A deterministic scheduling method and apparatus for latency-sensitive computing tasks

Abstract

This invention provides a deterministic scheduling method and apparatus for latency-sensitive computing tasks. The method includes the following steps: acquiring all tasks to be processed; calculating the priority weight of each task based on its task attributes; sorting all tasks to be processed based on the priority weight to obtain a task queue; traversing servers in the current network; pruning servers based on the shortest path between the current task and each server to obtain a candidate server set; constructing a transmission scheme set including multiple transmission schemes based on the servers in the candidate server set; for each transmission scheme, calculating multiple calculation factors based on the target server, the task transmission link, and the task start transmission time; calculating a comprehensive score for each transmission scheme based on the multiple calculation factors; selecting the target transmission scheme based on the comprehensive score; and implementing the transmission scheme in the network.

Description

Technical Field

[0001] This invention relates to the fields of time-sensitive networking and edge networking, and in particular to a deterministic scheduling method and apparatus for latency-sensitive computing tasks. Background Technology

[0002] With the deep integration of new-generation information technology and manufacturing, the Industrial Internet is accelerating its evolution towards collaboration and intelligence, becoming a strategic infrastructure supporting the development of new industrialization. In this process, industrial production scenarios are placing unprecedentedly stringent demands on network performance, especially in key applications such as robotic collaborative assembly, closed-loop motion control, real-time industrial vision inspection, and dynamic scheduling of flexible production lines. Numerous edge devices continuously generate massive data streams, relying on low-latency, highly reliable real-time computing for precise control and status monitoring. These edge-device-initiated computing tasks typically have strict time constraints, not only pursuing ultra-low latency at the microsecond level, but also, more importantly, ensuring the predictability and determinism of task completion time. This means that the end-to-end latency from task generation to result return must be strictly limited within a predetermined window. Such tasks are called "time-sensitive computing tasks," and their lifecycle is tightly coupled with the communication phase of data transmission from the edge device through the industrial network to the computing node, the computing phase of node execution, and the phase of transmitting the computing results back to the edge device. These three phases together constitute the end-to-end latency.

[0003] Ensuring determinism for time-sensitive computing tasks in industrial field networks presents challenges. Firstly, industrial network environments are complex, often exhibiting a heterogeneous convergence of wired and wireless technologies. In the wired domain, Time-Sensitive Networking (TSN), a key technology for ensuring deterministic communication in industrial networks, provides bounded latency, minimal jitter, and zero congestion loss guarantees for time-critical traffic through mechanisms such as precise time synchronization, traffic shaping, and path control. Among these, Cyclic Queuing and Forwarding (CQF), defined by the IEEE 802.1Qch protocol, is a crucial mechanism for achieving deterministic transmission in TSN. This mechanism, by configuring alternately opening and closing buffer queues at each pair of output ports, limits the forwarding latency of data packets at each switching node to a known time slot period, thereby providing stable transmission guarantees for periodic traffic. However, existing TSN scheduling research primarily focuses on network-level transmission determinism, emphasizing the effective planning of time slot resources within a single network domain. On the other hand, in the wireless domain, both 5G standards and new protocols such as Wi-Fi 6 / 7 have introduced Orthogonal Frequency Division Multiple Access (OFDMA) technology. Unlike traditional OFDMA, which can only serve a single user per time slot, OFDMA divides the channel into multiple subcarrier groups in the frequency domain and introduces finer-grained time slot division in the time domain, thus forming a two-dimensional "time-frequency resource block" grid in both the time and frequency domains. By quantifying resources in both the time and frequency domains, OFDMA significantly improves spectrum utilization and scheduling flexibility.

[0004] Currently, research has begun to focus on the collaborative scheduling problem of communication and computation for latency-sensitive computing tasks. The core idea of ​​the most similar scheme to this invention is to model the deterministic scheduling problem as a resource time slot allocation problem, considering both computation-level and network-level scheduling. Its system model abstracts the industrial IoT network as a directed graph containing terminal devices, routers, and servers. Each periodic computing task is defined by attributes such as its source device, generation time, period, deadline, computation duration, and uplink / downlink packet size. This scheme assumes that routers in the network do not have packet buffering capabilities and packets must be forwarded immediately upon arrival. The scheme first traverses all candidate servers for each task to be scheduled, estimating the earliest possible arrival time of data transmission based on the shortest path, and then derives the feasible range of computation start times by combining the task deadline. Using the computational resource conflict conditions proposed in this scheme, infeasible intervals that may overlap with the time slots of already scheduled tasks on the same server are quickly eliminated, thus obtaining a set of candidate computation start times. After obtaining the time slot allocation of computing resources, the computation scheme processes the transmission of uplink data (device to server) and downlink results (server to device) separately. For each transmission direction, a depth-first search is first used to enumerate all candidate routing paths that satisfy the constraints. For each candidate path, the occupied transmission time slots of each link on that path are calculated based on network resource conflict conditions, and the infeasible interval of the starting transmission time from the source device is derived in reverse, thus obtaining the set of feasible starting times. After both uplink and downlink are successfully allocated, a complete scheduling scheme is formed. In addition, considering factors such as maintenance costs, the existing scheme also aims to minimize the number of active servers when solving the scheduling problem.

[0005] However, existing technical solutions mainly rely on sorting and enumerating tasks according to computational cost and response latency when generating computation-level scheduling solutions, lacking a global priority sorting mechanism based on multi-dimensional task attributes before scheduling begins. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a deterministic scheduling method for latency-sensitive computing tasks to eliminate or improve one or more defects existing in the prior art.

