also provides a multi-type task scheduling method based on hierarchical embodied intelligence. Based on the above system, the method includes the following steps:

[0070] 1) Collect resource context, task context, and decision context information, and analyze them to obtain runtime context characteristics s t ;

[0071] 2) According to the runtime context characteristics s t , using macro-scheduling and micro-scheduling, outputting the coarse-grained decision of the system's multi-dimensional configuration knob ξ t and fine-grained decision τ t ;

[0072] 3) According to the current fine-grained decision τ t, execute the current stream data task T in the cloud-edge collaborative stream data processing system t and distribute the results to the runtime context-aware analysis module and the hierarchical embodied intelligent scheduling decision module;

[0073] 4) Repeat steps 1) to 3) until the stream data analysis is completed.

[0074] Beneficial effects of the present invention:

[0075] 1. Low decision-making overhead: This invention decomposes the exponential decision space in the multi-system knob scheduling scenario into two stages of coarse-grained decision-making and fine-grained decision-making through the collaboration of a deep reinforcement learning network in the macro-scheduling module and multiple independent negative feedback adjustment algorithms in the micro-scheduling module, thereby reducing decision-making complexity, thereby reducing the difficulty of model convergence and shortening scheduling time overhead.

[0076] 2. High dynamic adaptability: The present invention applies embodied intelligence to the scheduling of streaming data analysis systems. By perceiving the real-time runtime context of the system, calculating the scheduling decisions of the system knobs, and controlling the streaming data analysis and processing flow, a closed-loop scheduling loop with integrated sensing, computing, and control is formed, thus realizing dynamic iterative optimization of the Markov decision process and adapting to the dynamic and complex changes of the scene environment.

[0077] 3. High latency compliance rate: The present invention implements integrated sensing, computing and control streaming data task scheduling processing based on hierarchical embodied intelligence, which can realize adaptive adjustment of system configuration in unstable scenarios, thereby ensuring stable task execution performance and good user experience.

[0078] 4. Low long-tail latency: Through a hierarchical, intelligent streaming data scheduling mechanism, the present invention can maintain stable task processing performance in long-term streaming data analysis, and make timely adjustments under extreme load scenarios to ensure the service level indicators of streaming data analysis applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 This is a system architecture diagram of the present invention;

[0080] Figure 2 This is a structural diagram of the runtime context awareness analysis module in the present invention;

[0081] Figure 3 This is a structural diagram of the runtime context awareness module in the present invention;

[0082] Figure 4 This is a structural diagram of the runtime situation analysis module in the present invention;

[0083] Figure 5 This is a structural diagram of the hierarchical embodied intelligent scheduling decision module in the present invention;

[0084] Figure 6This is a decision flow chart of the macro scheduling module in the present invention;

[0085] Figure 7 This is a decision flow chart of the micro-scheduling module in the present invention;

[0086] Figure 8 This is a flowchart of network parameter learning for the macro scheduling module in the present invention;

[0087] Figure 9 This is a structural diagram of the stream data processing control module in the present invention;

[0088] Figure 10 This is a flow chart of stream data processing control in the present invention. DETAILED DESCRIPTION

[0089] To facilitate understanding by those skilled in the art, the present invention will be further described below using a streaming video analysis application case as an example in combination with embodiments and drawings. The contents mentioned in the implementation manner are not intended to limit the present invention.

[0090] Reference Figure 1 As shown, a multi-type task scheduling system based on hierarchical embodied intelligence of the present invention includes: a runtime context awareness analysis module, a hierarchical embodied intelligence scheduling decision module, and a stream data processing control module;

[0091] The runtime context-aware analysis module is used to collect runtime context information in the cloud-edge collaborative stream data processing system in real time, analyze and extract runtime context features;

[0092] A hierarchical embodied intelligent scheduling decision module is used to generate system multi-dimensional configuration knob scheduling decisions based on runtime context characteristics using macro-scheduling and micro-scheduling;

[0093] The stream data processing control module is used to update the system configuration according to the scheduling decision and control the processing flow of stream data in the cloud-edge collaborative stream data processing system.

[0094] Reference Figure 2 As shown, the runtime context awareness analysis module includes:

[0095] Runtime context awareness module, used to collect resource context, task context, and decision context information in the cloud-edge collaborative stream data processing system;

[0096] The runtime context analysis module is used to analyze the resource context, task context, and decision context information collected by the runtime context awareness module and extract the characteristics of the runtime context.

