Heterogeneous node digital orthogonal scheduling method and system adaptive to mixed data stream features

Abstract

The application discloses a heterogeneous node digital orthogonal scheduling method and system adapting to mixed data stream characteristics, relates to the field of data stream scheduling, and comprises the following steps: forming initial task units by classifying service attributes of mixed data streams, extracting multi-dimensional original features to construct orthogonal bases, and mapping different task units into orthogonal feature vectors; combining service attributes, original features and network states to calculate dynamic scheduling priority weights, generating node capability vectors according to attributes and running states of processing nodes, and matching and evaluating the node capability vectors with the orthogonal feature vectors to form a compatibility matrix; generating a candidate scheduling scheme by comprehensively considering the priority and the compatibility, calculating orthogonal conflict degrees of task units to be allocated to the same processing node, and applying scheduling penalties to screen or adjust the candidate allocation relationship to generate a scheduling matrix. Therefore, the resource adaptability and scheduling stability of the mixed data stream in a heterogeneous environment can be improved.

Description

Technical Field

[0001] This application relates to the field of data flow scheduling technology, and in particular to a heterogeneous node digital orthogonal scheduling method and system adapted to the characteristics of mixed data flows. Background Technology

[0002] With the continuous development of mobile edge computing, IoT sensing, industrial internet, intelligent video analytics, and cloud-edge-device collaborative systems, the types of data streams carried in distributed computing environments are becoming increasingly diverse. Real-world systems often need to simultaneously process multiple data streams with different business attributes and spatiotemporal characteristics, such as high-frequency, small-packet sensor sampling streams, high-bandwidth, continuously transmitted video streams, control command streams sensitive to response latency, and log or alarm streams with event-triggered characteristics. These data streams differ significantly in arrival patterns, data size, bandwidth usage, computational requirements, real-time requirements, and business priorities, leading to more complex resource coordination needs for distributed systems when managing data streams and scheduling tasks.

[0003] In such application scenarios, data streams typically need to be distributed to heterogeneous computing environments consisting of various processing nodes, such as general-purpose processing nodes, parallel computing nodes, hardware-accelerated nodes, and edge processing nodes located close to the data source. Different processing nodes vary in computing power, parallelism, cache capacity, hardware acceleration features, energy consumption, and network link status, and different data streams also have varying degrees of dependence on these resources. Therefore, in complex distributed environments, data stream scheduling involves not only arranging the order of task execution but also the allocation, coordination, and operational stability of multiple types of data streams across different processing resources.

[0004] Currently, data stream scheduling technologies typically make scheduling decisions based on task priority, node load, task dependencies, or historical operational data, enabling basic task allocation and resource utilization in general scenarios. However, in operating environments with continuously arriving mixed data streams, fluctuating network conditions, and complex processing node types, traditional scheduling methods often have limited ability to express the differences between different data streams, making it difficult to fully adapt to resource competition and latency fluctuations under conditions of multi-service concurrency. Summary of the Invention

[0005] This application provides a heterogeneous node digital orthogonal scheduling method, system, storage medium, computer program product, and electronic device adapted to the characteristics of mixed data streams, in order to at least solve the problem in the prior art that it is difficult to achieve stable, efficient, and fine-grained collaborative scheduling of mixed data streams in a heterogeneous distributed computing environment.

[0006] In a first aspect, embodiments of this application provide a heterogeneous node digital orthogonal scheduling method adapted to the characteristics of mixed data streams. The method includes: acquiring a mixed data stream to be processed, network status data of the current system, and node data of each processing node in a heterogeneous computing environment; classifying the mixed data stream according to business attributes to generate multiple initial task units; wherein the node data includes node attribute data and node running status data; extracting multidimensional original features for each initial task unit; generating an orthogonal basis for constructing an orthogonal feature space based on the multidimensional original features; and performing orthogonal mapping on the multidimensional original features based on the orthogonal basis to generate orthogonal feature vectors corresponding to each initial task unit; calculating the dynamic scheduling priority weight of each initial task unit based on the business attributes, the multidimensional original features, and the network status data; and generating node capability vectors corresponding to each processing node based on the node data of each processing node. The matching degree of the orthogonal feature vectors of each initial task unit and the node capability vector of each processing node is evaluated to generate a compatibility matrix for quantifying the degree of fit between the initial task units and the processing nodes. The dynamic scheduling priority weights and the compatibility matrix are combined to generate a candidate scheduling scheme containing candidate allocation relationships between the initial task units and candidate processing nodes. For at least two initial task units in the candidate scheduling scheme that are intended to be allocated to the same processing node, an orthogonal conflict degree is calculated based on their corresponding orthogonal feature vectors. If the orthogonal conflict degree meets a preset conflict condition, a scheduling penalty is applied to the candidate allocation relationships in the candidate scheduling scheme that are intended to allocate at least two initial task units to the same processing node. The candidate allocation relationships are then filtered or adjusted based on the scheduling penalty to obtain a target scheduling scheme. The target processing node for each initial task unit is determined according to the target scheduling scheme to generate and output a scheduling matrix.

[0007] Secondly, embodiments of this application provide a heterogeneous node digital orthogonal scheduling system adapted to the characteristics of mixed data streams. The system includes: a data acquisition and classification module, used to acquire the mixed data stream to be processed, the network status data of the current system, and the node data of each processing node in the heterogeneous computing environment, and classify the mixed data stream according to business attributes to generate multiple initial task units; wherein, the node data includes node attribute data and node running status data; an orthogonal feature mapping module, used to extract multidimensional original features for each initial task unit, generate an orthogonal basis for constructing an orthogonal feature space based on the multidimensional original features, and perform orthogonal mapping on the multidimensional original features based on the orthogonal basis to generate orthogonal feature vectors corresponding to each initial task unit; a priority weight calculation module, used to calculate the dynamic scheduling priority weight of each initial task unit based on the business attributes, the multidimensional original features, and the network status data of each initial task unit; and a compatibility matrix generation module, used to generate each processing node based on the node data of each processing node. The system generates a compatibility matrix to quantify the fit between the initial task units and the processing nodes by comparing the corresponding node capability vectors with the orthogonal feature vectors of each initial task unit with the node capability vectors of each processing node. A candidate scheduling generation module combines the dynamic scheduling priority weights with the compatibility matrix to generate a candidate scheduling scheme containing candidate allocation relationships between the initial task units and candidate processing nodes. For at least two initial task units in the candidate scheduling scheme that are intended to be allocated to the same processing node, an orthogonal conflict degree is calculated based on their corresponding orthogonal feature vectors. A conflict penalty adjustment module applies a scheduling penalty to the candidate allocation relationships in the candidate scheduling scheme that are intended to allocate at least two initial task units to the same processing node when the orthogonal conflict degree meets a preset conflict condition. Based on the scheduling penalty, the candidate allocation relationships are filtered or adjusted to obtain a target scheduling scheme. The target processing node for each initial task unit is determined according to the target scheduling scheme to generate and output a scheduling matrix.

[0008] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the heterogeneous node digital orthogonal scheduling method adapted to the hybrid data stream characteristics of any embodiment of the present application.

[0009] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the heterogeneous node digital orthogonal scheduling method adapted to the characteristics of mixed data streams according to any embodiment of this application.

[0010] Fifthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the heterogeneous node digital orthogonal scheduling method adapted to the hybrid data stream characteristics of any embodiment of this application.

[0011] The heterogeneous node digital orthogonal scheduling method and system adapted to the characteristics of mixed data streams provided in this application can achieve at least the following technical effects:

[0012] By jointly representing the business attributes, multi-dimensional original features, and network state of mixed data streams, and further orthogonally mapping the multi-dimensional original features, the differences in bandwidth usage, computational demands, and real-time requirements among different data streams can participate in scheduling decisions with higher distinguishability. Since the orthogonalized feature representation objectively weakens the coupling effect between different feature dimensions, the scheduling process can more accurately identify the main resource requirements and latency sensitivity of each initial task unit. Simultaneously, dynamic scheduling priority weights are generated based on network state, allowing the importance evaluation of tasks to be dynamically adjusted according to changes in operational states such as link congestion and latency fluctuations. Therefore, the system can form a task priority ranking and resource demand expression that fits the current operating environment when multiple types of data streams arrive concurrently, improving the granularity and real-time adaptability of mixed data stream scheduling decisions.

[0013] The orthogonal feature vectors of each initial task unit are matched with the node capability vectors of heterogeneous processing nodes to form a compatibility matrix for quantifying the suitability between tasks and nodes. In the scheduling decision, an orthogonal conflict degree judgment based on orthogonal feature vectors is introduced for multiple initial task units to be assigned to the same processing node. This allows the scheduling decision to not only consider the capability matching relationship between a single task and a single processing node, but also effectively assess the degree of competition for multidimensional computing resources when multiple tasks are assigned to the same processing node. When the orthogonal conflict degree meets the preset conflict conditions, scheduling penalties are applied to the corresponding candidate allocation relationships, and screening or adjustment is performed. This reduces the probability of tasks with highly overlapping resource requirements being concentrated on the same processing node, thereby reducing node congestion, queuing, and execution jitter, and improving the balance of resource utilization across heterogeneous nodes and the stability of task execution.

[0014] This technical solution utilizes orthogonal feature representation to transform the differentiated requirements of hybrid data streams into objective, quantifiable, and loosely coupled scheduling criteria. Based on this, it links dynamic scheduling priority weights, a compatibility matrix, and an orthogonal conflict penalty mechanism to form a collaborative scheduling architecture that encompasses task feature identification, precise node adaptation, and resource conflict suppression. This improves the matching accuracy between hybrid data streams and processing resources in heterogeneous distributed computing environments, enhances the overall scheduling scheme's adaptability to business differences and fluctuations in underlying operational states, and effectively ensures the stability of digital orthogonal scheduling in high-concurrency multi-service scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating an example of a heterogeneous node digital orthogonal scheduling method adapted to hybrid data stream characteristics according to an embodiment of this application is shown.

[0017] Figure 2 A flowchart illustrating an example of generating a compatibility matrix according to an embodiment of this application is shown.

[0018] Figure 3 A flowchart illustrating an example of determining a target scheduling scheme and generating a scheduling matrix according to an embodiment of this application is shown.

[0019] Figure 4 A schematic diagram illustrating the system operation mechanism of an example heterogeneous node digital orthogonal scheduling method adapted to hybrid data stream characteristics according to an embodiment of this application is shown.

[0020] Figure 5 A comparative simulation diagram showing the end-to-end delay accumulation distribution and long-tail characteristics of different scheduling methods is presented as an example.

[0021] Figure 6 A schematic diagram of a comparative experiment simulation showing the orthogonal isolation effect and system overhead trade-off performance of different scheduling methods in a scenario of sudden changes in background traffic is presented.

[0022] Figure 7 A simulation diagram illustrating an example of the trade-off between the parameter sensitivity of the orthogonal penalty coefficient and resource utilization is shown.

[0023] Figure 8A structural block diagram of an example of a heterogeneous node digital orthogonal scheduling system adapted to hybrid data stream characteristics, according to an embodiment of this application, is shown. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] It should be noted that, at the level of specific scheduling mechanisms, current technologies largely employ heuristic scheduling methods based on task graph representations, such as DAG (Directed Acyclic Graph) task modeling, HEFT (Heterogeneous Earliest Finish Time) list scheduling, DLS (Dynamic Level Scheduling), and clustering scheduling. These methods typically determine task priorities based on topological dependencies between tasks, estimated execution time, and communication costs, and then map tasks to corresponding processing nodes according to priority. In scenarios with relatively stable task relationships, small task sizes, or simple business types, these methods are characterized by simple implementation and low operational overhead.

[0026] However, mixed data streams often exhibit characteristics such as continuous arrival, sudden changes, periodic fluctuations, and dynamic changes in service priorities during operation, which differ from discrete tasks in a static task graph. Scheduling methods based on DAG, HEFT, or DLS typically focus more on task topology and processor execution costs, and have relatively limited capacity to express differences in data stream content type, timing patterns, bandwidth usage, real-time requirements, and service urgency. While clustering scheduling can reduce some communication overhead by grouping tasks, the grouping granularity usually relies on preset rules or human experience. When data stream types are complex, task scale increases, or network conditions fluctuate significantly, it can easily lead to local resource concentration or resource competition between different service streams.