[0007] One aspect of the present invention provides a deterministic scheduling method for latency-sensitive computing tasks, the method being applied to an industrial internet edge network, the method comprising the following steps: Obtain all pending tasks in the current network, calculate the priority weight of each task based on the task attributes of the pending tasks, and sort all pending tasks based on the priority weight to obtain a task queue. Get the task at the head of the current task queue, traverse the servers in the current network, perform server pruning based on the shortest path between the current task and each server, and obtain a set of candidate servers. Based on the servers in the candidate server set, a transmission scheme set including multiple transmission schemes is constructed. The transmission scheme includes a target server, a task transmission link, and a task start transmission time. For each of the transmission schemes, multiple calculation factors are calculated based on the target server, the task transmission link, and the task start transmission time. The calculation factors include a fragmentation factor calculated based on the transmission time slot of the task to be processed for the task transmission link. A comprehensive score for each transmission scheme is calculated based on multiple calculation factors. The target transmission scheme is selected based on the comprehensive score and then implemented in the network.

[0008] The above scheme first calculates the urgency of tasks based on their attributes to construct a task queue. In subsequent steps, candidate servers are pruned to reduce the computational load in the subsequent scheme selection steps. In the final step, multiple calculation factors are calculated for each transmission scheme, and a comprehensive score for each scheme is calculated to select the final transmission scheme before scheduling. Before scheduling begins, the scheme uses a global priority sorting mechanism based on the multi-dimensional attributes of the tasks to ensure processing efficiency.

[0009] In some embodiments of the present invention, the task attributes include task deadline, task generation time, task generation cycle, and estimated task computation time. In the step of calculating the priority weight of each task based on the task attributes of the task to be processed, a floating time is calculated based on the task deadline, task generation time, and estimated task computation time of each task to be processed, and a priority weight is calculated based on the task deadline, task generation cycle, and floating time of each task to be processed.

[0010] In some embodiments of the present invention, in the step of calculating the floating time based on the task deadline, task generation time, and estimated task computation time for each task to be processed, the floating time is calculated using the following formula: in, Indicates task The floating time, Indicates task Task deadline Indicates task Task generation time, Indicates task The estimated time consumed by the task is calculated.

[0011] In some embodiments of the present invention, in the step of calculating the priority weight based on the task deadline, task generation cycle and floating time of each task to be processed, normalization calculation is performed based on the task deadline, task generation cycle and floating time of all tasks to be processed to obtain the normalized parameters of the corresponding task deadline, task generation cycle and floating time. The priority weight of each normalized parameter is obtained by weighted calculation.

[0012] In some embodiments of the present invention, in the step of performing normalization calculations based on the task deadlines, task generation cycles, and float times of all pending tasks to obtain the normalized parameters for the corresponding task deadlines, task generation cycles, and float times, the following formula is used for the normalization calculation: in, This represents the normalized calculation result, which is the normalized calculation result of the task deadline, task generation cycle, or floating time. This represents the original value, which is the task deadline, task generation cycle, or floating time. This represents the minimum original value of all tasks yet to be processed. This represents the maximum value of the original values ​​of all pending tasks.

[0013] In the specific implementation process, when calculating different normalized calculation results, the corresponding original values ​​are substituted in.

[0014] In some embodiments of the present invention, the step of pruning servers based on the shortest path between the current task to be processed and each server to obtain a candidate server set includes: The task upload time and task return time are determined based on the shortest path, the corresponding estimated server computation time is determined based on the target server, and the first pruning process is performed based on the estimated server computation time, task upload time and task return time. The first-hop link of the task backhaul link is determined based on the shortest path. The time slot occupancy rate of the first-hop link is calculated. The average and standard deviation of the time slot occupancy rates of all servers are calculated. Based on the average and standard deviation, it is determined whether the current network is in a light-load state. If it is in a light-load state, the servers with time slot occupancy rates greater than the preset occupancy rate threshold are subjected to a second pruning process.

[0015] In some embodiments of the present invention, in the step of determining the first-hop link of the task backhaul link based on the shortest path and calculating the time slot occupancy rate of the first-hop link, the time slot occupancy rate is the proportion of the number of time slots already allocated in the first-hop link of the task backhaul link to the total number of time slots.

[0016] In some embodiments of the present invention, in the step of determining whether the current network is in a light load state based on the average value and standard deviation, if the average value is greater than a preset first threshold and the standard deviation is less than a preset second threshold, then the current network is determined to be in a high load state; otherwise, it is in a light load state.

[0017] In some embodiments of the present invention, in the step of calculating multiple calculation factors based on the target server, the task transmission link, and the task start transmission time for each of the transmission schemes, for each sub-link in the transmission link, the time slot compactness of the sub-link is calculated based on the time slot occupied by the transmission scheme in each sub-link, and the occupancy of the sub-link in the preceding and following time slots of the time slot occupied by the transmission scheme in each sub-link. The fragmentation factor of the transmission scheme is calculated based on the time slot compactness of each sub-link.

[0018] In some embodiments of the present invention, in the step of calculating the time slot compactness of a sub-link based on the time slots occupied by the transmission scheme in each sub-link, and the occupancy of the preceding and succeeding time slots of the time slots occupied by the transmission scheme in each sub-link, the time slot compactness of the sub-link is calculated using the following formula: in, Indicates task Sublink in the transmission scheme The compactness of time slots; Indicates task Sublink in the transmission scheme The preceding time slot of the time slot occupied by the current time slot is whether it is occupied; if occupied, it is 1; if not occupied, it is 0. Indicates task Sublink in the transmission scheme The time slot following the one occupied is whether it is occupied; if occupied, it is 1.

[0019] In some embodiments of the present invention, the calculation factors further include a latency factor and a resource load factor. In the step of calculating multiple calculation factors based on the target server, the task transmission link, and the task start transmission time for each transmission scheme, the task upload time and task return time are determined based on the task transmission link of the transmission scheme; the estimated server computation time is determined based on the target server in the transmission scheme; the completion time is determined based on the task upload time, the task upload time, the estimated server computation time, and the task return time; the latency factor is calculated based on the completion time, the task deadline time, and the task generation time of the task attributes; and the resource load factor is based on the ratio of the time slots that have been scheduled to the total available time slots of the target server in the transmission scheme.