[0097] Reference Figure 3 As shown, the runtime context awareness module includes: a resource context monitor, a task context extractor, and a decision context memory;

[0098] Resource context monitor, used to collect resource context information of cross-cloud-edge network communication bandwidth in cloud-edge collaborative stream data processing system in real time R,t ;

[0099] The task context extractor is used to extract task context information C including task execution delay, number of targets in the task, and target size from the cloud-edge collaborative stream data processing system in real time. T,t ;

[0100] Decision context storage, used to collect decision context information C in the cloud-edge collaborative stream data processing system D,t .

[0101] Reference Figure 4 As shown, the analysis process of the runtime situation analysis module is specifically as follows:

[0102] Save resource context information, task context information, and decision context information to the corresponding context information queue;

[0103] Each context information queue is updated using a sliding window method, and the length of all context information queues is maintained at 2l. When the number of context information in the context information queue reaches the upper limit and new context information arrives, the oldest context information stored in the context information queue is removed and the newly arrived context information is added to the context information queue;

[0104] Each context information queue is sampled in an evenly spaced manner, and the length of each context information sequence obtained by sampling is l;

[0105] The sampled context information sequences are divided into three groups according to resource context, task context, and decision context. Each group uses an independent one-dimensional convolutional neural network to extract features and generate feature embedding vectors representing resources, tasks, and decisions.

[0106] The three feature embedding vectors are cross-fused using a fully connected neural network layer to extract a low-dimensional feature embedding vector s t , which represents the analysis results of the runtime scenario of the current cloud-edge collaborative stream data processing system.

[0107] Reference Figure 5 As shown, the hierarchical embodied intelligent scheduling decision module includes: a macro scheduling module and a micro scheduling module;

[0108] The macro scheduling module is used to extract the runtime context features s from the runtime context awareness analysis module. t , using a deep reinforcement learning network to output the coarse-grained decision ξ of the system's multidimensional configuration knob t ;

[0109] The micro-scheduling module is used to make coarse-grained decisions based on the output of the macro-scheduling module. t , for each scheduling knob (i.e., an adjustable variable containing several values), an independent negative feedback adjustment algorithm is used to calculate the fine-grained decision τ of the output configuration knob t , and serves as the decision output of the hierarchical embodied intelligent scheduling decision module.

[0110] In the example, refer to Figure 6 As shown in FIG, the decision-making process of the macro scheduling module is:

[0111] Load the pre-trained parameters W into the deep reinforcement learning network θ , and outputs a coarse-grained decision ξ for the system's multi-dimensional configuration knob based on the input runtime context characteristics and network parameters. t , which contains the decisions of monotonic knobs and non-monotonic knobs, and outputs a value of -1, 0, or 1 for each knob;

[0112] For the monotonicity knob (Knob performance changes linearly with knob value, such as task end-to-end latency increases with video frame resolution). Deep reinforcement learning network output monotonicity adjustment direction Output Represents a decrease, Represents unchanged, Represents an increase; for the monotonic knob (Knob performance changes linearly with knob value, such as task end-to-end latency increases with video frame resolution). Deep reinforcement learning network output monotonicity adjustment direction Output Represents a decrease, Represents unchanged, Representing the increase, the coarse-grained decision of the monotonicity knob includes the resolution decision Frame rate decision Frame Packet Size Decision

[0113] For non-monotonic knobs (Knob performance varies nonlinearly with knob value. For example, the end-to-end latency of a task has no linear relationship with the offload location and depends on multiple factors such as network conditions and node computing power.) The deep reinforcement learning network outputs preliminary coarse-grained adjustments. For collaborative split point decision in task offloading, Indicates the offloading decision of multi-stage processing tasks on the edge and cloud. It corresponds to three split point values ​​0, 1, and 2, which represent the subscript of the array composed of multi-stage processing tasks. The processing tasks on the left side of the split point are offloaded to the edge for execution, and the processing tasks on the right side of the split point are offloaded to the cloud for execution. When the streaming data task has two processing stages (detection stage and classification stage), (Split point is 0) means that both the detection stage and the classification stage are performed in the cloud. (Split point 1) means the detection phase is performed on the edge and the classification phase is performed on the cloud. (The split point is 2) means that both the detection stage and the classification stage are performed on the edge;

[0114] The decisions of all monotonic knobs and non-monotonic knobs in the output results of the macro scheduling module constitute the coarse-grained decision, that is,

[0115] In the example, refer to Figure 7 As shown in FIG, the decision-making process of the micro-scheduling module is:

[0116] According to the coarse-grained decision ξ output by the macro scheduling module t ,The micro-scheduling module assigns a separate negative feedback adjustment algorithm to each knob, and calculates the final adjustment value of each knob as the fine-grained decision τ t , which includes the decisions of monotonic and non-monotonic knobs;

[0117] For the monotonicity knob Negative feedback adjustment algorithm calculates fine-grained decision values ​​based on the AIMD (aggregate growth, multiplicative reduction) principle Monotonicity knob fine-grained decisions including resolution decisions Frame rate decision The frame packet size is determined by the following calculation formula:

[0118]

[0119] in, Knobs arranged in ascending order according to linear influence An array of values, For knob In the current task T t The value subscript on , For knob In the previous task T t-1 The value subscript on ;

[0120] For non-monotonic knobs Negative feedback adjustment algorithm calculates fine-grained decision values ​​based on the performance of the previous task Non-monotonic knob fine-grained decision-making includes cloud-edge collaborative split point decision-making Region of interest allocation decision in, Depends on The calculation formula is as follows:

[0121]

[0122] Among them, n t Indicates the current task T t The number of regions of interest detected in represents the number of ROIs offloaded to the cloud server when the classification stage is offloaded to the cloud for execution (i.e., all ROIs are offloaded to the cloud server for processing), represents the number of interest regions offloaded to the edge node μ when the classification stage is offloaded to the edge execution, Indicates the previous task T t-1 Analysis time at edge node μ;

[0123] The decisions of all monotonic knobs and non-monotonic knobs in the output results of the micro-scheduling module constitute fine-grained decisions as the decision results of the hierarchical embodied intelligent scheduling decision module, namely

[0124] In the example, refer to Figure 8 As shown, the pre-trained parameters W in the macro scheduling module θ The learning process is:

[0125] (11) Based on the runtime context information collected by the runtime context awareness analysis module, extract the stream data task T t The runtime context characteristics at that moment t ;

[0126] (12) According to the runtime context characteristics s t , outputting the coarse-grained decision of the system's multi-dimensional configuration knob through the hierarchical embodied intelligent scheduling decision module ξ t and fine-grained decision τ t ;

[0127] (13) According to the decision result τ t Adjust the system knob to schedule the current streaming data task T t Execute in the stream data processing control module and get task T t Execution results, and based on the runtime context awareness analysis module to collect runtime context information, extract the stream data task T t+1 The runtime context at that moment t+1 ;

[0128] (14) According to task T t Execution result calculation scheduling decision reward value r t , the reward value calculation formula is as follows:

[0129]

[0130] Among them, Δ L,t Represents task T t The deviation between the delay constraint and the actual end-to-end execution delay, f t Indicates the current data source frame rate used to calculate the task delay constraint, L t Represents the end-to-end delay result of the task, A t Represents the accuracy results of the task, α, β, θ and γ are preset hyperparameters that control the dimension and range;

[0131] (15) According to the quadruple of Markov decision process MDP t ,ξ t ,s t+1 ,r t >, update the deep reinforcement learning network parameters W θ ;

[0132] (16) Repeat steps (11) to (15) until the deep reinforcement learning network parameter W θ convergence.

[0133] Reference Figure 9 As shown, the stream data processing control module includes: a task generator, a control forwarder, an application processor and a storage distributor;

[0134] The task generator is deployed on the edge node and is used to obtain real-time streaming data from the streaming data source address. It preprocesses and packages the streaming data according to the scheduling decision output by the hierarchical embodied intelligent scheduling decision module. Specifically, it adjusts the quality, frequency, and encoding configuration of the streaming data, packages it into a streaming data file, uses the data configuration decision and task offloading decision information in the scheduling decision as metadata information, and packages the streaming data file and metadata information into a streaming data task package.

[0135] ​The control forwarder is deployed on all edge nodes and cloud servers and is used to forward the stream data task package according to the metadata information in the stream data task package. The forwarding logic is executed based on the current stage of the stream data task package as follows: if the unloading node of the current stage is the same as the node where the control forwarder is located, the stream data task package is forwarded to the corresponding stage application processor on the current node; if the unloading node of the current stage of the stream data task package is different from the node where the control forwarder is located, the stream data task package is forwarded to the control forwarder of the unloading node; if all current stages of the stream data task package have been processed, the stream data task package is forwarded to the storage distributor on the cloud server;

[0136] Application processors, deployed on all edge nodes and cloud servers, are used to analyze streaming data task packets and store the task processing results in the streaming data task packets. Each application processor handles a single stage of streaming data analysis and performs analysis on the streaming data task packets based on an encapsulated neural network algorithm.