[0027] To improve the adaptability of scheduling strategies, some studies have further introduced methods such as AI scheduling, DRL (Deep Reinforcement Learning) scheduling, and GA / QGA (Genetic Algorithm / Quantum Genetic Algorithm) to attempt to learn task allocation strategies through historical running data or search for better resource mapping results under multi-objective constraints. These methods have certain advantages in handling complex constraints and global optimization, but they typically rely on a large number of training samples, iterative searches, or parameter tuning processes, resulting in limited adaptability when faced with sudden surges in business or new data flows. Furthermore, their scheduling processes are often not intuitive enough, making it difficult to clearly explain the correspondence between scheduling results and data flow characteristics and node capabilities.

[0028] For edge computing and multi-source data processing scenarios, some research has proposed low-latency scheduling frameworks such as MDS (Multi-source Data Scheduling) for 5G environments. These frameworks typically combine dynamic prioritization, node load assessment, network state awareness, and cache management to adjust task execution order, which can improve average processing latency and resource utilization in edge-cloud collaborative environments to some extent. However, their focus is usually concentrated on task order adjustment, cache coordination, and load correction, with limited characterization of the timing characteristics, burst intensity, bandwidth sensitivity, and mutual influence between different types of service flows.

[0029] In heterogeneous processing platforms, current technologies typically integrate different processing resources such as CPUs, GPUs, FPGAs, and DSPs into a unified scheduling system through a unified software framework, driver abstraction layer, or resource management layer, and map tasks based on computation time, communication latency, or node load. While this approach achieves basic unified management of heterogeneous resources, scheduling decisions often focus more on general performance metrics at the processor level, neglecting the compatibility between data flow characteristics and specific hardware capabilities. For example, different data flows have varying dependencies on parallel computing, pipelined processing, cache space, link stability, and low-latency response. Estimating only coarse-grained execution time or communication costs is insufficient to accurately reflect the actual processing adaptability of different data flows on heterogeneous nodes.

[0030] Furthermore, in multi-service concurrent environments, the impact between different data streams is not limited to priority competition; it can also manifest as overlapping bandwidth usage, cache congestion, computing channel conflicts, and interference with critical services from sudden traffic surges. Currently, these impacts are typically mitigated through queuing, load balancing, resource reservation, or cache replacement. However, these methods provide a rather coarse representation of the differences and interactions between service streams. If scheduling strategies do not adequately consider service differences, critical data streams may be affected by high-throughput or sudden traffic surges; conversely, overly stringent resource constraints may lead to underutilization of some heterogeneous resources. Therefore, in complex heterogeneous environments, data stream scheduling still needs to develop more refined scheduling support that balances the representation of service differences, resource adaptation, and multi-stream collaborative operation.

[0031] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application. Furthermore, the technical solutions described in the above-mentioned relevant technologies are not prior art, and may also be undisclosed technical solutions, such as those under research or in the laboratory stage.

[0032] Figure 1 A flowchart illustrating an example of a heterogeneous node digital orthogonal scheduling method adapted to hybrid data stream characteristics according to an embodiment of this application is shown.

[0033] Regarding the execution entity of the method in the embodiments of this application, it can be any controller or processor with computing or processing capabilities, such as a cloud-edge-device collaborative scheduling platform controller, a distributed computing resource management platform controller, an edge computing gateway controller, an industrial internet task scheduling controller, or a heterogeneous computing cluster scheduling controller. It implements the method of the embodiments of this application by running programs or instructions stored in a storage medium. In some examples, it can be integrated and configured in an electronic device or terminal through software, hardware, or a combination of both, and the type of terminal or electronic device can be diverse.

[0034] like Figure 1 As shown, in step S110, the mixed data stream to be processed, the network status data of the current system, and the node data of each processing node in the heterogeneous computing environment are acquired, and the mixed data stream is classified according to business attributes to generate multiple initial task units. The node data includes node attribute data and node running status data.

[0035] Here, the mixed data stream can include multiple types of data with different sources, uses, and processing requirements. For example, in a smart manufacturing production line, the system may simultaneously receive periodic sensor data uploaded by robotic arms or tooling equipment, image or video detection data uploaded by machine vision equipment, control command data generated by PLCs or motion controllers, alarm event data generated by equipment status monitoring modules, and production process record data. These data streams differ in terms of data scale, arrival frequency, real-time requirements, reliability requirements, and business importance. Therefore, this embodiment does not directly schedule them as completely homogeneous data objects, but first identifies and classifies them according to their business attributes.

[0036] In some implementations, the system can receive and parse the mixed data stream at the data access side, and determine the service attributes of the data stream based on information such as data source identifier, service tag, communication protocol type, port information, packet header fields, timestamp information, quality of service identifier, or preset service configuration table. Based on the service attributes, the system divides the continuously or interleaved mixed data stream into multiple logically independent initial task units. An initial task unit can be a set of data packets within a time window, an image detection task to be executed, a set of control instructions, an alarm processing object, or a processing task formed by aggregating similar data streams. Thus, data streams of different types and granularities are converted into task units that the scheduling system can uniformly identify and operate.

[0037] Simultaneously, the system acquires current network status data and node data for each processing node in the heterogeneous computing environment. Network status data may include link bandwidth utilization, end-to-end transmission latency, queue backlog, packet loss, retransmission status, or network congestion level. Node attribute data may include the processor type, computing power, concurrent processing capacity, supported data stream types, hardware acceleration capabilities, storage capacity, and communication interface capabilities of the processing node. Node runtime status data may include real-time load rate, available memory, remaining computing resources, task queue length, online status, and current service status. In this way, the system obtains the basic status of the task side, network side, and node side, integrating the scheduling object and runtime environment into a unified scheduling view.

[0038] Therefore, the mixed data streams with complex sources and different business attributes are organized into multiple schedulable initial task units, and the network status and node status are acquired simultaneously. This enables the scheduling system to have basic perception capabilities of task type, network environment and node carrying capacity before executing task allocation, thereby reducing scheduling deviations caused by chaotic data stream granularity or unknown node status.

[0039] In step S120, multidimensional original features are extracted for each initial task unit, an orthogonal basis for constructing an orthogonal feature space is generated based on the multidimensional original features, and the multidimensional original features are orthogonally mapped based on the orthogonal basis to generate the orthogonal feature vector corresponding to each initial task unit.

[0040] In some examples, the system can extract multi-dimensional raw features based on the business type and data structure of the initial task unit. For sensor acquisition tasks, features such as sampling frequency, data packet interval, data duration, and timeliness requirements can be extracted; for visual inspection tasks, features such as data scale, frame rate, image resolution, and processing complexity can be extracted; for control command tasks, features such as controlled object type, response time requirements, and business priority can be extracted; and for alarm event tasks, features such as alarm level, trigger time, and associated device range can be extracted. In this way, the above features can describe the processing requirements and scheduling characteristics of the initial task unit from different perspectives.

[0041] It should be noted that, due to potential differences in scale and correlations between different feature dimensions—for example, tasks with large data volumes often have high bandwidth or caching requirements—directly using multidimensional raw features for scheduling and matching can easily lead to redundant emphasis on related features. Therefore, this embodiment generates an orthogonal basis for constructing an orthogonal feature space based on the multidimensional raw features, and then uses this orthogonal basis to orthogonalize the multidimensional raw features. This transforms the raw features of the initial task units into relatively independent and easily comparable feature representations, allowing feature differences between different tasks to be represented in a unified vector form.

[0042] After orthogonalization mapping, each initial task unit generates an orthogonal feature vector. This orthogonal feature vector represents the feature state of the corresponding initial task unit in the orthogonal feature space. It's important to note that orthogonal feature vectors do not require all tasks to be completely orthogonal to each other. Rather, they mean that the features of each task are mapped to a space represented by an orthogonal basis, allowing for a more stable measurement of the similarity, differences, and potential feature coupling relationships between tasks. For example, for multiple visual detection tasks, orthogonal feature representations can help distinguish differences in data scale, timeliness requirements, and processing complexity; for control command and sensor acquisition tasks, orthogonal feature representations can preserve scheduling-related features such as high frequency and low latency. This weakens the interference of correlations between different dimensions in the multidimensional original features on scheduling decisions, allowing the scheduling characteristics of the initial task units to be presented in a unified and comparable orthogonal feature vector form, thereby improving the stability and computability of task feature representation.

[0043] In step S130, the dynamic scheduling priority weight of each initial task unit is calculated based on the service attributes, multidimensional original features, and network status data of each initial task unit.

[0044] In some examples, dynamic scheduling priority weights are used to indicate the degree to which an initial task unit should be prioritized within the current scheduling cycle. These weights are not solely determined by a preset service level, but rather by a comprehensive consideration of the initial task unit's service attributes, multidimensional original characteristics, and current network status. For instance, in industrial control scenarios, closed-loop control commands typically have high real-time requirements; visual inspection data is usually large in volume and has high processing demands; and background logs or statistical data typically have less reliance on immediate processing. By dynamically calculating priority weights, the system can determine task scheduling preferences based on the current operating state rather than fixed rules.

[0045] In practical implementation, the system can determine the business category, business level, real-time requirements, or service quality requirements of the initial task unit based on business attributes; determine its data scale, processing complexity, resource requirements, time constraints, or arrival frequency based on multi-dimensional original characteristics; and determine whether there are issues such as increased latency, intensified congestion, or unstable transmission on the current link based on network status data. Based on the above information, the system generates corresponding dynamic scheduling priority weights, giving task units with high business urgency, strong time constraints, or significant network status influence a higher scheduling tendency within the current scheduling cycle.

[0046] For example, when the network is relatively stable, the system can establish a relatively balanced scheduling order based on the data size, processing requirements, and business level of each task. When the network deteriorates or the latency of critical links increases, the system can increase the scheduling priority of low-latency control tasks, alarm event tasks, or security interlock tasks, while appropriately reducing the immediate allocation preference for background tasks with lower timeliness requirements. This allows the scheduling weights to be updated according to changes in business needs and the operating environment, avoiding the distortion of fixed priorities in dynamic scenarios.

[0047] In this embodiment, the importance of the task itself, resource requirements, and changes in the network environment are all incorporated into the priority evaluation, so that the scheduling system can still form a task processing order that conforms to the current operating state when there are network fluctuations, sudden task outbreaks, or changes in business load, thereby improving the stability of critical task response.

[0048] In step S140, a node capability vector corresponding to each processing node is generated based on the node data of each processing node, and the matching degree of the orthogonal feature vector of each initial task unit and the node capability vector of each processing node is evaluated to generate a compatibility matrix for quantifying the degree of adaptation between the initial task unit and the processing node.

[0049] In some examples, processing nodes in a heterogeneous computing environment can include general-purpose CPU nodes, GPU-accelerated nodes, FPGA-accelerated nodes, edge control nodes, cloud-edge collaborative nodes, or embedded computing nodes, with different processing nodes suited to different types of tasks. For example, GPU-accelerated nodes are typically suitable for image detection, video analysis, or deep learning inference tasks; FPGA or dedicated acceleration nodes can be suitable for low-latency data preprocessing, protocol parsing, or specific signal processing tasks; edge control nodes can be suitable for near-end control tasks; and general-purpose CPU nodes can be suitable for logic orchestration, medium-scale data processing, or task coordination.

[0050] The system generates node capability vectors based on node attribute data and node operational status data. Node attribute data reflects the inherent capabilities of a node, such as its computing architecture, parallel processing capabilities, hardware acceleration capabilities, supported data types, and communication interface capabilities. Node operational status data reflects the node's current actual available capabilities, such as load level, available memory, remaining computing resources, queue backlog, and online status. Through node capability vectors, the system converts the hardware capabilities and real-time operational status of processing nodes into an expression that can be used for matching degree evaluation.

[0051] Subsequently, the system evaluates the matching degree between the orthogonal feature vectors of each initial task unit and the node capability vectors of each processing node, generating a compatibility matrix. The elements in the compatibility matrix represent the degree of fit between a given initial task unit and a given processing node, which can be influenced by factors such as the processing node's hardware capabilities, current load, task resource requirements, communication overhead, and data stream type support. For example, high-resolution visual inspection tasks can have a high degree of fit with nodes that have image processing acceleration capabilities and low load; low-latency control tasks can have a high degree of fit with edge control nodes deployed near the edge and with minimal queue backlog; if a processing node does not support a certain type of data stream processing, the compatibility between that task and that node can be reduced or eliminated.

[0052] In step S150, by combining the dynamic scheduling priority weight and the compatibility matrix, a candidate scheduling scheme containing the candidate allocation relationship between the initial task units and the candidate processing nodes is generated. For at least two initial task units in the candidate scheduling scheme that are to be allocated to the same processing node, the orthogonal conflict degree is calculated based on the corresponding orthogonal feature vectors.