[0020] In some embodiments of the present invention, in the step of calculating the delay factor based on the completion time point, the task deadline time and the task generation time of the task attributes, the delay factor is calculated using the following formula: in, Indicates task The delay factor in a transmission scheme, Indicates task The completion time point, Indicates task Task deadline Indicates task Task generation time.

[0021] A second aspect of the present invention also provides a deterministic scheduling apparatus for latency-sensitive computing tasks. The apparatus includes a computer device, the computer device including a processor and a memory, the memory storing computer instructions, and the processor executing the computer instructions stored in the memory. When the computer instructions are executed by the processor, the apparatus performs the steps implemented by the method described above.

[0022] A third aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned deterministic scheduling method for latency-sensitive computing tasks.

[0023] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.

[0024] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0025] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0026] Figure 1 This is a schematic diagram illustrating one implementation of the deterministic scheduling method for latency-sensitive computing tasks in this scheme. Figure 2 This is a schematic diagram of the network system architecture of this solution; Figure 3 This is a schematic diagram illustrating how the proposed scheme sorts all tasks to be processed based on the aforementioned priority weights. Figure 4 This is a schematic diagram of the overall processing flow of this solution; Figure 5 This is a schematic diagram of the scheduling process in this scheme; Figure 6 This is a schematic diagram of the pruning treatment in this scheme. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0028] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0029] The drawback of existing technical solutions is that they are primarily based on wired time-sensitive network architectures, with data transmission mainly relying on time slot allocation through routers and switches supporting the IEEE 802.1Qbv protocol. However, in real-world industrial internet edge scenarios, network environments are often a mix of wired and wireless, with wireless edge devices accessing the network via wireless protocols such as Wi-Fi.

[0030] Furthermore, existing technical solutions primarily rely on sorting and enumerating tasks based on computational cost and response latency when generating computationally-level scheduling solutions, lacking a global priority sorting mechanism based on multi-dimensional task attributes before scheduling begins. This approach fails to comprehensively consider the combined impact of task deadline urgency, task cycle frequency, and task float time on scheduling success rate. Consequently, in multi-task concurrency scenarios, it may fail to prioritize the acquisition of resources for high-urgency or high-flexibility tasks. In contrast, the lack of normalized weight calculation based on deadline, cycle, and float time results in coarse-grained optimization of scheduling order, making it difficult to maximize the overall scheduling success rate when resource contention is intense. Meanwhile, the optimization goals of existing technical solutions are mainly focused on minimizing the number of active servers and the cost of network structure upgrades. Although this reduces deployment costs to some extent, in the process of selecting specific scheduling solutions, there is a lack of a comprehensive scoring function that can simultaneously balance end-to-end latency, server computing load, and resource fragmentation. Its heuristic strategy tends to share servers to reduce the number of servers, but it does not explicitly evaluate the degree of fragmentation and load balancing after resource allocation. This may lead to some servers being overloaded while other resources are idle, or time slot allocation being too dispersed, thereby reducing the scheduling space for subsequent tasks.

[0031] To address the shortcomings of existing technologies, this solution aims to provide a joint scheduling method for time-sensitive computing tasks, resolving issues such as the inability of existing solutions to effectively adapt to wired-wireless converged network architectures, insufficiently granular task scheduling priority division, and a singular resource optimization objective. This solution constructs a scheduling model adapted to wired-wireless converged edge networks, simultaneously satisfying the cyclic queuing and forwarding constraints of the wired TSN domain and the OFDMA time-frequency resource allocation constraints of the wireless domain, achieving global collaborative scheduling of computing and communication resources. This solution determines scheduling priorities by comprehensively considering the deadlines, periods, and float times of computing tasks, ensuring that urgent and flexible tasks receive resources first, thereby improving the overall scheduling success rate. Furthermore, this solution establishes a comprehensive scoring function that includes latency factors, computing resource load factors, and resource fragmentation factors. During the scheduling process, it dynamically balances low latency, load balancing, and resource compactness, optimizing system resource utilization efficiency while meeting the strict deadlines of all time-sensitive tasks, and enhancing the deterministic guarantee capability of industrial internet systems for critical computing services.

[0032] This invention proposes a deterministic scheduling method for latency-sensitive computing tasks, which is applied to the edge network of the Industrial Internet. The physical topology diagram of the industrial internet edge network to which this solution applies is shown below. Figure 2 As shown, it includes two parts: the wired domain and the wireless domain. Figure 2 In , and This represents servers deployed in the Industrial Internet. , , and They are all network forwarding devices, among which What makes it special is that it is a wireless access point, through which wireless devices in the network access each other. Establish a wireless connection. Figure 2 In the diagram, straight lines represent wired connections in the network, while dashed lines represent wireless connections. Wired domain network switching nodes employ a circular queuing forwarding model. This model maintains a circular queuing forwarding queue with two buffer queues on each output port of each switching node. Within a data transmission cycle, only one queue has the authority to send a stream, while the other queue can only receive streams. If a switching node receives a data packet from the link within this transmission cycle, it must output it to the link in the next transmission cycle. The wireless domain uses a wireless network supporting OFDMA technology. OFDMA technology quantizes transmission resources into time-frequency resource blocks in both the time and frequency domains, and supports multi-user concurrent communication. The set of time-sensitive computing tasks in the network is denoted as , and each task has task attributes, including task generation time. Task cycle Task deadline Calculate the estimated time consumed by the task The device that initiated the task and the size of the uplink and downlink data packets.