[0137] The storage distributor is deployed on the cloud server and is used to collect streaming data tasks that have been processed in all stages, store the task processing results in the streaming data task package into the database, and distribute them to the runtime context-aware analysis module and the hierarchical embodied intelligent scheduling decision module.

[0138] Reference Figure 10 As shown, the processing process of the stream data processing control module is specifically as follows:

[0139] (21) The task generator binds the stream data source address and obtains real-time stream data from the stream data source address;

[0140] (22) The task generator requests scheduling decision information from the hierarchical embodied intelligent scheduling decision module (the scheduling decision information includes data configuration decision and task offloading decision), and pre-processes and packages the stream data according to the data configuration decision in the obtained scheduling decision information to form a stream data task package, and stores the scheduling decision information in the metadata information of the stream data task package;

[0141] (23) According to the task offloading decision of the first processing stage in the metadata information in the stream data task packet, the task generator forwards the current stream data task packet to the control forwarder on the corresponding device for processing;

[0142] (24) After the control forwarder obtains the stream data task packet, it forwards the stream data task packet according to the unloading decision in the metadata information in the stream data task packet;

[0143] If the unloading node of the current stage of the stream data task packet is the same as the node where the control forwarder is located, the stream data task packet is forwarded to the corresponding stage application processor on the current node, the stream data analysis of this stage is completed, and the process goes to step (25);

[0144] If the unloading node of the current stage of the stream data task packet is different from the node where the control forwarder is located, the stream data task packet is forwarded to the control forwarder of the unloading node, and the process returns to step (24);

[0145] If all current stages of the stream data task package have been processed, the stream data task package is forwarded to the storage distributor on the cloud server and the process goes to step (26);

[0146] (25) The application processor processes the corresponding stream data analysis application according to the obtained stream data task packet, stores the processing result in the stream data task packet, forwards it to the control forwarder on the current node, and returns to step (24);

[0147] (26) The storage distributor stores the obtained stream data task package in the database and distributes the task processing result information to the runtime context awareness analysis module and the hierarchical embodied intelligent scheduling decision module;

[0148] (27) Repeat steps (21) to (26) until the stream data analysis is completed.

[0149] In addition, the present invention also provides a multi-type task scheduling method based on hierarchical embodied intelligence. Based on the above system, the method includes the following steps:

[0150] 1) Collect resource context, task context, and decision context information, and analyze them to obtain runtime context characteristics s t ;

[0151] 2) According to the runtime context characteristics s t , using macro-scheduling and micro-scheduling, outputting the coarse-grained decision of the system's multi-dimensional configuration knob ξ t and fine-grained decision τ t ;

[0152] 3) According to the current fine-grained decision τ t , execute the current stream data task T in the cloud-edge collaborative stream data processing system t and distribute the results to the runtime context-aware analysis module and the hierarchical embodied intelligent scheduling decision module;

[0153] 4) Repeat steps 1) to 3) until the stream data analysis is completed.

[0154] The present invention has many specific application paths. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements can be made without departing from the principles of the present invention. These improvements should also be considered as the scope of protection of the present invention.

Claims

1. A multi-type task scheduling system based on hierarchical embodied intelligence, characterized by: include: Runtime context-aware analysis module, hierarchical embodied intelligent scheduling decision module, and stream data processing and control module; The runtime context-aware analysis module is used to collect runtime context information in the cloud-edge collaborative stream data processing system in real time, analyze and extract runtime context features; A hierarchical embodied intelligent scheduling decision module is used to generate system multi-dimensional configuration knob scheduling decisions based on runtime context characteristics using macro-scheduling and micro-scheduling; The stream data processing control module is used to update the system configuration according to the scheduling decision and control the processing flow of stream data in the cloud-edge collaborative stream data processing system.

2. The multi-type task scheduling system based on hierarchical embodied intelligence according to claim 1 is characterized in that: The runtime context awareness analysis module includes: Runtime context awareness module, used to collect resource context, task context, and decision context information in the cloud-edge collaborative stream data processing system; The runtime context analysis module is used to analyze the resource context, task context, and decision context information collected by the runtime context awareness module and extract runtime context features.