[0053] In some examples, the system can determine the order or priority of initial task units participating in scheduling based on dynamic scheduling priority weights, and determine the candidate processing nodes available for each initial task unit based on the compatibility matrix. For task-node combinations with high compatibility, the system can include them in the candidate allocation relationship; for combinations with low compatibility or that do not meet the processing conditions, the system can reduce the probability of their selection or exclude them. The resulting candidate scheduling scheme can represent a set of feasible task allocation schemes under the current system state.

[0054] However, a high degree of compatibility between tasks and nodes does not necessarily mean that multiple tasks are suitable for being shared with the same node. For example, multiple high-resolution visual inspection tasks may all be highly compatible with the same GPU node, but if they run simultaneously, they may compete for GPU memory, inference queues, or memory bandwidth; multiple high-frequency sensor data tasks may all be suitable for processing by the same edge node, but concentrated execution may cause increased latency in short-cycle data processing. Therefore, the system further calculates the orthogonal conflict degree for multiple initial task units in the candidate scheduling scheme that are intended to be assigned to the same processing node.

[0055] Orthogonal conflict degree is determined based on the orthogonal feature vectors of the corresponding initial task units. It is used to characterize the degree of feature overlap or resource competition risk that may occur when multiple tasks are executed on the same processing node. If multiple initial task units show high similarity in the orthogonal feature space, it can be considered that their data scale, computing requirements, bandwidth usage, latency constraints, or other resource request patterns have strong overlap, and execution on the same node may produce high conflict. If their orthogonal features differ greatly, it indicates that the resource requirements between the corresponding tasks may be more complementary, and the risk of conflict when executed on the same node is relatively low.

[0056] In this embodiment, potential conflicts between tasks on the same node are further identified in the candidate scheduling schemes, so that the scheduling system can not only consider the compatibility between a single task and a node, but also evaluate the mutual influence when multiple tasks share the same processing node, thereby improving the reliability of the candidate scheduling schemes.

[0057] In step S160, if the orthogonal conflict degree meets the preset conflict condition, a scheduling penalty is applied to the candidate allocation relationship in the candidate scheduling scheme used to allocate at least two initial task units to the same processing node, and the candidate allocation relationship is screened or adjusted based on the scheduling penalty to obtain the target scheduling scheme. The target processing node of each initial task unit is determined according to the target scheduling scheme to generate and output the scheduling matrix.

[0058] In some examples, preset conflict conditions are used to determine whether there is a risk of conflict that needs to be suppressed among multiple initial task units on the same processing node. When the orthogonal conflict degree does not meet the preset conflict condition, the system can treat the corresponding candidate allocation relationship as a reservable scheduling relationship; when the orthogonal conflict degree meets the preset conflict condition, the system can determine that the candidate allocation relationship has a high risk of resource contention. For example, when multiple visual detection tasks tend to be allocated to the same image processing node and have a high orthogonal conflict degree, it may cause memory or inference queue contention; when multiple short-cycle sampling tasks tend to be allocated to the same edge node and have a high orthogonal conflict degree, it may cause task queuing or increased response latency.

[0059] When preset conflict conditions are met, the system applies a scheduling penalty to the corresponding candidate allocation relationship. This penalty can manifest as increasing the cost of the candidate allocation relationship in the scheduling evaluation, reducing its probability of being retained in the target scheduling scheme, or prompting the system to adjust one or more initial task units to other candidate processing nodes with sufficient fit and low conflict levels. This scheduling penalty targets candidate allocation relationships whose orthogonal conflict degree meets the conflict determination condition, rather than excluding all concurrent execution relationships on the same node. Therefore, the system can still retain allocation relationships on the same node with complementary features or low conflict levels to maintain the resource utilization of processing nodes.

[0060] The system filters or adjusts candidate allocation relationships based on scheduling penalties to obtain a target scheduling scheme, which represents the allocation result after comprehensively weighing task priority, task-node suitability, and the degree of conflict between tasks. For example, in a candidate scheduling scheme, if two vision detection tasks both tend to be on the same GPU node but have a high degree of conflict, the system can retain the task that is more suitable for that node or has a higher priority, and adjust the other task to another node that meets the processing conditions; if a low-latency control task and a low-priority background task compete for the same near-edge node, the system can prioritize the allocation of the target node to the control task.

[0061] After obtaining the target scheduling scheme, the system determines the target processing node corresponding to each initial task unit based on the scheme and generates a scheduling matrix. In some examples, the scheduling matrix can be represented by a matrix, table, mapping table, array, or other machine-readable data structure to indicate the final correspondence between the initial task units and the target processing nodes. Furthermore, after the scheduling matrix is ​​output, it can be read by the task assignment module, resource control module, or execution scheduling module, causing the corresponding processing node to execute the initial task unit assigned to it. This allows for targeted suppression of highly conflicting candidate assignment relationships, achieving a balance between task priority, node suitability, and the risk of conflict between tasks on the same node in the final scheduling result. This reduces resource contention and queuing risks caused by the concentrated execution of homogeneous tasks and improves the executability of the scheduling matrix in heterogeneous computing environments.

[0062] Regarding the implementation details of generating orthogonal feature vectors corresponding to each initial task unit in the method of this application's embodiments, in some examples of this application's embodiments, multidimensional original features of each initial task unit are extracted. These multidimensional original features include at least time-domain features, bandwidth features, and service priority features. The multidimensional original features are then normalized to obtain the normalized feature vectors corresponding to each initial task unit. .

[0063] In heterogeneous distributed computing scenarios, the physical characteristics of different data streams may differ significantly. For example, time-domain features can be used to characterize the arrival time interval, duration, or lifecycle of a data stream; bandwidth features can be used to characterize the instantaneous packet size, throughput requirements, or transmission occupancy; and service priority features can be used to characterize the relative importance of different service tasks during scheduling. Because these features are inconsistent in terms of dimensions, numerical range, and physical meaning, directly involving them in the construction of orthogonal spaces may lead to a situation where feature dimensions with larger numerical ranges have an excessively strong influence on the mapping results.

[0064] In a specific example, extremum normalization can be used to make the original multidimensional features dimensionless. Let the first... The initial task unit in the first The original feature values ​​of each dimension are The corresponding normalized feature elements are If there exists a range of non-zero values ​​in this dimension, then It can be calculated using the following formula:

[0065] Equation (1)

[0066] In the formula, and These represent the initial task units participating in the current scheduling calculation at the [number]th [time]. Minimum and maximum values ​​of features in each dimension. If a feature appears in a certain feature dimension... In such cases, the normalized feature elements of that dimension can be uniformly set to a preset constant, or a preset small positive number can be introduced into the normalization denominator to avoid the denominator being zero. In this way, the original features of different dimensions are transformed to a uniform numerical scale, thereby obtaining the normalized feature vectors corresponding to each initial task unit. .

[0067] Then, a feature weight matrix is ​​constructed to represent the weight of different feature dimensions in the scheduling decision. Among them, the feature weight matrix It is a positive definite diagonal matrix, and the eigenweight matrix is... The diagonal elements are positively correlated with the latency sensitivity and / or business priority of the corresponding feature dimensions.

[0068] In some examples, it is assumed that the normalized feature vector has If there are multiple dimensions, then the feature weight matrix It can be represented as:

[0069] Equation (2)

[0070] in, Indicates the first The weight coefficients corresponding to each feature dimension, and By Setting it to a positive definite diagonal matrix ensures that the weighted inner product based on this matrix maintains a non-negative metric basis, and allows adjustment of the influence of different feature dimensions on orthogonalization processing based on their importance in the current business scenario. For example, feature dimensions with high latency sensitivity, or those that significantly differentiate high-priority alarm services, can be configured with relatively large weight coefficients; while feature dimensions with weaker impact on scheduling differences can be configured with relatively small weight coefficients. Thus, orthogonalization processing can reflect the scheduling sensitivity of different business feature dimensions within a unified mathematical expression.

[0071] Then, from each normalized feature vector Multiple linearly independent baseline eigenvectors are selected or constructed as support sets, and based on the feature weight matrix... The weighted Gram-Schmidt algorithm is used to generate orthogonal bases. Specifically, during the generation of orthogonal bases, for the current baseline eigenvectors, the feature weight matrix is ​​used to generate the orthogonal base. The weighted correlation between the current baseline eigenvector and the generated orthogonal basis vectors is measured, and the weighted projection components of the current baseline eigenvector along the directions of the generated orthogonal basis vectors are removed to obtain the orthogonal intermediate vector; subsequently, based on the eigenweight matrix... Normalize the orthogonal intermediate vectors to obtain the corresponding orthogonal basis vectors.

[0072] For example, orthogonal bases include The nth orthogonal basis vector in this orthogonal basis orthogonal basis vectors Generate according to the following formula:

[0073] Equation (3)

[0074] Equation (4)

[0075] In the formula, Indicates the first in the support set One baseline feature vector, This represents the orthogonal intermediate vector during the orthogonalization iteration process. Indicates the number of generated [items]. orthogonal basis vectors , For the summation index, This represents the transpose of a vector, and It is a positive integer not greater than the dimension of the normalized feature vector.

[0076] In the above process, the system targets the current benchmark feature vector. Calculate its relative to the generated orthogonal basis vectors in sequence. The weighted projection components are then extracted from the current reference eigenvectors, and these projection components are removed to obtain orthogonal intermediate vectors that are independent of the generated orthogonal basis vectors in the sense of weighted inner product. Subsequently, the system is based on right Weighted normalization is performed to obtain orthogonal basis vectors. The resulting orthogonal basis vectors satisfy a weighted orthogonality relation, ensuring that different basis vectors are aligned by the characteristic weight matrix. The defined metric space exhibits low correlation. If, during the support set construction process, a candidate baseline eigenvector is found to be linearly independent from the selected baseline eigenvector, the candidate vector can be skipped, or an alternative baseline eigenvector can be constructed based on the existing normalized eigenvector to ensure the stable execution of the orthogonal basis construction process.

[0077] Then, the normalized feature vectors are... Project them respectively onto the orthogonal eigenspace composed of orthogonal basis vectors, and use each normalized eigenvector as a reference. The weighted projection result on the orthogonal basis vectors is used as the corresponding orthogonal feature vector; where the orthogonal feature vector is used to calculate the orthogonal conflict degree between different initial task units.

[0078] For example, each normalized feature vector Projected onto the orthogonal basis vectors respectively In the constructed orthogonal feature space, the orthogonal feature vectors corresponding to each initial task unit are obtained. :

[0079] Equation (5)

[0080] In the formula, This represents the feature representation of the initial task unit in the orthogonal feature space. Through the projection process described above, the original multidimensional features of the initial task unit are transformed into coordinate representations in the orthogonal feature space, where each coordinate component reflects the weighted mapping intensity of the initial task unit along the corresponding orthogonal basis direction. This orthogonal feature vector can serve as a unified feature representation of the initial task units, used to measure the feature differences and feature overlap between different initial task units.

[0081] Through the processing described in this embodiment, original features with different dimensions, numerical scales, and business meanings are transformed into normalized feature vectors at a unified scale, and further mapped into orthogonal feature vectors represented by weighted orthogonal bases. This weakens the correlation and repetitive effects between original features, ensuring that feature dimensions with high latency sensitivity or high business priority are appropriately reflected in the orthogonalized representation. Simultaneously, it enables stable measurement of feature relationships between different initial task units through vector relationships. Therefore, the consistency, comparability, and anti-interference capabilities of mixed data stream feature representations can be improved, providing a reliable quantitative basis for identifying initial task units with similar resource request patterns.

[0082] Regarding the implementation details of calculating the dynamic scheduling priority weight of the initial task unit in the method of this application embodiment, in some examples of this application embodiment, for each initial task unit, feature data representing the corresponding data stream acquisition frequency, data resolution, and / or data packet size are parsed from multidimensional original features and normalized to construct a complexity index; resource requirement index is constructed based on feature data representing computing and storage requirements in multidimensional original features; remaining deadline constraints are determined based on time-domain expiration information in multidimensional original features; a preset emergency level score is matched based on service attributes to determine the service urgency index; and network congestion factors related to the corresponding initial task unit are extracted from network status data.

[0083] It should be noted that in heterogeneous distributed computing scenarios, the scheduling urgency of different initial task units does not solely depend on the preset business priority. For example, visual inspection tasks may have a large data scale and high computational requirements, control instruction tasks may have strong time constraints, alarm event tasks may have high business urgency, while background log tasks, although they may have a certain amount of data, usually have lower requirements for immediate processing. Therefore, scheduling based solely on static priority labels is insufficient to accurately reflect the actual processing needs of tasks under the current network conditions and resource environment.