[0033] Specifically, in the wired-wireless converged edge network model constructed in this scheme, three core constraints are defined to ensure the transmission of latency-sensitive streams: global constraints, wired domain TSN network constraints, and wireless domain network constraints. These constraints together constitute the basic boundary conditions for joint scheduling of time-sensitive computing tasks. Firstly, at the global constraint level, data transmission must follow strict order constraints; that is, data packets must be fully received on the previous hop link before being transmitted on the next hop link, ensuring the continuity of the data flow. Simultaneously, the transmission start time is strictly limited by the task lifecycle. When the computational data to be processed is forwarded from the wireless edge device to the server, the first hop transmission naturally needs to occur after the computational task is generated. Conversely, when the computational task completes and the result is returned to the wireless domain edge device, the start time of the transmission phase must be after the computational task has finished. Furthermore, the transmission deadline constraint requires that the data upload phase must be completed before the planned start time of the computational task, and the result return phase must be completed before the task deadline, thus satisfying the upper bound of the tolerable latency of the computational task and ensuring the determinism of end-to-end latency. To address the constraints of wired TSN networks, the time slot period planning based on a cyclic queuing forwarding queue imposes specific limitations on time slot resource sharing and transmission windows. The time slot period size has upper and lower bound constraints. The upper bound is the greatest common divisor of all periods in the task set, while the lower bound ensures that any frame can be transmitted within one time slot period, taking into account the transmission delay of all data frames in the transmission queue and link-related delays. In terms of the transmission mechanism, since the basic time slot unit of the cyclic queuing forwarding mechanism is fixed, the transmission start time is set to an integer multiple of the time slot period, and the transmission duration is defined as the time difference between the time when the last transmission is completed within the current transmission period and the transmission start time. More critically, there are constraints on time slot resource allocation: each time slot period can only be allocated to one transmission task on the same link at the TSN switch's output port to avoid resource conflicts. Simultaneously, to meet transmission load requirements, the transmission duration must include sufficient time slot resources to transmit its effective load, determined jointly by link bandwidth and packet load. Regarding wireless domain network constraints, specific limitations are imposed primarily on wireless transmission and dynamic scheduling of time-frequency grids. These limitations depend on the wireless network architecture and the OFDMA resource scheduling mechanism. In OFDMA-based transmission, the wireless network allows multiple users to schedule on multiple sub-frequency channels in the frequency domain, while resource allocation in the time domain is still performed through time slots. Each stream's data frames can only be scheduled between specific time-frequency resource blocks. During time-frequency resource allocation, each time-frequency grid resource element can only be allocated to one user to ensure transmission orthogonality. Finally, wireless transmission resource constraints require that, to meet the transmission load demands of the stream, each scheduled time-frequency resource allocation must contain sufficient transmission resources to support the payload of the transmitted data packets.

[0034] like Figure 1 and Figure 4 As shown, the steps of the method include: Step S100: Obtain all pending tasks in the current network, calculate the priority weight of each task based on the task attributes of the pending tasks, and sort all pending tasks based on the priority weight to obtain a task queue. In the specific implementation process, all tasks to be processed are sorted in ascending order based on the priority weights to obtain a task queue.

[0035] Step S200: Obtain the task to be processed at the head of the current task queue, traverse the servers in the current network, perform server pruning based on the shortest path between the current task to be processed and each server, and obtain a set of candidate servers. In the specific implementation process, for all candidate servers, this scheme proposes a pruning strategy based on latency constraints and link load to eliminate servers that cannot meet the latency constraints. The specific pruning strategy based on latency constraints and link load will be described in the next section. After obtaining the set of available candidate servers, in addition to calculating the shortest route path for each server, a suboptimal shortest path is also calculated for backup. Based on this, the algorithm further constructs the corresponding communication transmission scheduling combination. The communication transmission combination consists of the start time and duration of the transmission. After constructing the corresponding transmission combination, in the wired domain, according to the cyclic queuing forwarding model, it checks whether there are idle time slots on each wired link in the path during the corresponding transmission time period, which must satisfy the TDMA constraint, that is, the same link and the same time slot can only be allocated to one task. In the wireless domain, the algorithm, according to the OFDMA resource scheduling mechanism, checks whether there are idle time-frequency resource blocks on the wireless links during the corresponding time period, which must satisfy the orthogonality constraint, that is, resource blocks of the same frequency band at the same time can only be allocated to one transmitting user. If the communication resource check passes, a complete feasible scheduling scheme is formed, including server selection, calculation time, uplink path and time slot, and downlink path and time slot.

[0036] Step S300: Based on the servers in the candidate server set, construct a transmission scheme set including multiple transmission schemes. The transmission scheme includes a target server, a task transmission link, and a task start transmission time. Step S400: For each of the transmission schemes, calculate multiple calculation factors based on the target server, the task transmission link, and the task start transmission time. The calculation factors include a fragmentation factor calculated based on the transmission time slot of the task to be processed by the task transmission link. like Figure 5As shown, in step S500, a comprehensive score for each transmission scheme is calculated based on multiple calculation factors, a target transmission scheme is selected based on the comprehensive score, and the transmission scheme is implemented in the network.

[0037] In the specific implementation process, in the step of calculating the comprehensive score of each transmission scheme based on multiple calculation factors, the comprehensive score is calculated using a weighted summation method that is the same as the calculation priority weight.

[0038] In practice, the algorithm selects the optimal scheduling result based on the comprehensive score, incorporates it into the final solution set, and immediately performs a resource status update operation. This update operation includes locking the computation time slice on the selected server, occupying the time slot on the selected wired link, and allocating the time-frequency resource block on the selected wireless link. Real-time resource updates ensure that subsequent task scheduling decisions are aware of the current resource consumption, thereby avoiding resource conflicts. If a feasible scheduling solution cannot be found on any candidate server, the task is marked as scheduling failed. After all tasks have been traversed, the algorithm outputs the final joint scheduling solution set, completing the entire scheduling process.

[0039] In practical implementation, after completing the task attribute-based scheduling and obtaining the sorted set of computational tasks, this scheme executes a joint scheduling method for time-sensitive computational tasks. The inputs are the sorted set of computational tasks, the wired-wireless converged network topology, and the server set. The output is a specific scheduling solution for each task, including the selected target server, computation start time, uplink and downlink transmission paths, and resource occupancy time slots on each link. This method, while satisfying the wired cyclic queuing and forwarding constraints and the wireless time-frequency resource orthogonality constraints, uses a comprehensive scoring function to select the optimal scheduling combination, thereby achieving optimized allocation of computational and communication resources.