3. The multi-type task scheduling system based on hierarchical embodied intelligence according to claim 2 is characterized in that: The runtime context awareness module includes: a resource context monitor, a task context extractor, and a decision context memory; Resource context monitor, used to collect resource context information of cross-cloud-edge network communication bandwidth in cloud-edge collaborative stream data processing system in real time R,t ; The task context extractor is used to extract task context information C including task execution delay, number of targets in the task, and target size from the cloud-edge collaborative stream data processing system in real time. T,t ; Decision context storage, used to collect decision context information C in the cloud-edge collaborative stream data processing system D,t .

4. The multi-type task scheduling system based on hierarchical embodied intelligence according to claim 2 is characterized in that: The analysis process of the runtime context analysis module is specifically as follows: Save resource context information, task context information, and decision context information to the corresponding context information queue; Each context information queue is updated using a sliding window method, and the length of all context information queues is maintained at 2l. When the number of context information in the context information queue reaches the upper limit and new context information arrives, the oldest context information stored in the context information queue is removed and the newly arrived context information is added to the context information queue; Each context information queue is sampled in an evenly spaced manner, and the length of each context information sequence obtained by sampling is l; The sampled context information sequences are divided into three groups according to resource context, task context, and decision context. Each group uses an independent one-dimensional convolutional neural network to extract features and generate feature embedding vectors representing resources, tasks, and decisions. The three feature embedding vectors are cross-fused using a fully connected neural network layer to extract a low-dimensional feature embedding vector s t , which represents the analysis results of the runtime scenario of the current cloud-edge collaborative stream data processing system.

5. The multi-type task scheduling system based on hierarchical embodied intelligence according to claim 1 is characterized in that: The hierarchical embodied intelligent scheduling decision module includes: a macro scheduling module and a micro scheduling module; The macro scheduling module is used to extract the runtime context features s from the runtime context awareness analysis module. t , using a deep reinforcement learning network to output the coarse-grained decision ξ of the system's multidimensional configuration knob t ; The micro-scheduling module is used to make coarse-grained decisions based on the output of the macro-scheduling module. t , an independent negative feedback adjustment algorithm is used for each scheduling knob to calculate the fine-grained decision τ of the output configuration knob t , and serves as the decision output of the hierarchical embodied intelligent scheduling decision module.

6. The multi-type task scheduling system based on hierarchical embodied intelligence according to claim 5 is characterized in that: The decision-making process of the macro-scheduling module is: Load the pre-trained parameters W into the deep reinforcement learning network θ , and outputs a coarse-grained decision ξ for the system's multi-dimensional configuration knob based on the input runtime context characteristics and network parameters. t , which contains the decisions of monotonic knobs and non-monotonic knobs, and outputs a value of -1, 0, or 1 for each knob; For the monotonicity knob Deep reinforcement learning network output monotonicity adjustment direction Output Represents a decrease, Represents unchanged, Represents elevation; For non-monotonic knobs Preliminary coarse-grained regulation of deep reinforcement learning network outputs For collaborative split point decision in task offloading, Indicates the offloading decision of multi-stage processing tasks on the edge and cloud. It corresponds to three split points with values ​​of 0, 1, and 2, which represent the subscript of the array composed of multi-stage processing tasks. The processing tasks on the left side of the split point are offloaded to the edge for execution, and the processing tasks on the right side of the split point are offloaded to the cloud for execution. When the streaming data task has two processing stages, Indicates that both the detection and classification stages are performed in the cloud. Indicates that the detection phase is performed on the edge and the classification phase is performed on the cloud. Indicates that both the detection and classification stages are performed on the edge; The decisions of all monotonic knobs and non-monotonic knobs in the output results of the macro scheduling module constitute a coarse-grained decision.

7. The multi-type task scheduling system based on hierarchical embodied intelligence according to claim 6 is characterized in that: The decision-making process of the micro-scheduling module is: According to the coarse-grained decision ξ output by the macro scheduling module t , assign a separate negative feedback adjustment algorithm to each knob, and calculate the final adjustment value of each knob as the fine-grained decision τ t , which contains the decisions of monotonic knob and non-monotonic knob; For the monotonicity knob Negative feedback adjustment algorithm calculates fine-grained decision values ​​based on AIMD principle The calculation formula is as follows: in, Knobs arranged in ascending order according to linear influence An array of values, For knob In the current task T t The value subscript on , For knob In the previous task T t-1 The value subscript on ; For non-monotonic knobs Negative feedback adjustment algorithm calculates fine-grained decision values ​​based on the performance of the previous task The calculation formula is as follows: in, Display knob The coarse-grained decision value, n t Indicates the current task T t The number of regions of interest detected in represents the number of interest regions offloaded to the cloud server when the classification stage is offloaded to the cloud for execution, μ represents the edge node number, and U represents the maximum edge node number. represents the number of interest regions offloaded to the edge node μ when the classification stage is offloaded to the edge execution, Indicates the previous task T t-1 Analysis time at edge node μ; Based on the decisions of all monotonic knobs and non-monotonic knobs in the output results of the micro-scheduling module, fine-grained decisions are formed as the decision results of the hierarchical embodied intelligent scheduling decision module.