[0084] In this embodiment, the complexity index can be used to characterize the data processing complexity of the initial task unit itself, for example, it can be obtained by normalizing features such as data stream acquisition frequency, data resolution, data packet size, and number of data frames; the resource requirement index can be used to characterize the initial task unit's demand for computing resources, storage resources, or cache resources; the remaining deadline constraint can be used to characterize the time remaining of the initial task unit before the processing deadline; the business urgency index can be obtained by matching business attributes with preset urgency level scores; the network congestion factor can be determined based on network status data such as link bandwidth utilization, transmission delay, packet loss rate, retransmission rate, or queue backlog. By unifying the above indicators into a comparable numerical scale, initial task units of different business types can participate in priority calculation under the same evaluation framework.

[0085] Then, a positive evaluation quantity for dynamic scheduling priority weights is constructed based on complexity indicators, resource requirement indicators, and business urgency indicators. A suppressive evaluation quantity for dynamic scheduling priority weights is constructed based on the remaining deadline constraint and network congestion factor. This ensures that initial task units with higher business urgency and shorter remaining deadlines have a higher scheduling priority tendency, and that network congestion status has a suppressive effect on dynamic scheduling priority weights. The complexity indicators, resource requirement indicators, and business urgency indicators are each configured with a weight coefficient that is dynamically adjusted according to the system's operating status, and each weight coefficient is subject to normalization constraints.

[0086] For example, the following weighted evaluation model is used to calculate the first... Dynamic scheduling priority weights for each initial task unit :

[0087] Equation (6)

[0088] In the formula, Indicates the first Dynamic scheduling priority weights for each initial task unit This represents a complexity metric. Indicators representing resource demand. Indicators representing the urgency of business operations This indicates the remaining deadline constraint. Represents the network congestion factor, and , ; , , This is a weighting coefficient that is dynamically adjusted according to the system's operating status, and , , , .in, , , , and It can be a normalized or standardized value to reduce the impact of different dimensions on the weight calculation.

[0089] In the weighted evaluation model described in equation (6) above, the numerator is used to characterize the positive scheduling requirements of the initial task unit. The complexity index, resource requirement index, and business urgency index participate in the priority evaluation from three aspects: data processing complexity, resource consumption requirements, and business importance, respectively. The weighting coefficient is used to adjust the relative influence of different indicators under the current system state, so that the system can adjust the priority calculation tendency according to the operating environment and business strategy.

[0090] The denominator term in equation (6) is used to constrain the dynamic scheduling priority weight. The smaller the remaining deadline constraint, the closer the task is to its processing deadline. Under the same conditions, the corresponding dynamic scheduling priority weight can be increased. The larger the network congestion factor, the more unfavorable the current communication environment. The denominator term has a suppressive effect on the weight, thereby preventing tasks with high data size or resource requirements from continuously occupying scheduling advantages in a congested environment. The square root operation is used to scale the denominator term, making the influence of the remaining deadline constraint and the network congestion factor on the weight smoother, reducing the risk of excessive fluctuations in the weight value due to changes in a single factor.

[0091] Furthermore, an adaptive adjustment mechanism for weighting parameters is constructed based on network status data. When network status data indicates that network congestion is intensifying, the weighting parameter adaptive adjustment mechanism is used to increase the value of the weighting coefficient corresponding to the business urgency index. The values ​​of the weighting coefficients corresponding to the complexity index and / or resource demand index are normalized and adjusted to improve the relative influence of the business urgency index in the dynamic scheduling priority weight calculation.

[0092] In some examples, when the system detects system-level network congestion evaluation metrics... Exceeding the preset congestion threshold When this happens, the following linear compensation ratio equation can be used to update the proportion coefficients of each item:

[0093] Equation (7)

[0094] Equation (8)

[0095] Equation (9)

[0096] In the formula, , and These represent the updated weighting coefficients for business urgency, complexity, and resource requirements, respectively. , and These represent the weighting coefficients before the update. The preset compensation sensitivity constant and , To preset the congestion threshold, A system-level congestion evaluation index used to characterize the overall network status of the current system.

[0097] In specific implementation, when At this time, the weighting coefficients can be kept constant, or minor adjustments can be made according to the system's preset strategy. When this happens, the system increases the weighting coefficient of the business urgency indicator and proportionally compresses the weighting coefficients of the complexity and resource requirement indicators, ensuring that the updated weighting coefficients still meet the normalization constraints. If this occurs... In extreme cases, it can maintain , , Alternatively, a preset lower limit protection parameter can be used for the denominator. Corrections were made to avoid division by zero during the scaling process.

[0098] By adopting the aforementioned weighting coefficient update method, the system can reduce the dominance of complexity and resource requirement indicators on scheduling priority and increase the relative influence of business urgency indicators when network congestion increases. For example, within the same scheduling cycle, if there are large-scale visual inspection tasks, background data statistics tasks, and control command tasks simultaneously, the control command task with higher business urgency can obtain a higher relative priority when network congestion increases. Tasks with large data volumes but weak timeliness requirements will not continue to occupy a priority position solely due to higher complexity or resource requirements.

[0099] Through the embodiments of this application, the dynamic scheduling priority weight can simultaneously reflect the data complexity, resource consumption requirements, business urgency, remaining deadline and network congestion status of the initial task unit. It can reduce the scheduling deviation caused by the inconsistency between static priority labels and real-time operating status, so that different types of initial task units can obtain a scheduling priority order that is more in line with the current operating conditions when the network status changes, and improve the smoothness and interpretability of priority calculation results.

[0100] Figure 2 A flowchart illustrating an example of generating a compatibility matrix in a method according to an embodiment of this application is shown.

[0101] like Figure 2 As shown, in step S210, based on the computing power, concurrent processing degree, hardware acceleration characteristics and available resource status contained in the node data of the processing node, a node capability vector corresponding to each processing node is generated, and the node capability vector is dimensionally aligned so that the node capability vector can be evaluated for matching degree with orthogonal feature vectors.

[0102] It should be noted that in heterogeneous distributed computing scenarios, different processing nodes may employ different hardware architectures, such as general-purpose CPU nodes, GPU-accelerated nodes, FPGA nodes, edge control nodes, or cloud-edge collaborative nodes. These various processing nodes differ in terms of computing frequency, concurrent processing capabilities, hardware acceleration capabilities, available memory, and real-time load status. To enable these heterogeneous hardware capabilities to participate in unified task adaptation computation, the system can first extract indicators characterizing the processing node's carrying capacity from the node data and organize them into a raw processing capacity vector.

[0103] For example, suppose the first The original processing capacity vector of each processing node is It can be represented as:

[0104] Equation (10)

[0105] In the formula, This indicates the baseline computation frequency of the processing node. This indicates the number of concurrent cores in the processing node. This represents the hardware acceleration factor of the processing node. This indicates the real-time available memory capacity of the processing node. In practical implementation, It can be a standardized, normalized, or ranked index value to reduce the impact of different physical dimensions on subsequent vector calculations.

[0106] To enable the node capability vectors to be matched with the orthogonal feature vectors of the initial task units for degree evaluation, a pre-defined dimension alignment matrix can be used. For the original processing capability vector Mapping is performed to obtain the node capability vector. :

[0107] Equation (11)

[0108] In the formula, Indicates the first The original processing capacity vector of each processing node. The preset dimension alignment matrix has dimensions of . ,in Original processing capability vector The number of feature dimensions, The number of spatial dimensions of the orthogonal feature space. This is the node capability vector after dimension alignment, with dimensions of... In the example above, if the original processing capability vector includes... Four indicators, then Through this dimensional alignment process, the hardware capabilities and operational status of the processing node can be converted into a vector representation consistent with the orthogonal feature vector dimension.

[0109] In step S220, for any initial task unit and any processing node, a non-negative matching degree is determined based on the matching relationship between the orthogonal feature vector of the initial task unit and the node capability vector of the processing node.

[0110] For example, for the first The initial task unit and the first Each processing node is based on orthogonal feature vectors. With node capability vector The inner product is used to calculate the nonnegative matching degree. :

[0111] Equation (12)

[0112] In the formula, Indicates the first Orthogonal eigenvectors of the initial task units Indicates the first The node capability vector of each processing node. This represents the transpose of a vector. This represents the non-negative matching degree between the corresponding initial task unit and the processing node.

[0113] The inner product in equation (12) above can be used to characterize the degree of directional matching between the feature requirements of the initial task unit and the capability expression of the processing node. When the inner product result is positive, it indicates that the capability expression of the processing node has a positive adaptation contribution to the initial task unit; when the inner product result is less than or equal to zero, it indicates that the matching relationship does not reflect a positive adaptation contribution. By processing the non-positive matching result to zero through the maximum value truncation function, the negative correlation matching relationship can be avoided from producing reverse cancellation in the compatibility calculation, thereby keeping the non-negative matching degree monotonic and comparable numerical meaning.

[0114] In step S230, the matrix elements in the compatibility matrix are calculated based on the non-negative matching degree and the dynamic adaptation coefficient. The dynamic adaptation coefficient characterizes the attenuation effect of the real-time load rate of the processing node and the initial task unit allocation or migration cost on the degree of adaptation, such that the higher the real-time load rate of the processing node and the greater the allocation or migration cost, the lower the corresponding degree of adaptation.

[0115] For example, the matrix elements at the corresponding positions in the compatibility matrix are calculated according to the following formula. :

[0116] Equation (13)

[0117] Equation (14)

[0118] In the formula, This indicates the compatibility matrix corresponding to the first... The initial task unit and the first The matrix elements of each processing node; Indicates non-negative matching degree; This represents the dynamic adaptation coefficient, used to characterize the attenuation effect of load rate and allocation or migration costs on the degree of adaptation. This represents the preset basic adaptation weights, and ; Represents the natural constant; The first node in the node's running status data The real-time load rate of each processing node, and ; Indicates the first The initial task unit is assigned or migrated to the first Cost parameters for each processing node, and ; This represents the preset load attenuation coefficient. This represents the preset cost sensitivity coefficient, and , .

[0119] In equation (14) above, the non-negative matching degree The dynamic adaptation coefficient reflects the degree of basic capability matching between the initial task unit and the processing node. This reflects the impact of the current operating status on the degree of adaptation. When the real-time load rate of the processing node... An increase indicates a decrease in the resources available for the processing node to accept new tasks, corresponding to a decrease in the dynamic adaptation coefficient; when the allocation or migration cost parameter... An increase in the dynamic adaptation coefficient indicates that allocating or migrating the initial task unit to this processing node may result in higher communication overhead, cache rebuilding overhead, or data migration overhead, and the corresponding dynamic adaptation coefficient also decreases. A negative exponential form allows the load rate and cost parameters to have a smooth decay effect on the degree of adaptation, avoiding abrupt changes in compatibility due to variations in a single operational state factor.

[0120] By adjusting the load attenuation coefficient and cost sensitivity coefficient The system can adjust its sensitivity to node load and migration costs based on the application scenario. For example, in edge control scenarios, sensitivity to migration costs can be increased to reduce latency caused by cross-node migration; in computationally intensive visual analytics scenarios, sensitivity to real-time load rates can be increased to avoid concentrating multiple high-load tasks on the same processing node. Therefore, the elements of the compatibility matrix can simultaneously reflect node capability matching relationships and runtime resource status.

[0121] In step S240, all initial task units and all processing nodes are traversed, the corresponding matrix elements are calculated, and a compatibility matrix is ​​constructed based on all matrix elements. Specifically, when the node attribute data contained in the node data indicates that the corresponding processing node does not support the data flow type of the corresponding initial task unit in its hardware architecture, the matrix element in the compatibility matrix corresponding to that initial task unit and that processing node is set to zero.

[0122] In practice, the system can iterate through the initial task unit set and the processing node set, calculating the non-negative matching degree, dynamic adaptation coefficient, and compatibility matrix elements for each set of initial task units and processing nodes. After calculation, all matrix elements are arranged according to task index and node index to form a compatibility matrix. In some examples, each row of this compatibility matrix can correspond to an initial task unit, each column can correspond to a processing node, and the matrix elements are used to represent the degree of adaptation between the corresponding task and the corresponding node.

[0123] For situations where the hardware architecture does not support the task, the system can directly set the corresponding matrix element to zero. For example, if an initial task unit belongs to a data stream that requires processing by a dedicated neural network acceleration unit, and a processing node does not have the corresponding acceleration capability or instruction support, then the compatibility between the task and the node can be set to zero. Similarly, if a task requires a low-latency control interface, and the corresponding node does not have the required interface or its deployment location does not meet the requirements, the corresponding matrix element can also be set to zero. Through this processing, combinations of task nodes that do not meet the actual execution conditions can be excluded from the compatibility matrix, preventing them from participating in the compatibility comparison.