[0040] The specific execution flow of the scheduling method first involves initialization, establishing an empty set of feasible scheduling solutions, and loading the current network resource occupancy state graph. This graph records the time slot occupancy of all wired links and the time-frequency resource block allocation of wireless links. Then, tasks to be processed are retrieved sequentially according to the task priority queue. For each task, the algorithm enters a server candidate evaluation loop, traversing all available edge computing servers in the network.

[0041] The above scheme first calculates the urgency of tasks based on their attributes to construct a task queue. In subsequent steps, candidate servers are pruned to reduce the computational load of subsequent scheme selection steps. In the final step, multiple calculation factors are calculated for each transmission scheme, and a comprehensive score for each transmission scheme is calculated to select the final transmission scheme before scheduling. Before scheduling begins, this scheme uses a global priority sorting mechanism based on the multi-dimensional attributes of tasks to ensure processing efficiency.

[0042] like Figure 3 As shown, in some embodiments of the present invention, the task attributes include task deadline, task generation time, task generation cycle, and estimated task computation time. In the step of calculating the priority weight of each task based on the task attributes of the task to be processed, the floating time is calculated based on the task deadline, task generation time, and estimated task computation time of each task to be processed, and the priority weight is calculated based on the task deadline, task generation cycle, and floating time of each task to be processed.

[0043] In practical implementation, within the industrial internet edge network environment, facing multiple concurrent time-sensitive computing tasks, different scheduling orders directly affect the outcome of resource contention, thus impacting the overall scheduling success rate. Using random order or single-attribute sorting might cause urgent tasks to fail to schedule due to resources being occupied by non-urgent tasks. Therefore, before executing the joint scheduling method, this solution first employs a task attribute-based scheduling and sorting method to preprocess the task set, generating a task queue with priority order. The core of this method lies in comprehensively considering three key characteristic factors: the deadline, task cycle, and float time of the computing tasks. It utilizes normalization methods to eliminate the influence of different dimensions, calculates the priority weight of each task, and ensures that urgent and inflexible tasks receive resources first.

[0044] Specifically, this scheme first defines three core attributes that affect scheduling priority. The first attribute is the deadline of the computation task. The shorter the deadline, the more urgent the task and the lower its tolerance for latency; therefore, it should be assigned a higher scheduling priority weight. The second attribute is the cycle time of the computation task. Computation tasks with shorter cycles mean that their data transmission and resource requests are more frequent, resulting in a higher frequency of system resource consumption. Therefore, short-cycle flows should be given higher scheduling priority to prevent congestion caused by frequent task accumulation. The third attribute is the float time of the computation task. This float time is the task's deadline minus its generation time and then its computation time. The float time determines the size of the time window that the task can use for transmission. The smaller the float time, the less time the task has to be scheduled, and it should be processed first.

[0045] In some embodiments of the present invention, in the step of calculating the floating time based on the task deadline, task generation time, and estimated task computation time for each task to be processed, the floating time is calculated using the following formula: in, Indicates task The floating time, Indicates task Task deadline Indicates task Task generation time, Indicates task The estimated time consumed by the task is calculated.

[0046] In some embodiments of the present invention, in the step of calculating the priority weight based on the task deadline, task generation cycle and floating time of each task to be processed, normalization calculation is performed based on the task deadline, task generation cycle and floating time of all tasks to be processed to obtain the normalized parameters of the corresponding task deadline, task generation cycle and floating time. The priority weight of each normalized parameter is obtained by weighted calculation.

[0047] In the specific implementation process, in the step of calculating the priority weight based on the task deadline, task generation cycle, and float time of each pending task, the priority weight is calculated by weighted summation using the following formula: in, Indicates task Priority weights, This represents the normalized parameter corresponding to the task deadline. This represents the normalized parameter corresponding to the task generation cycle. This represents the normalization parameter corresponding to the float time. , and These are the corresponding weighting coefficients.

[0048] In the specific implementation process, , and This setting can be configured according to specific application scenarios to balance the impact of different parameters on the priority calculation of computational tasks. After calculating the priority factor, this scheme sorts the set of computational tasks based on the scheduling priority factor. As mentioned earlier, computational tasks with shorter deadlines, shorter sending periods, and smaller float times will receive higher scheduling priority, and the scheduling priority factors for these tasks will be correspondingly smaller. Therefore, the smaller the scheduling priority factor, the higher the priority. The system will rearrange the task queue according to the ascending order of scheduling priority factors, and finally output the sorted set of computational tasks.

[0049] In some embodiments of the present invention, in the step of performing normalization calculations based on the task deadlines, task generation cycles, and float times of all pending tasks to obtain the normalized parameters for the corresponding task deadlines, task generation cycles, and float times, the following formula is used for the normalization calculation: in, This represents the normalized calculation result, which is the normalized calculation result of the task deadline, task generation cycle, or floating time. This represents the original value, which is the task deadline, task generation cycle, or floating time. This represents the minimum original value of all tasks yet to be processed. This represents the maximum value of the original values ​​of all pending tasks.

[0050] In the specific implementation process, when calculating different normalized calculation results, the corresponding original values ​​are substituted in.

[0051] By adopting the above scheme, a normalization method is used to perform linear operations on the original data, mapping the results to a unified interval. This operation does not change the relationship between the original data.

[0052] like Figure 6 As shown, in some embodiments of the present invention, the step of pruning servers based on the shortest path between the current task to be processed and each server to obtain a candidate server set includes: The task upload time and task return time are determined based on the shortest path, the corresponding estimated server computation time is determined based on the target server, and the first pruning process is performed based on the estimated server computation time, task upload time and task return time. Specifically, if the server's estimated computation time + task upload time + task return time > task deadline, then the server will be pruned.