8. The multi-type task scheduling system based on hierarchical embodied intelligence according to claim 6 is characterized in that: The pre-trained parameter W in the macro scheduling module θ The learning process is: (11) Based on the runtime context information collected by the runtime context awareness analysis module, the runtime context feature s of the stream data task Tt is extracted t ; (12) According to the runtime context characteristics s t , outputting the coarse-grained decision of the system's multi-dimensional configuration knob through the hierarchical embodied intelligent scheduling decision module ξ t and fine-grained decision τ t ; (13) According to the decision result τ t Adjust the system knob to schedule the current streaming data task T t Execute in the stream data processing control module and get task T t Execution results, and based on the runtime context awareness analysis module to collect runtime context information, extract the stream data task T t+1 The runtime context at that moment t+1 ; (14) According to task T t Execution result calculation scheduling decision reward value r t , the reward value calculation formula is as follows: Among them, Δ L,t Represents task T t The deviation between the delay constraint and the actual end-to-end execution delay, f t Indicates the current data source frame rate used to calculate the task delay constraint, L t Represents the end-to-end delay result of the task, A t Represents the accuracy results of the task, α, β, θ and γ are preset hyperparameters that control the dimension and range; (15) According to the quadruple of Markov decision process MDP t ,ξ t ,s t+1 ,r t >, update the deep reinforcement learning network parameters W θ ;​ (16) Repeat steps (11) to (15) until the deep reinforcement learning network parameter W θ convergence.

9. The multi-type task scheduling system based on hierarchical embodied intelligence according to claim 1 is characterized in that: The stream data processing control module includes: a task generator, a control forwarder, an application processor and a storage distributor; The task generator is deployed on the edge node and is used to obtain real-time streaming data from the streaming data source address. It preprocesses and packages the streaming data according to the scheduling decision output by the hierarchical embodied intelligent scheduling decision module. Specifically, it adjusts the quality, frequency, and encoding configuration of the streaming data, packages it into a streaming data file, uses the data configuration decision and task offloading decision information in the scheduling decision as metadata information, and packages the streaming data file and metadata information into a streaming data task package. The control forwarder is deployed on all edge nodes and cloud servers and is used to forward the stream data task package according to the metadata information in the stream data task package. The forwarding logic is executed based on the current stage of the stream data task package as follows: if the unloading node of the current stage is the same as the node where the control forwarder is located, the stream data task package is forwarded to the corresponding stage application processor on the current node; if the unloading node of the current stage of the stream data task package is different from the node where the control forwarder is located, the stream data task package is forwarded to the control forwarder of the unloading node; if all current stages of the stream data task package have been processed, the stream data task package is forwarded to the storage distributor on the cloud server; Application processors, deployed on all edge nodes and cloud servers, are used to analyze streaming data task packets and store the task processing results in the streaming data task packets. Each application processor handles a single stage of streaming data analysis and performs analysis on the streaming data task packets based on an encapsulated neural network algorithm. The storage distributor is deployed on the cloud server and is used to collect streaming data tasks that have been processed in all stages, store the task processing results in the streaming data task package into the database, and distribute them to the runtime context-aware analysis module and the hierarchical embodied intelligent scheduling decision module.

10. A multi-type task scheduling method based on hierarchical embodied intelligence, based on the system according to any one of claims 1 to 9, characterized in that: The method comprises the following steps: 1) Collect resource context, task context, and decision context information, and analyze them to obtain runtime context characteristics s t ; 2) According to the runtime context characteristics s t , using macro-scheduling and micro-scheduling, outputting the coarse-grained decision of the system's multi-dimensional configuration knob ξ t and fine-grained decision τ t ; 3) According to the current fine-grained decision τ t , execute the current stream data task T in the cloud-edge collaborative stream data processing system t and distribute the results to the runtime context-aware analysis module and the hierarchical embodied intelligent scheduling decision module; 4) Repeat steps 1) to 3) until the stream data analysis is completed.

Images