[0124] In this embodiment, the static hardware capabilities, dynamic operating status, real-time load, and task allocation or migration costs of heterogeneous processing nodes are uniformly incorporated into the calculation process of the compatibility matrix. The non-negative matching degree reflects the positive adaptation relationship between task characteristics and node capabilities, the dynamic adaptation coefficient reflects the impact of runtime load and cost on the degree of adaptation, and the zero-value setting when hardware is not supported excludes combinations of task nodes that do not meet the execution conditions. Therefore, the generated compatibility matrix has good comparability and runtime sensitivity, and can more accurately represent the degree of adaptation between the initial task unit and the processing node.

[0125] Regarding the implementation details of calculating the orthogonal conflict degree in the method of this application embodiment, in some examples of this application embodiment, for any two initial task units to be assigned to the same processing node in the candidate scheduling scheme, the orthogonal conflict degree is determined based on the normalized directional similarity between the corresponding orthogonal feature vectors; wherein, the normalized directional similarity is determined based on the absolute correlation of the inner product between the two orthogonal feature vectors, and the influence of vector magnitude difference on the orthogonal conflict degree is weakened by normalization processing.

[0126] It should be noted that in scenarios such as industrial automation edge control, heterogeneous distributed communication, or concurrent processing of multiple service streams, although multiple initial task units may have a high degree of compatibility with the same processing node, their underlying resource request patterns may still overlap. For example, multiple high-frequency sampling tasks may simultaneously occupy the data preprocessing queue of a near-edge node, and multiple image detection tasks may simultaneously occupy the GPU memory, cache, or inference computing unit of the same acceleration node. Therefore, allocating resources solely based on the compatibility between a single task and a single processing node is insufficient to fully reflect the resource competition that may arise when multiple tasks are executed concurrently on the same processing node.

[0127] In this embodiment, the orthogonal feature vectors corresponding to each initial task unit are used to evaluate the directional similarity of any two initial task units to be assigned to the same processing node. Since the orthogonal feature vectors have already orthogonally represented the multidimensional original features of the initial task units, the directional similarity between two orthogonal feature vectors can be used to characterize the degree of overlap between the corresponding two initial task units in terms of resource request patterns, latency requirements, or business processing characteristics. A higher directional similarity indicates a higher probability of resource contention when the two tasks are executed simultaneously on the same processing node.

[0128] For example, based on orthogonal eigenvectors and The absolute value of the normalized inner product between them determines the first... The initial task unit and the first Orthogonal conflict degree between initial task units :

[0129] Equation (15)

[0130] In the formula, Describing the vector norm, Represents the absolute value of the algebraic dot product of vectors. This is a preset positive number used to avoid a denominator of zero.

[0131] In equation (15) above, the inner product is used to reflect the correlation between two orthogonal feature vectors in terms of direction. Using the absolute value of the inner product allows for the simultaneous identification of non-orthogonal relationships in both positive and negative correlation directions, avoiding underestimation of the feature coupling between tasks due to the cancellation of positive and negative directions. Normalization using vector norm products can weaken the impact of differences in the magnitudes of orthogonal feature vectors on the calculation of conflict degree, making the orthogonal conflict degree more focused on reflecting the directional similarity between two tasks. This is achieved by introducing a preset positive number. This can improve computational stability when zero vectors or vectors with minimal magnitudes are present, and avoid anomalies caused by zero denominator.

[0132] Accordingly, regarding the implementation details of determining the target scheduling scheme in the method of the embodiments of this application, in some examples of the embodiments of this application, the orthogonal conflict degree is compared with a preset conflict threshold, and when the orthogonal conflict degree reaches or exceeds the preset conflict threshold, the candidate allocation relationship of the two initial task units to be allocated to the same processing node is determined as the candidate allocation relationship to be penalized; a global scheduling cost function including the basic scheduling cost term and the orthogonal conflict penalty term is constructed.

[0133] For example, the global scheduling cost function It can be expressed by the following formula:

[0134] Equation (16)

[0135] In the formula, , represents the set of optional candidate assignment relationships; This represents the global cost evaluation value of the candidate scheduling scheme; and Assign state variables to Boolean-type scheduling, when the corresponding initial task unit is to be assigned to the first... The value is 1 when processing a node, and 0 otherwise. ,and This represents the total number of initial task units; To process the index of the node, and This represents the total number of nodes to be processed. These are the matrix elements in the compatibility matrix; It is a delay-sensitive factor; For the first The initial task unit in the first The expected execution delay on each processing node; For the first The remaining deadline constraints for each initial task unit, and ; Let be the orthogonal penalty coefficient; representing the condition for satisfying . The initial task units are combined and summed using a one-way traversal. As a conflict indicator variable, when orthogonal conflict degree When the preset conflict conditions are met, ,otherwise Among them, the preset conflict conditions include , This is a preset conflict threshold.

[0136] Here, the basic scheduling cost term is negatively correlated with the fit between the initial task unit and the processing node, and is related to the expected execution delay and remaining deadline constraints of the initial task unit on the corresponding processing node; the orthogonal conflict penalty term is used to improve the global cost evaluation value of the candidate scheduling scheme containing the candidate allocation relationship to be penalized.

[0137] Refer to the global scheduling cost function in equation (16) The example consists of two parts: the left side is the basic scheduling cost term, which reflects the matching degree between tasks and nodes based on the inverse of the compatibility matrix, and quantitatively describes the time pressure of tasks by combining the ratio of expected execution delay to deadline; the right side is the orthogonal conflict penalty term, which is indicated by an indicator variable. and conflict level The implicit feature interference between multiple tasks under the same node was quantified, and the penalty coefficient was used. Adjust the intensity of intervention.

[0138] To make the optimization solution boundary more explicit, the above allocation constraints can be expressed in the solution conditions at the system's underlying level as follows:

[0139] Equation (17)

[0140] Equation (18)

[0141] Equation (19)

[0142] In the formula, Indicates the first Resource consumption requirements of each initial task unit Indicates the first Available resource capacity of each processing node.

[0143] During the optimization process, the system aims to reduce the global scheduling cost evaluation value and determines the target scheduling scheme within the candidate allocation relationship set and the aforementioned allocation constraints. If a candidate allocation relationship contains a task combination whose orthogonal conflict degree reaches or exceeds a preset conflict threshold, its corresponding orthogonal conflict penalty term will increase the global cost evaluation value of the candidate scheduling scheme. This guides the algorithm to actively avoid such high-conflict concurrent allocation methods during optimization, ultimately achieving comprehensive coordination between task priority, node adaptability, and the risk of conflict between tasks on the same node.

[0144] Figure 3 A flowchart illustrating an example of determining a target scheduling scheme and generating a scheduling matrix in a method according to an embodiment of this application is shown.

[0145] like Figure 3 As shown, in step S310, the scheduling allocation state, the set of allocated initial task units for each processing node, and the remaining available resource capacity are initialized, and a task scheduling queue is constructed in descending order of the dynamic scheduling priority weight of each initial task unit.

[0146] Specifically, at the start of the current scheduling cycle, the system initializes the scheduling allocation status between each initial task unit and each processing node to an unallocated state, and initializes the allocated initial task unit set of each processing node to an empty set. Simultaneously, the system determines the remaining available resource capacity of each processing node based on node running status data. Subsequently, the system sorts each initial task unit according to its dynamic scheduling priority weight from high to low, prioritizing initial task units with higher urgency, shorter remaining deadlines, or greater susceptibility to network conditions.

[0147] In step S320, each initial task unit is traversed according to the sorting order of the task scheduling queue. For the currently traversed initial task unit, effective processing nodes are extracted from the optional candidate allocation relationship indicated by the compatibility matrix, and a set of candidate nodes is selected from the effective processing nodes based on the basic cost increment corresponding to the basic scheduling cost item.

[0148] For example, for the currently traversed th The system can select from the compatibility matrix a set of initial task units that meet the following requirements: The processing node is determined as the valid processing node, and the processing node is calculated respectively. The base cost increment corresponding to each initial task unit allocated to each effective processing node :

[0149] Equation (20)

[0150] In the formula, This indicates the compatibility matrix corresponding to the first... The initial task unit and the first The matrix elements of each effective processing node. Indicates the delay sensitivity factor. Indicates the first The initial task unit in the first The expected execution latency on each effective processing node Indicates the first The remaining deadline constraints for each initial task unit, and .in, Used to reflect the impact of fit on the base cost. This is used to reflect the pressure of the expected execution delay relative to the remaining deadline. The system can select a preset number of valid processing nodes to form a candidate node set in ascending order of the base cost increment.

[0151] In step S330, for each candidate processing node in the candidate node set, the scheduling penalty increment is determined based on the orthogonal conflict degree between the initial task unit set already allocated on the candidate processing node and the initial task unit currently being traversed, and the total marginal cost of the corresponding candidate processing node is determined in combination with the basic cost increment.

[0152] For example, scheduling penalty increment It can be expressed by the following formula:

[0153] Equation (21)

[0154] In the formula, Indicates the current candidate processing node The initial set of assigned task units. For set The index of the initial task unit has been allocated. Denotes the orthogonal penalty coefficient. Indicates the first The initial task unit and the first The orthogonal conflict degree between the assigned initial task units This indicates a conflict indicator variable.

[0155] In some examples, conflict indicator variables It can be determined according to the following conditions:

[0156] Equation (22)

[0157] In the formula, This represents the preset conflict threshold. When no initially assigned task unit exists on the candidate processing node, the scheduling penalty increment can be zero. Correspondingly, the system calculates the penalty based on the base cost increment. With scheduling penalty increment Calculate the total marginal cost of the corresponding candidate processing node. :

[0158] Equation (23)

[0159] Therefore, the total marginal cost reflects both the basic adaptation cost between the current initial task unit and the candidate processing node, and the orthogonal conflict impact between the current initial task unit and the tasks already assigned on the candidate processing node.

[0160] In step S340, the candidate processing node with the lowest total marginal cost and satisfying the available resource capacity constraint is selected from the candidate node set as the target processing node; if there is no candidate processing node that satisfies the available resource capacity constraint in the candidate node set, the processing node that satisfies the available resource capacity constraint and has the lowest total marginal cost is selected from the remaining valid processing nodes other than the candidate node set as the target processing node.

[0161] Specifically, the system can first determine whether the remaining available resource capacity of each candidate processing node in the candidate node set can meet the resource requirements of the current initial task unit, and then select the node with the lowest total marginal cost among the candidate processing nodes that meet the capacity constraint. If none of the processing nodes in the candidate node set meet the capacity constraint, the system can continue to evaluate the marginal cost and determine the capacity constraint of the remaining valid processing nodes, and select the processing node that meets the capacity constraint and has the lowest total marginal cost as the target processing node. If there is still no processing node that meets the conditions, the current initial task unit can be placed in the waiting queue and delayed to the next scheduling cycle for processing, or a resource expansion operation can be triggered, according to the system configuration.

[0162] In step S350, the currently traversed initial task units are assigned to the target processing node, and the corresponding scheduling allocation status, the set of allocated initial task units of the target processing node, and the remaining available resource capacity are updated.

[0163] Specifically, after determining the target processing node, the system updates the scheduling allocation status between the current initial task unit and the target processing node to an allocated status, and adds the current initial task unit to the allocated initial task unit set of the target processing node. Simultaneously, the system updates the remaining available resource capacity of the target processing node based on the resource requirements of the current initial task unit, ensuring that the node status within the same scheduling cycle changes synchronously with the task allocation process.

[0164] In step S360, the allocation traversal of all initial task units in the task scheduling queue is completed, the target scheduling scheme is obtained based on all updated scheduling allocation states, and the scheduling matrix is ​​generated and output according to the target scheduling scheme.

[0165] Specifically, the system continuously executes the node selection and status update process described above according to the task scheduling queue until all initial task units have completed their allocation or the system's configured waiting, delay, or exception handling conditions are met. After the traversal is complete, all updated scheduling allocation states together constitute the target scheduling scheme. The system can convert the target scheduling scheme into a scheduling matrix, where the elements of the scheduling matrix represent whether a final allocation relationship exists between the initial task units and the processing nodes. In some examples, the generated scheduling matrix can be read by the task distribution module, execution scheduling module, or resource control module to complete the allocation control of the corresponding tasks.

[0166] Through the embodiments of this application, the system determines the task processing order according to the dynamic scheduling priority, and comprehensively considers the basic cost increment, scheduling penalty increment and available resource capacity constraints during the node selection process of each task. This can reduce the selection of low-fitness nodes or nodes with insufficient resources, and reduce the possibility of high-conflict tasks being concentrated on the same processing node, so that the generated scheduling matrix is ​​more in line with the task requirements and node status of the current heterogeneous computing environment.