[0053] The first-hop link of the task backhaul link is determined based on the shortest path. The time slot occupancy rate of the first-hop link is calculated. The average and standard deviation of the time slot occupancy rates of all servers are calculated. Based on the average and standard deviation, it is determined whether the current network is in a light-load state. If it is in a light-load state, the servers with time slot occupancy rates greater than the preset occupancy rate threshold are subjected to a second pruning process.

[0054] In some embodiments of the present invention, in the step of determining the first-hop link of the task backhaul link based on the shortest path and calculating the time slot occupancy rate of the first-hop link, the time slot occupancy rate is the proportion of the number of time slots already allocated in the first-hop link of the task backhaul link to the total number of time slots.

[0055] In some embodiments of the present invention, in the step of determining whether the current network is in a light load state based on the average value and standard deviation, if the average value is greater than a preset first threshold and the standard deviation is less than a preset second threshold, then the current network is determined to be in a high load state; otherwise, it is in a light load state.

[0056] In the specific implementation process, to avoid the high time complexity of traversing all servers when scheduling specific computing tasks, this solution proposes a pruning strategy based on latency constraints and link load to remove some servers that do not meet the scheduling requirements or are difficult to schedule. The pruning strategy uses the shortest path algorithm to calculate the uplink transmission path from the task source edge device to the candidate server, and the downlink transmission path from the candidate server back to the task source edge device, based on the current network topology. After determining the shortest path, the upload data time and the result return time can be calculated. Then, the total time (task generation time plus upload data time, task processing time, and data return time) is checked to see if it exceeds the task deadline. If it does, the server does not meet the scheduling requirements regardless, and the pruning strategy removes these servers; this is the first round of pruning.

[0057] Next, for each candidate server, the pruning strategy focuses on the load of a critical link. The strategy prioritizes the first wired link along the shortest path from the server back to the source wireless access point, i.e., the server's downlink first-hop link. We extract the time slot occupancy rate of this downlink first-hop link, which is the proportion of allocated time slots on that link to the total number of time slots. This occupancy rate directly reflects the server's "egress congestion level." A higher occupancy rate indicates a more congested backhaul channel for the server's computational results. After collecting the downlink first-hop link load of all remaining servers after the first round of pruning, we calculate the average and standard deviation of the downlink first-hop link occupancy rates for all candidate servers. The average reflects the overall congestion level of the current network downlink channel, and the standard deviation reflects the degree of load difference between different servers. Based on the average and standard deviation, we determine the overall load level. When the overall network load is relatively light and there are many resources available for scheduling, a group of servers with high first-hop link loads are removed, thereby alleviating the network load pressure on these servers. This is the second round of pruning. When the overall network load is heavy, and considering the limited resources available for scheduling in the network, a second round of pruning will not be performed in order to maximize the scheduling success rate.

[0058] In some embodiments of the present invention, in the step of calculating multiple calculation factors based on the target server, the task transmission link, and the task start transmission time for each of the transmission schemes, for each sub-link in the transmission link, the time slot compactness of the sub-link is calculated based on the time slot occupied by the transmission scheme in each sub-link, and the occupancy of the sub-link in the preceding and following time slots of the time slot occupied by the transmission scheme in each sub-link. The fragmentation factor of the transmission scheme is calculated based on the time slot compactness of each sub-link.

[0059] In the specific implementation process, in the step of calculating the fragmentation factor of the transmission scheme based on the time slot compactness of each sub-link, the average value of the time slot compactness of the sub-link is calculated and then divided by 2 to obtain the fragmentation factor.

[0060] In some embodiments of the present invention, in the step of calculating the time slot compactness of a sub-link based on the time slots occupied by the transmission scheme in each sub-link, and the occupancy of the preceding and succeeding time slots of the time slots occupied by the transmission scheme in each sub-link, the time slot compactness of the sub-link is calculated using the following formula: in, Indicates task Sublink in the transmission scheme The compactness of time slots; Indicates task Sublink in the transmission scheme The preceding time slot of the time slot occupied by the current time slot is whether it is occupied; if occupied, it is 1; if not occupied, it is 0. Indicates task Sublink in the transmission scheme The time slot following the one occupied is whether it is occupied; if occupied, it is 1.

[0061] In some embodiments of the present invention, the calculation factors further include a latency factor and a resource load factor. In the step of calculating multiple calculation factors based on the target server, the task transmission link, and the task start transmission time for each transmission scheme, the task upload time and task return time are determined based on the task transmission link of the transmission scheme; the estimated server computation time is determined based on the target server in the transmission scheme; the completion time is determined based on the task upload time, the task upload time, the estimated server computation time, and the task return time; the latency factor is calculated based on the completion time, the task deadline time, and the task generation time of the task attributes; and the resource load factor is based on the ratio of the time slots that have been scheduled to the total available time slots of the target server in the transmission scheme.

[0062] Using the above scheme, the latency factor reflects the speed of task completion; the smaller the value, the earlier the task is completed, which is more conducive to meeting the deadline. The computational resource load factor is calculated by the ratio of the server's scheduled time slots to all available time slots. The computational resource load factor measures the server's load, and servers with lighter loads are more likely to be selected. The resource fragmentation factor is designed to measure the compactness between newly allocated time slots and occupied time slots. In the allocation of network transmission time slots, it is desirable for newly allocated time slots to be as close as possible to occupied time slots to reduce resource fragmentation and reserve continuous resource blocks for subsequent tasks.

[0063] In some embodiments of the present invention, in the step of calculating the delay factor based on the completion time point, the task deadline time and the task generation time of the task attributes, the delay factor is calculated using the following formula: in, Indicates task The delay factor in a transmission scheme, Indicates task The completion time point, Indicates task Task deadline Indicates task Task generation time.