[0167] In some examples of embodiments of this application, after generating and outputting the scheduling matrix, a closed-loop state feedback operation can also be performed.

[0168] It should be noted that, since the business load, node resource status, and network transmission status in a heterogeneous computing environment change over time, if the orthogonal penalty coefficient remains fixed for a long period, two types of problems may arise: First, when task conflicts are strong or transmission anomalies increase, the fixed penalty strength may be insufficient to suppress the concentrated allocation of tasks with similar characteristics; second, when system resource utilization is low and task execution status is relatively stable, an excessively high penalty strength may cause some processing nodes to be underutilized. Therefore, this embodiment periodically updates the orthogonal penalty coefficient based on actual execution results, allowing it to adjust with changes in operating status.

[0169] Specifically, firstly, the system periodically collects data on task completion long-tail latency, task retransmission rate, and overall resource utilization in a heterogeneous computing environment. The task completion long-tail latency is a preset quantile latency for task completion, such as the 95th or 99th quantile latency. Compared to the average latency, the preset quantile latency more fully reflects the execution status of tasks with higher latency. The task retransmission rate characterizes the frequency of retries, retransmissions, or anomaly recovery during task transmission or execution, while the overall resource utilization characterizes the degree to which the overall resources of the processing nodes are occupied. By simultaneously collecting these metrics, the current scheduling results can be evaluated from three aspects: execution latency, transmission stability, and resource utilization status.

[0170] Then, within the current feedback control cycle, obtain the current long-tail delay value. Current retransmission rate value and current resource utilization rate In some examples, the system can perform time window statistics, moving averages, or filtering on the acquired raw telemetry data to reduce the impact of instantaneous fluctuations on parameter updates, so that the indicators used in the current feedback control cycle can more stably reflect the system's operating status.

[0171] Then, based on the current long-tail latency value Current retransmission rate value and current resource utilization rate The orthogonal penalty coefficients in the global scheduling cost function are updated online.

[0172] For example, the orthogonal penalty coefficients in the global scheduling cost function can be adjusted according to the following dynamic evolution function. Perform online updates:

[0173] Equation (24)

[0174] Equation (25)

[0175] Equation (26)

[0176] Equation (27)

[0177] Equation (28)

[0178] In the formula, This represents the orthogonal penalty coefficient corresponding to the next feedback control cycle. This represents the orthogonal penalty coefficient corresponding to the current feedback control cycle. This is the index of the current feedback control cycle. and These are the preset lower and upper limits of the penalty coefficient, respectively. ; This indicates the increment for correcting the penalty coefficient; Indicates the retransmission penalty driver; Indicates the delay penalty driver; This represents the driver of utilization decay; This represents the retransmission rate threshold; This indicates the upper limit threshold for long-tail latency; This indicates the lower limit threshold for resource utilization. , , This is a preset dynamic adjustment step size factor with dimensional alignment function, and , , ; This is a status indicator function. It takes the value 1 when the logical condition within the parentheses is true, and 0 otherwise.

[0179] In the aforementioned dynamic evolution function, the retransmission penalty driving term and delay penalty driver This is used to add an orthogonal penalty coefficient when the current retransmission rate or the current long-tail delay value exceeds a corresponding threshold. Specifically, when... When this occurs, it indicates that the task retransmission or abnormal recovery is significant, and the system can then... The degree of deviation increases the orthogonal penalty coefficient; when When this occurs, it indicates that the long-tail delay in task completion exceeds the preset limit, and the system can then... The degree of deviation increases the orthogonal penalty coefficient. Therefore, the orthogonal conflict degree has a stronger impact on the global scheduling cost function, and the scheduling cost increases when tasks with similar characteristics are assigned to the same processing node.

[0180] Utilization decay driver This is used to reduce the orthogonality penalty coefficient when resource utilization is low and long-tail delay and retransmission rate are within limits. Specifically, when , and If the system's current resource utilization is low, while latency and retransmission metrics remain within acceptable limits, the orthogonal penalty coefficient can be appropriately reduced to mitigate resource idleness caused by excessive task distribution. This conditional constraint prevents the reduction of conflict penalty intensity simply to improve resource utilization when latency or retransmission is already abnormal.

[0181] During the update process, , , This is used to adjust the degree of influence of different monitoring indicators on the correction magnitude of the orthogonal penalty coefficient. If the system prioritizes transmission stability, the value can be appropriately increased. If the system is more concerned about task completion delay, the time limit can be increased appropriately. If the system focuses more on resource utilization, the capacity can be appropriately increased. In some examples, the step size factor mentioned above can be preset based on the business type, node resource scale, or historical operating experience.

[0182] Therefore, the updated orthogonal penalty coefficient is limited to the lower limit of the preset penalty coefficient. With the preset penalty coefficient upper limit The updated orthogonal penalty coefficients are then used to optimize the global scheduling cost function in the next feedback control cycle. By limiting the amplitude, the orthogonal penalty coefficients are prevented from continuously increasing or decreasing under prolonged congestion or idle conditions, thus maintaining the numerical stability of the orthogonal conflict penalty term in the global scheduling cost function. The updated orthogonal penalty coefficients can be used to optimize the global scheduling cost function in the next feedback control cycle, allowing the scheduling results to be adjusted according to changes in the operating state.

[0183] Through the embodiments of this application, the system can update the orthogonal penalty coefficient online based on the long-tail latency of task completion, the task retransmission rate, and the overall resource utilization. When the execution status exhibits high latency or high retransmission risk, the orthogonal penalty coefficient increases, raising the cost of centrally allocating tasks with similar characteristics. When the resource utilization is low and the latency and retransmission indicators are within the allowable range, the orthogonal penalty coefficient decreases, making the scheduling results more resource-utilization-oriented. This improves the consistency between the orthogonal penalty coefficient and the actual operating status, and enhances the stability and adaptability of the scheduling results in dynamic load environments.

[0184] Figure 4 A schematic diagram illustrating the system operation mechanism of an example of a heterogeneous node digital orthogonal scheduling method adapted to hybrid data stream characteristics according to an embodiment of this application is shown.

[0185] like Figure 4 As shown, the system first receives heterogeneous service data from diverse sources, such as sensor streams, video streams, and control commands, at the data access side. This data is then aggregated and classified into mixed data streams to be scheduled, providing standardized initial task units for subsequent unified scheduling. Subsequently, the system sequentially performs orthogonal feature extraction, dynamic weight calculation, and compatibility matrix generation on each initial task unit. Through low-level feature transformation, the complex and intertwined service requirements are accurately transformed into mathematically decoupled orthogonal feature representations. Combined with real-time network status and underlying hardware capabilities, a multi-dimensional quantified task priority weight and node matching degree matrix is ​​output.

[0186] Based on this, the digital orthogonal scheduler makes optimal decisions based on the parameters output above. To address the risk of local resource contention arising from multiple similar tasks being assigned to the same processing node, the scheduler introduces an orthogonal conflict penalty mechanism for interception and cost constraints, thus preventing excessive overlap and mutual interference of underlying physical resources. After cost evaluation and adjustment, the final scheduling scheme is transformed into precise scheduling instructions, driving heterogeneous nodes such as CPUs and GPUs to complete accurate task allocation and high-concurrency parallel execution. Furthermore, the system mechanism also deploys a resource and status monitoring module to construct a globally closed-loop dynamic adjustment feedback link. This monitoring module collects key physical states during the underlying cluster operation in real time and continuously transmits the evaluation results back through the status feedback and adjustment link. This allows the system to adaptively adjust scheduling parameters such as dynamic weights online when congestion changes or load shifts occur in the operating environment, enabling the heterogeneous node scheduling mechanism to approach an optimal operating range that balances throughput and isolation.

[0187] At the level of overall collaborative operation mechanism Figure 4 The demonstrated architecture also embodies the two-layer decoupling and synergy between the forward real-time scheduling pipeline and the reverse periodic parameter evolution loop on the time scale. Forward high-concurrency data processing and digital orthogonal scheduling decisions are driven by the instantaneous arrival of task units, executing rapid responses to ensure that the mixed data streams can be allocated adaptively in real time. Meanwhile, reverse state feedback and adjustments evolve slowly based on a preset control cycle, progressively correcting the boundaries of the global scheduling cost function through smoothed and filtered statistical indicators. The two-layer timescale design avoids frequent oscillations in the parameter adjustment loop due to transient network noise, ensuring the control stability of the scheduling system during long-term operation. It also significantly reduces the additional computing power overhead of monitoring calculations on forward business processing, achieving efficient and robust on-demand allocation of distributed computing resources.

[0188] To verify the effectiveness and boundary conditions of the heterogeneous node digital orthogonal scheduling method adapted to hybrid data flow characteristics provided in this application embodiment under real-world complex environments, the system designed multiple sets of comparative experiments based on a discrete event simulator that combines cloud and edge computing underlying logic. For the selection of benchmark algorithms, the experiments used the classic Heterogeneous Earliest Finish Time (HEFT) scheduling algorithm and the dynamic priority scheduling algorithm as comparative references. Unlike traditional small-scale static task graph tests, the experimental environment in this embodiment focuses on simulating a high-concurrency, dynamically arriving heterogeneous collaborative computing environment.

[0189] In the specific hardware topology and data stream generation settings, the simulation environment constructed a computing cluster containing 50 heterogeneous processing nodes, specifically including 30 general-purpose CPU nodes, 10 GPU acceleration nodes, and 10 field-programmable gate arrays (FPGAs) or edge low-power nodes. The network bandwidth and transmission latency between nodes were configured to dynamically fluctuate. On the data stream access side, the experiment used a Poisson process to simulate the dynamic arrival characteristics of mixed data streams, processing approximately 10,000 initial task units per simulation cycle. Simultaneously, the task types were strictly distributed according to the business ratios of real industrial and communication scenarios, with high-frequency sensor streams accounting for 40%, high-throughput video streams for 30%, emergency control commands for 20%, and background log streams for 10%.

[0190] To objectively reflect the processing quality and isolation effect of multi-dimensional task scheduling on heterogeneous nodes, the evaluation metrics in this comparative experiment not only include the conventional average processing latency, but also focus on introducing long-tail latency of task completion (such as 95th percentile latency), isolation and anti-interference performance when heterogeneous tasks are concurrent on the same node, and the scheduler's own algorithm overhead as core comparison dimensions. Through the above parameter settings that closely resemble real-world business conditions, the actual performance gain and robustness of the orthogonal conflict penalty mechanism and dynamic compatibility matching strategy described in this application under extremely high load conditions can be comprehensively and rigorously quantified.

[0191] Figure 5 This diagram illustrates a comparative simulation of the end-to-end latency cumulative distribution and long-tail characteristics of different scheduling methods. The simulation results, presented in the form of a cumulative distribution function, demonstrate the differences in end-to-end latency distribution characteristics between the Orthogonal Digital Scheduling Algorithm (ODSA) provided in this application and the benchmark Heterogeneous Earliest Complete Time (HEFT) and Dynamic Priority (DP) scheduling algorithms under a high-concurrency mixed data flow scenario where the system computational load reaches 80%.

[0192] like Figure 5 As shown in the cumulative probability curves, in terms of performance in routine task processing, the median latency of the HEFT and DP algorithms, i.e., the latency corresponding to a cumulative probability of 50%, is approximately 35ms; while the median latency of the ODSA method in this embodiment is significantly lower than the two benchmark methods, at approximately 20ms. Therefore, it is evident that the ODSA method performs better in routine end-to-end latency control for most tasks.

[0193] Furthermore, for the long-tail latency characteristic that is more sensitive to critical business operations, namely the 95th percentile latency corresponding to a cumulative probability of 95%, the three methods exhibit more significant differences. The ODSA method in this embodiment achieves a 95% cumulative probability at a latency of approximately 45ms; in contrast, the DP algorithm and the HEFT algorithm achieve a 95% cumulative probability at approximately 71ms and 82ms, respectively. This indicates that in high-load mixed data flow scenarios, the latency distribution curve of the ODSA method is generally closer to the low-latency range and can effectively shorten the long-tail latency.

[0194] It should be noted that the HEFT algorithm mainly focuses on the estimated completion time of tasks on heterogeneous processing nodes, while the DP algorithm mainly focuses on dynamic priority ranking. Neither of them adequately considers the overlapping characteristics of multiple concurrent data streams on the same processing node or shared resource path. In high-concurrency scenarios, high-throughput video streams, high-frequency sensor streams, and low-latency control command streams may be concentrated on the same processing node or adjacent resource paths, leading to cache contention, queue backlog, communication resource occupation, or execution waiting, thereby lengthening the end-to-end completion time of some tasks.

[0195] In contrast, the ODSA method in this application, based on task and node adaptation evaluation, introduces orthogonal feature vectors, a compatibility matrix, and orthogonal conflict penalties to quantify the feature overlap between initial task units to be assigned to the same processing node and suppress high-conflict candidate assignment relationships. As a result, the system can reduce the concentrated allocation of feature-similar tasks on the same processing node, making the task completion delay distribution more concentrated and reducing the risk of long-tail delays under high load conditions. This verifies the delay control effect of the method in this application in mixed data stream concurrent scheduling scenarios.

[0196] Figure 6 This figure illustrates a comparative simulation of the orthogonal isolation effect and system overhead trade-off between different scheduling methods in a scenario of sudden changes in background traffic. The main body of the figure uses a dual Y-axis multi-line trend graph to show the impact of sudden changes in background video stream load on the real-time average latency of control commands. The left primary Y-axis represents the background video stream throughput in Mbps; the right secondary Y-axis represents the real-time average latency of control commands in milliseconds (ms); and the horizontal axis represents the simulation time in seconds (s). The figure compares the latency changes of the Digital Orthogonal Scheduling Method (ODSA), the Heterogeneous Earliest Complete Time (HEFT) scheduling algorithm, and the Dynamic Priority Scheduling (DP) algorithm provided in the embodiments of this application.

[0197] like Figure 6As shown, at approximately 10 seconds into the simulation, the system simulates a sudden load surge in the background video stream, increasing the throughput from approximately 500 Mbps to approximately 1500 Mbps, roughly three times the original load. Under this sudden load condition, the average real-time latency of control commands for both the HEFT and DP algorithms shows a significant increase, fluctuating primarily between approximately 60 ms and 90 ms in the subsequent period. In contrast, the control command latency of the ODSA method in this embodiment only shows a brief increase after the load surge, subsequently remaining at a low level of approximately 20 ms.

[0198] The reason for these differences lies in the fact that the HEFT algorithm mainly schedules tasks based on the estimated completion time on heterogeneous nodes, while the DP algorithm mainly schedules tasks based on dynamic priority. Neither of these algorithms adequately considers the overlapping characteristics of background video streams and control command streams on the same processing node or shared resource path. When the load on the background video stream suddenly increases, if the control command stream and the high-throughput background video stream compete for resources such as cache, communication queues, memory bandwidth, or computing units, it can easily lead to increased control command latency and fluctuations.

[0199] In contrast, the ODSA method in this application, based on task and node adaptation evaluation, introduces a compatibility matrix and orthogonal conflict penalty to quantify the degree of feature overlap between different service flows and suppress high-conflict candidate allocation relationships. As a result, the system can reduce the possibility of high-throughput background video streams and low-latency control command streams occupying the same processing resources, ensuring that control commands maintain a low latency level even under sudden changes in background traffic, thus demonstrating better orthogonal isolation and resistance to load disturbances.

[0200] also, Figure 6 The embedded bar chart in the lower right corner compares the time taken for a single scheduling decision by different methods. As shown in the chart, the HEFT algorithm takes approximately 0.4 ms for a single scheduling decision, while the ODSA method takes approximately 2.5 ms. This result indicates that the ODSA method introduces a certain scheduling computational overhead compared to the HEFT algorithm due to the need to perform orthogonal feature processing, compatibility matrix evaluation, and orthogonal conflict penalty calculations.

[0201] therefore, Figure 6 The technical trade-offs of the proposed method are illustrated from the perspectives of latency stability and scheduling computational overhead: In scenarios with sudden changes in background traffic and high-concurrency mixed data streams, the ODSA method, at the cost of increasing the time required for a single scheduling decision, reduces the impact of sudden changes in background video stream load on the control command stream, thus keeping the average real-time latency of control commands within a low range. These experimental results support the effectiveness of the proposed method in improving the stability of low-latency services in complex heterogeneous concurrent service scenarios.

[0202] Figure 7 The simulation results illustrate an example of the trade-off between the parameter sensitivity of the orthogonal penalty coefficient and resource utilization. The simulation results are presented as a two-dimensional contour heatmap, showing the relationship between the network load factor and the orthogonal penalty coefficient. The combined impact on the overall system return rate. Here, the horizontal axis represents the network load factor, and the vertical axis represents the orthogonal penalty coefficient. The color bar on the right represents the overall return on the system, and the contour lines in the figure are used to represent the relative changes in the overall return on the system under different parameter combinations.

[0203] like Figure 7 As shown, the orthogonal penalty coefficient If the value is too low or too high, the overall system return rate may decrease. When the orthogonal penalty coefficient is... When the network load factor is low, the scheduling process is insufficient to suppress overlapping tasks. As the network load factor increases, multiple data streams with similar resource requirements are prone to clustering at the same processing node or sharing resource paths, resulting in a low overall system return rate. It should be noted that... Figure 7 The region where the algorithm degenerates into greedy matching, corresponding to the low-to-medium penalty range, reflects this type of situation.

[0204] On the other hand, when the orthogonal penalty coefficient At higher levels, the system tends to excessively suppress candidate assignment relationships with certain similarities. While this reduces the risk of resource contention within the same node, some available processing nodes may not be fully utilized due to overly strong isolation constraints, leading to decreased resource utilization. The "extremely isolation-emphasized" region corresponding to the high penalty range in the figure indicates that the system may experience insufficient resource utilization under these parameter conditions, for example, resource utilization below 65%.

[0205] Furthermore, from Figure 7 The highlighted areas show that, under medium to high network load factors, the orthogonal penalty coefficient... When the overall system return rate is around 0.35, a high contour ridge is formed. This region indicates that a moderate orthogonal penalty strength can achieve a good trade-off between suppressing the concentrated allocation of highly conflicting tasks and maintaining the resource utilization of processing nodes. If the penalty strength is too low, the system will have difficulty effectively suppressing the concentration of homogeneous tasks; if the penalty strength is too high, it may lead to excessive isolation and resource idleness.

[0206] Therefore, a fixed orthogonal penalty coefficient is insufficient to simultaneously adapt to scheduling requirements under different network load conditions. This simulation result supports the technical design in this application embodiment, which involves online updates of the orthogonal penalty coefficient based on task completion tail latency, task retransmission rate, and overall resource utilization. By dynamically adjusting the orthogonal penalty coefficient, the system can approach a more optimal benefit range when network load changes, maintaining a more reasonable balance between orthogonal isolation effectiveness and resource utilization in the scheduling results.

[0207] In summary, this application provides a heterogeneous node digital orthogonal scheduling method and system adapted to the characteristics of mixed data streams. It overcomes the limitations of traditional heuristic scheduling algorithms in modeling the characteristics of mixed data streams by introducing a weighted Gram-Schmidt orthogonalization mechanism and a dynamic compatibility matrix. This accurately transforms the complex and intertwined business attributes and resource requirements of data streams into mathematically decoupled orthogonal feature representations. Furthermore, by introducing an explicit orthogonal penalty term into the global scheduling cost function, coupled with a lightweight marginal cost assessment and closed-loop state feedback mechanism, the system can reduce the risk of homogeneous task concentration and resource contention in heterogeneous computing resource allocation, and improve the stability of scheduling results in multi-service concurrent scenarios.

[0208] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of combined actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0209] Figure 8 A structural block diagram of an example of a heterogeneous node digital orthogonal scheduling system adapted to hybrid data stream characteristics, according to an embodiment of this application, is shown.

[0210] like Figure 8 As shown, the heterogeneous node digital orthogonal scheduling system 800 adapted to the characteristics of mixed data streams includes a data acquisition and classification module 810, an orthogonal feature mapping module 820, a priority weight calculation module 830, a compatibility matrix generation module 840, a candidate scheduling generation module 850, and a conflict penalty adjustment module 860.

[0211] The data acquisition and classification module 810 is used to acquire the mixed data stream to be processed, the network status data of the current system, and the node data of each processing node in the heterogeneous computing environment, and classify the mixed data stream according to business attributes to generate multiple initial task units; wherein, the node data includes node attribute data and node running status data.

[0212] The orthogonal feature mapping module 820 is used to extract multidimensional original features for each of the initial task units, generate an orthogonal basis for constructing an orthogonal feature space based on the multidimensional original features, and perform orthogonal mapping on the multidimensional original features based on the orthogonal basis to generate orthogonal feature vectors corresponding to each of the initial task units.

[0213] The priority weight calculation module 830 is used to calculate the dynamic scheduling priority weight of each initial task unit based on the service attributes, the multidimensional original features, and the network status data of each initial task unit.

[0214] The compatibility matrix generation module 840 is used to generate a node capability vector corresponding to each processing node based on the node data of each processing node, and to evaluate the matching degree between the orthogonal feature vector of each initial task unit and the node capability vector of each processing node, thereby generating a compatibility matrix for quantifying the degree of adaptation between the initial task unit and the processing node.

[0215] The candidate scheduling generation module 850 is used to combine the dynamic scheduling priority weight and the compatibility matrix to generate a candidate scheduling scheme that includes the candidate allocation relationship between the initial task unit and the candidate processing node, and to calculate the orthogonal conflict degree based on the corresponding orthogonal feature vector for at least two initial task units in the candidate scheduling scheme that are to be allocated to the same processing node.

[0216] The conflict penalty adjustment module 860 is used to apply a scheduling penalty to the candidate allocation relationship in the candidate scheduling scheme that is intended to allocate at least two initial task units to the same processing node when the orthogonal conflict degree meets the preset conflict conditions, and to filter or adjust the candidate allocation relationship based on the scheduling penalty to obtain the target scheduling scheme, and to determine the target processing node of each initial task unit according to the target scheduling scheme to generate and output the scheduling matrix.

[0217] In some embodiments, this application provides a non-volatile computer-readable storage medium storing one or more programs including execution instructions. The execution instructions can be read and executed by an electronic device (including but not limited to a computer, server, or network device) to perform the steps of any of the heterogeneous node digital orthogonal scheduling methods adapted to the characteristics of mixed data streams described above.

[0218] In some embodiments, this application also provides a computer program product, the computer program product including a computer program stored on a non-volatile computer-readable storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the steps of any of the above-described heterogeneous node digital orthogonal scheduling methods adapted to hybrid data stream characteristics.

[0219] In some embodiments, this application also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform steps of a heterogeneous node digital orthogonal scheduling method adapted to hybrid data stream characteristics.

[0220] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0221] The electronic devices in this application can exist in various forms, including but not limited to: mobile communication devices, ultra-mobile personal computer devices, portable entertainment devices, or other airborne electronic devices with data interaction functions.

[0222] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0223] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0224] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A heterogeneous node digital orthogonal scheduling method adapted to the characteristics of mixed data streams, characterized in that, The method includes: The system acquires the mixed data stream to be processed, the network status data of the current system, and the node data of each processing node in the heterogeneous computing environment, and classifies the mixed data stream according to business attributes to generate multiple initial task units; wherein, the node data includes node attribute data and node running status data; For each of the initial task units, multidimensional original features are extracted. Based on the multidimensional original features, an orthogonal basis for constructing an orthogonal feature space is generated. Based on the orthogonal basis, the multidimensional original features are orthogonally mapped to generate orthogonal feature vectors corresponding to each of the initial task units. Based on the business attributes, multidimensional original features, and network status data of each initial task unit, the dynamic scheduling priority weight of each initial task unit is calculated. Based on the node data of each processing node, a node capability vector corresponding to each processing node is generated. The matching degree of the orthogonal feature vector of each initial task unit and the node capability vector of each processing node is evaluated to generate a compatibility matrix for quantifying the degree of adaptation between the initial task unit and the processing node. Specifically, for any initial task unit and any processing node, a non-negative matching degree is determined based on the matching relationship between the orthogonal feature vector of the initial task unit and the node capability vector of the processing node. The matrix elements in the compatibility matrix are calculated based on the non-negative matching degree and a dynamic adaptation coefficient. The dynamic adaptation coefficient characterizes the attenuation effect of the real-time load rate of the processing node and the allocation or migration cost of the initial task unit on the degree of adaptation, such that the higher the real-time load rate of the processing node and the greater the allocation or migration cost, the lower the corresponding degree of adaptation. By combining the dynamic scheduling priority weights and the compatibility matrix, a candidate scheduling scheme is generated that includes the candidate allocation relationship between the initial task units and the candidate processing nodes. For at least two initial task units in the candidate scheduling scheme that are to be allocated to the same processing node, the orthogonal conflict degree is calculated based on the corresponding orthogonal feature vectors. When the orthogonal conflict degree meets the preset conflict condition, a scheduling penalty is applied to the candidate allocation relationship in the candidate scheduling scheme that is intended to allocate at least two initial task units to the same processing node, and the candidate allocation relationship is filtered or adjusted based on the scheduling penalty to obtain the target scheduling scheme. The target processing node of each initial task unit is determined according to the target scheduling scheme to generate and output the scheduling matrix.