[0064] In summary, this scheme integrates the circular queuing forwarding mechanism of wired time-sensitive networks and the time-frequency resource allocation mechanism of orthogonal frequency division multiple access (OFDMA) in wireless networks into a unified scheduling model, achieving deterministic guarantees for the entire lifecycle of tasks (data upload, computation processing, and result feedback). During the specific scheduling phase, this scheme constructs a priority ranking mechanism based on the multi-dimensional attributes of tasks (deadline, period, and float), using a normalization method to quantify the urgency of tasks, ensuring that high-urgency tasks acquire resources first, thus reducing the probability of tasks missing deadlines from the outset. During the scheduling execution phase, this scheme calculates the shortest transmission path for each candidate server and constructs strict time window constraints to ensure precise matching between the computation start time and data arrival time, guaranteeing timely feedback of computation results. Finally, a comprehensive scoring function including latency factors, computational resource load factors, and resource fragmentation factors is introduced to select the optimal solution from multiple feasible scheduling combinations, thereby balancing low latency, load balancing, and resource allocation effects while meeting strict deadlines.

[0065] This invention is adaptable to heterogeneous network architectures that integrate wired and wireless technologies, while existing solutions are mainly based on wired time-sensitive network architectures, neglecting the widespread wireless access needs in industrial edge scenarios. This invention constructs a unified mathematical model that includes a wired cyclic queuing forwarding model and wireless time-frequency resource allocation, incorporating the transmission opportunity constraints of wireless links, channel conditions, and the time slot period constraints of wired links into the same scheduling framework.

[0066] Secondly, this invention proposes a normalized scheduling and sorting mechanism based on multi-dimensional task attributes, which effectively improves the scheduling success rate of high-urgency tasks and overcomes the shortcomings of coarse-grained task scheduling order optimization in existing technologies. This allows tasks with short deadlines, frequent cycles, and small fluctuation times to obtain higher scheduling priorities. This sorting mechanism, based on the inherent logical relationship of task attributes, ensures that in situations of intense resource competition, the system can prioritize the most urgent and least flexible tasks for resource allocation, thereby enhancing the system's ability to guarantee service for critical business operations.

[0067] Furthermore, this invention proposes a two-round dynamic pruning strategy based on latency constraints and link load, significantly reducing the high time complexity of traversing servers in large-scale networks. This strategy first estimates the total time taken from uploading, computation, to backhaul based on the shortest path, preemptively eliminating candidate servers destined to time out to drastically reduce the search space. Subsequently, the algorithm precisely focuses on the core bottleneck of the "computation result backhaul channel," extracting the time slot occupancy rate of the downlink first-hop link for each server, and dynamically assessing the overall network load using the average and standard deviation of the global occupancy rate. This adaptive mechanism based on relative statistical characteristics completely abandons rigid fixed thresholds: under light network load, the system decisively eliminates congested nodes using the average and standard deviation to accelerate the solution; while under heavy load and resource scarcity, the system automatically skips the second round of pruning to retain all available candidates. This mechanism greatly improves the algorithm's solution efficiency while perfectly ensuring the global scheduling success rate under different network pressures.

[0068] Finally, this invention establishes a comprehensive scoring function that incorporates latency, load, and resource fragmentation, achieving a balance between resource utilization efficiency and system stability. This avoids the resource hotspots or fragmentation problems caused by the single optimization objective in existing technologies. Existing solutions tend to minimize the number of active servers to reduce deployment costs during the selection of scheduling solutions, which may lead to overloading of some servers or overly dispersed time slot allocation, thus reducing the scheduling space for subsequent tasks. This invention, in determining the optimal scheduling combination, considers not only end-to-end latency but also introduces a resource load factor and a resource fragmentation factor. Through weighted scoring, it selects a scheduling scheme that satisfies both low latency requirements and ensures even server load distribution and compact time slot allocation. This multi-objective collaborative optimization approach reduces resource fragmentation, reserving more continuously available resource blocks for subsequent tasks, thereby improving scheduling success rate and system robustness.

[0069] The core claim of this invention lies in a joint scheduling method for time-sensitive computing tasks in wired-wireless converged edge networks. This method constructs a system architecture that integrates a time-sensitive network cyclic queuing forwarding model and a wireless OFDMA time-frequency resource model. It proposes a priority ranking mechanism based on task multi-dimensional attribute normalization and combines it with a comprehensive scoring function that includes latency, load, and resource fragmentation. This achieves global collaborative optimization scheduling of computing and communication resources, thereby ensuring end-to-end deterministic response for time-sensitive computing tasks in heterogeneous network environments. The key points of this invention are as follows: 1. This invention extracts three feature factors from a task: deadline, task duration, and float time. It uses a normalization method to map features of different dimensions to a unified interval and obtains a scheduling priority factor through weighted calculation. This method differs from existing technologies that prioritize tasks based solely on computational cost or response delay. It ensures that urgent tasks with short deadlines, frequent durations, and small float times receive priority access to resources, thereby maximizing the overall scheduling success rate of the system when resource competition is intense.

[0070] 2. This invention employs a two-stage progressive dynamic pruning approach to rapidly eliminate invalid servers before precise scheduling. Its core lies in extracting the end-to-end latency limit of the task and the relative load of the downlink first-hop link as cascading evaluation factors: the first round of protection uses a mechanism that eliminates servers based on a hard comparison of the estimated total latency and the task deadline; the second round of focused protection is based on adaptive soft pruning logic with global situational awareness, specifically extracting the time slot occupancy rate of the downlink first-hop link as an egress congestion indicator, and determining the overall network load level by calculating the average and standard deviation of the occupancy rates of all candidate nodes. This invention protects this differentiated decision-making process: under lightly loaded networks, the average and standard deviation are used to generate a relative threshold to eliminate congested nodes, while under heavily loaded networks, pruning operations are automatically suspended to preserve scheduling opportunities, thus differentiating it from traditional exhaustive search or absolute threshold filtering, achieving a dynamic balance between computational latency and scheduling success rate.

[0071] 3. The scoring function of this invention not only includes an end-to-end latency factor, but also introduces a computational resource load factor and a resource fragmentation factor. By using a weighted scoring-guided scheduling method, a scheme that can meet low latency requirements while ensuring even distribution of server load and compact time slot allocation is selected, thus solving the problems of local resource hotspots or severe resource fragmentation that may occur in existing technologies, affecting the scheduling space of subsequent tasks.