2. The method according to claim 1, characterized in that, The process of extracting multidimensional original features for each initial task unit, generating an orthogonal basis for constructing an orthogonal feature space based on the multidimensional original features, and performing orthogonal mapping on the multidimensional original features based on the orthogonal basis to generate an orthogonal feature vector corresponding to each initial task unit includes: Extract the multidimensional original features of each initial task unit, and normalize the multidimensional original features to obtain the normalized feature vector corresponding to each initial task unit; the multidimensional original features include at least time domain features, bandwidth features, and service priority features. Construct a feature weight matrix to characterize the weight of different feature dimensions in scheduling decisions; wherein the feature weight matrix is ​​a positive definite diagonal matrix, and the diagonal elements of the feature weight matrix are positively correlated with the latency sensitivity and / or service priority of the corresponding feature dimension; Multiple linearly independent reference eigenvectors are selected or constructed from the normalized eigenvectors as a support set, and an orthogonal basis is generated based on the weighted Graham-Schmidt algorithm using the eigenweight matrix. In the process of generating the orthogonal basis, for the current reference feature vector, the weighted correlation between the current reference feature vector and the generated orthogonal basis vectors is measured based on the feature weight matrix, and the weighted projection component of the current reference feature vector in the direction of the generated orthogonal basis vectors is removed to obtain the orthogonal intermediate vector; then the orthogonal intermediate vector is normalized based on the feature weight matrix to obtain the corresponding orthogonal basis vector. Each of the normalized feature vectors is projected onto the orthogonal feature space formed by the orthogonal basis vectors, and the weighted projection result of each of the normalized feature vectors onto the orthogonal basis vectors is used as the corresponding orthogonal feature vector; wherein, the orthogonal feature vector is used to calculate the orthogonal conflict degree between different initial task units.

3. The method according to claim 1, characterized in that, The calculation of the dynamic scheduling priority weight for each initial task unit based on its service attributes, multidimensional original features, and network state data includes: For each initial task unit, feature data representing the corresponding data stream acquisition frequency, data resolution, and / or data packet size are parsed from multidimensional raw features and normalized to construct a complexity index; resource requirement index is constructed based on feature data representing computational and storage requirements from the multidimensional raw features; remaining deadline constraints are determined based on time-domain expiration information from the multidimensional raw features; a preset urgency level score is matched based on business attributes to determine the business urgency index; and network congestion factors related to the corresponding initial task unit are extracted from network status data. A positive evaluation metric for dynamic scheduling priority weights is constructed based on the complexity index, resource requirement index, and business urgency index. A suppressive evaluation metric for dynamic scheduling priority weights is constructed based on the remaining deadline constraint and the network congestion factor. This ensures that initial task units with higher business urgency and shorter remaining deadlines have a higher scheduling priority tendency, and that network congestion status exerts a suppressive correction on the dynamic scheduling priority weights. The complexity index, resource requirement index, and business urgency index are each configured with a weighting coefficient that dynamically adjusts with the system's operating status, and each weighting coefficient is subject to normalization constraints. Based on the network status data, an adaptive adjustment mechanism for weighting parameters is constructed. When the network status data indicates that the network congestion level is aggravated, the weighting parameter adaptive adjustment mechanism increases the value of the weighting coefficient corresponding to the business urgency index. The weighting coefficients corresponding to the complexity index and / or the resource demand index are normalized and adjusted to improve the relative influence of the business urgency index in the dynamic scheduling priority weight calculation.

4. The method according to claim 1, characterized in that, The process involves generating a node capability vector corresponding to each processing node based on the node data of each processing node, and evaluating the matching degree between the orthogonal feature vector of each initial task unit and the node capability vector of each processing node to generate a compatibility matrix for quantifying the degree of adaptation between the initial task unit and the processing node, including: Based on the computing power, concurrency, hardware acceleration features, and available resource status contained in the node data of the processing nodes, a node capability vector corresponding to each processing node is generated, and the node capability vector is dimensionally aligned so that the node capability vector can be evaluated for matching degree with orthogonal feature vectors. Traverse all initial task units and all processing nodes, calculate the corresponding matrix elements, and construct a compatibility matrix based on all the matrix elements; wherein, when the node attribute data contained in the node data indicates that the corresponding processing node does not support the data flow type of the corresponding initial task unit in the hardware architecture, the matrix element in the compatibility matrix corresponding to the initial task unit and the processing node is set to zero.

5. The method according to claim 4, characterized in that, The step of calculating the orthogonal conflict degree based on the corresponding orthogonal feature vectors for at least two initial task units in the candidate scheduling scheme that are intended to be assigned to the same processing node includes: For any two initial task units in the candidate scheduling scheme that are to be assigned to the same processing node, the orthogonal conflict degree is determined based on the normalized directional similarity between their corresponding orthogonal feature vectors; wherein, the normalized directional similarity is determined based on the absolute correlation of the inner product between the two orthogonal feature vectors, and the influence of vector magnitude difference on the orthogonal conflict degree is weakened by normalization processing. Accordingly, when the orthogonal conflict degree satisfies the preset conflict condition, the step of applying a scheduling penalty to the candidate allocation relationship in the candidate scheduling scheme used to allocate at least two initial task units to the same processing node, and filtering or adjusting the candidate allocation relationship based on the scheduling penalty to obtain the target scheduling scheme includes: The orthogonal conflict degree is compared with a preset conflict threshold. When the orthogonal conflict degree reaches or exceeds the preset conflict threshold, the candidate allocation relationship of the two initial task units to be assigned to the same processing node is determined as the candidate allocation relationship to be penalized. A global scheduling cost function is constructed, which includes a basic scheduling cost term and an orthogonal conflict penalty term. The basic scheduling cost term is negatively correlated with the fit between the initial task unit and the processing node, and is related to the expected execution delay and remaining deadline constraints of the initial task unit on the corresponding processing node. The orthogonal conflict penalty term is used to improve the global cost evaluation value of the candidate scheduling scheme that includes the candidate allocation relationship to be penalized. The global scheduling cost function is optimized under the condition of satisfying the allocation constraints, and the candidate allocation relationship to be penalized is screened or adjusted based on the orthogonal conflict penalty term to obtain the target scheduling scheme, and a scheduling matrix is ​​generated based on the target scheduling scheme; wherein, the allocation constraints include: each initial task unit is assigned to a candidate processing node; when the compatibility matrix indicates that there is no alternative candidate allocation relationship between the initial task unit and the processing node, the corresponding initial task unit is prohibited from being assigned to the processing node; and the resource occupancy corresponding to each processing node does not exceed the available resource capacity of the processing node.

6. The method according to claim 5, characterized in that, The step of filtering or adjusting the candidate allocation relationships to be penalized based on the orthogonal conflict penalty term to obtain the target scheduling scheme, and generating a scheduling matrix based on the target scheduling scheme, includes: Initialize the scheduling allocation state, the set of allocated initial task units for each processing node, and the remaining available resource capacity, and build a task scheduling queue in descending order of the dynamic scheduling priority weight of each initial task unit. According to the sorting order of the task scheduling queue, each initial task unit is traversed. For the current initial task unit, effective processing nodes are extracted from the optional candidate allocation relationship indicated by the compatibility matrix, and a set of candidate nodes is selected from the effective processing nodes based on the basic cost increment corresponding to the basic scheduling cost item. For each candidate processing node in the candidate node set, the scheduling penalty increment is determined based on the orthogonal conflict degree between the initial task unit set already allocated on the candidate processing node and the current traversed initial task unit, and the total marginal cost of the corresponding candidate processing node is determined in combination with the basic cost increment. The candidate processing node with the lowest total marginal cost and satisfying the available resource capacity constraint is selected from the candidate node set as the target processing node; if there is no candidate processing node in the candidate node set that satisfies the available resource capacity constraint and has the lowest total marginal cost, then the remaining valid processing nodes in the valid processing nodes other than the candidate node set are selected as the target processing node. Assign the currently traversed initial task unit to the target processing node, and update the corresponding scheduling allocation status, the set of allocated initial task units of the target processing node, and the remaining available resource capacity. The process continues until all initial task units in the task scheduling queue have been allocated and traversed. Based on all updated scheduling allocation states, the target scheduling scheme is obtained, and a scheduling matrix is ​​generated and output according to the target scheduling scheme.

7. The method according to claim 5 or 6, characterized in that, After generating and outputting the scheduling matrix, the method further includes a closed-loop state feedback operation; wherein the closed-loop state feedback operation includes: Periodically collect task completion long-tail latency, task retransmission rate, and overall resource utilization in a heterogeneous computing environment; wherein, the task completion long-tail latency is a preset quantile latency of the task completion latency; Within the current feedback control cycle, obtain the current long-tail delay value, the current retransmission rate value, and the current resource utilization value; Based on the current long-tail delay value, the current retransmission rate value, and the current resource utilization value, the orthogonal penalty coefficient in the global scheduling cost function is updated online. Specifically, when the current retransmission rate value is greater than the retransmission rate threshold, or the current long-tail delay value is greater than the long-tail delay upper limit threshold, the orthogonal penalty coefficient is increased to enhance the penalty intensity for orthogonal collision degree. If the current resource utilization value is less than the lower limit threshold of resource utilization, the current long-tail latency value is not greater than the upper limit threshold of long-tail latency, and the current retransmission rate value is not greater than the retransmission rate threshold, the orthogonal penalty coefficient is reduced to reduce the degree of resource idleness caused by excessive isolation. The updated orthogonal penalty coefficient is limited to between a preset lower limit and a preset upper limit, and the updated orthogonal penalty coefficient is used to optimize the solution of the global scheduling cost function in the next feedback control cycle.

8. A heterogeneous node digital orthogonal scheduling system adapted to the characteristics of mixed data streams, characterized in that, The system includes: The data acquisition and classification module is used to acquire the mixed data stream to be processed, the network status data of the current system, and the node data of each processing node in the heterogeneous computing environment, and classify the mixed data stream according to business attributes to generate multiple initial task units; wherein, the node data includes node attribute data and node running status data; The orthogonal feature mapping module is used to extract multidimensional original features for each of the initial task units, generate an orthogonal basis for constructing an orthogonal feature space based on the multidimensional original features, and perform orthogonal mapping on the multidimensional original features based on the orthogonal basis to generate orthogonal feature vectors corresponding to each of the initial task units. The priority weight calculation module is used to calculate the dynamic scheduling priority weight of each initial task unit based on the service attributes, multidimensional original features and network status data of each initial task unit. A compatibility matrix generation module is used to generate node capability vectors corresponding to each processing node based on the node data of each processing node, and to evaluate the matching degree between the orthogonal feature vectors of each initial task unit and the node capability vectors of each processing node, thereby generating a compatibility matrix for quantifying the degree of adaptation between the initial task unit and the processing node. Specifically, for any initial task unit and any processing node, a non-negative matching degree is determined based on the matching relationship between the orthogonal feature vector of the initial task unit and the node capability vector of the processing node. The matrix elements in the compatibility matrix are calculated based on the non-negative matching degree and a dynamic adaptation coefficient. The dynamic adaptation coefficient characterizes the attenuation effect of the real-time load rate of the processing node and the allocation or migration cost of the initial task unit on the degree of adaptation, such that the higher the real-time load rate of the processing node and the greater the allocation or migration cost, the lower the corresponding degree of adaptation. The candidate scheduling generation module is used to combine the dynamic scheduling priority weight and the compatibility matrix to generate a candidate scheduling scheme that includes the candidate allocation relationship between the initial task units and the candidate processing nodes, and to calculate the orthogonal conflict degree based on the corresponding orthogonal feature vector for at least two initial task units in the candidate scheduling scheme that are to be allocated to the same processing node. The conflict penalty adjustment module is used to apply a scheduling penalty to the candidate allocation relationship in the candidate scheduling scheme that is intended to allocate at least two initial task units to the same processing node when the orthogonal conflict degree meets the preset conflict conditions, and to filter or adjust the candidate allocation relationship based on the scheduling penalty to obtain the target scheduling scheme, and to determine the target processing node of each initial task unit according to the target scheduling scheme to generate and output the scheduling matrix.

9. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor, characterized in that the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the method as described in any one of claims 1-7.

10. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

Images