[0072] This invention also provides a deterministic scheduling device for latency-sensitive computing tasks. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method described above.

[0073] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned deterministic scheduling method for time-sensitive computing tasks. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0074] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0075] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0076] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deterministic scheduling method for latency-sensitive computing tasks, characterized in that, The method is applied to the edge network of the Industrial Internet, and the steps of the method include: Obtain all pending tasks in the current network, calculate the priority weight of each task based on the task attributes of the pending tasks, and sort all pending tasks based on the priority weight to obtain a task queue. Get the task at the head of the current task queue, traverse the servers in the current network, perform server pruning based on the shortest path between the current task and each server, and obtain a set of candidate servers. Based on the servers in the candidate server set, a transmission scheme set including multiple transmission schemes is constructed. The transmission scheme includes a target server, a task transmission link, and a task start transmission time. For each of the transmission schemes, multiple calculation factors are calculated based on the target server, the task transmission link, and the task start transmission time. The calculation factors include a fragmentation factor calculated based on the transmission time slot of the task to be processed for the task transmission link. A comprehensive score for each transmission scheme is calculated based on multiple calculation factors. The target transmission scheme is selected based on the comprehensive score and then implemented in the network.

2. The deterministic scheduling method for latency-sensitive computing tasks according to claim 1, characterized in that, The task attributes include task deadline, task generation time, task generation cycle, and estimated task computation time. In the step of calculating the priority weight of each task based on the task attributes of the tasks to be processed, the floating time is calculated based on the task deadline, task generation time, and estimated task computation time of each task to be processed, and the priority weight is calculated based on the task deadline, task generation cycle, and floating time of each task to be processed.

3. The deterministic scheduling method for latency-sensitive computing tasks according to claim 2, characterized in that, In the step of calculating the floating time based on the task deadline, task generation time, and estimated computation time of each pending task, the floating time is calculated using the following formula: in, Indicates task The floating time, Indicates task Task deadline Indicates task Task generation time, Indicates task The estimated time consumed by the task is calculated.

4. The deterministic scheduling method for latency-sensitive computing tasks according to claim 3, characterized in that, In the step of calculating the priority weight based on the task deadline, task generation cycle, and floating time of each pending task, a normalization calculation is performed based on the task deadline, task generation cycle, and floating time of all pending tasks to obtain the normalized parameters of the corresponding task deadline, task generation cycle, and floating time. The weighted calculation is then performed on each normalized parameter to obtain the priority weight of each pending task.

5. The deterministic scheduling method for latency-sensitive computing tasks according to claim 4, characterized in that, In the step of normalizing the task deadlines, task generation cycles, and float times based on all pending tasks to obtain the normalized parameters for the corresponding task deadlines, task generation cycles, and float times, the following formula is used for the normalization calculation: in, This represents the normalized calculation result, which is the normalized calculation result of the task deadline, task generation cycle, or floating time. This represents the original value, which is the task deadline, task generation cycle, or floating time. This represents the minimum original value of all tasks yet to be processed. This represents the maximum value of the original values ​​of all pending tasks.

6. The deterministic scheduling method for latency-sensitive computing tasks according to claim 1, characterized in that, The steps for pruning servers based on the shortest path between the current task and each server to obtain a candidate server set include: The task upload time and task return time are determined based on the shortest path, the corresponding estimated server computation time is determined based on the target server, and the first pruning process is performed based on the estimated server computation time, task upload time and task return time. The first-hop link of the task backhaul link is determined based on the shortest path. The time slot occupancy rate of the first-hop link is calculated. The average and standard deviation of the time slot occupancy rates of all servers are calculated. Based on the average and standard deviation, it is determined whether the current network is in a light-load state. If it is in a light-load state, the servers with time slot occupancy rates greater than the preset occupancy rate threshold are subjected to a second pruning process.

7. The deterministic scheduling method for latency-sensitive computing tasks according to claim 6, characterized in that, In the step of determining the first-hop link of the task backhaul link based on the shortest path and calculating the time slot occupancy rate of the first-hop link, the time slot occupancy rate is the proportion of the number of time slots allocated in the first-hop link of the task backhaul link to the total number of time slots; in the step of determining whether the current network is in a light-load state based on the average value and standard deviation, if the average value is greater than a preset first threshold and the standard deviation is less than a preset second threshold, then the current network is determined to be in a high-load state; otherwise, it is in a light-load state.

8. The deterministic scheduling method for latency-sensitive computing tasks according to claim 1, characterized in that, In the step of calculating multiple calculation factors based on the target server, task transmission link, and task start transmission time for each of the transmission schemes, for each sub-link in the transmission link, the time slot compactness of the sub-link is calculated based on the time slot occupied by the transmission scheme in each sub-link, and the occupancy of the sub-link in the preceding and following time slots of the time slot occupied by the transmission scheme in each sub-link. The fragmentation factor of the transmission scheme is calculated based on the time slot compactness of each sub-link.

9. The deterministic scheduling method for latency-sensitive computing tasks according to claim 8, characterized in that, In the step of calculating the time slot compactness of a sub-link based on the time slots occupied by the transmission scheme in each sub-link, and the occupancy of the preceding and succeeding time slots of the time slots occupied by the transmission scheme in each sub-link, the time slot compactness of the sub-link is calculated using the following formula: in, Indicates task Sublink in the transmission scheme The compactness of time slots; Indicates task Sublink in the transmission scheme The preceding time slot of the time slot occupied by the current time slot is whether it is occupied; if occupied, it is 1; if not occupied, it is 0. Indicates task Sublink in the transmission scheme The time slot following the one occupied is whether it is occupied; if occupied, it is 1.

10. A deterministic scheduling device for latency-sensitive computing tasks, characterized in that, The device includes a computer device, which includes a processor and a memory, wherein computer instructions are stored in the memory, and the processor is configured to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method as described in any one of claims 1 to 9.

Images