A multi-core chip task scheduling optimization method and system

Abstract

The application discloses a kind of multi-core chip task scheduling optimization method and system, it is related to computer science technical field, including, the attribute information of each task in the set of tasks to be scheduled is obtained, and task model is constructed based on the attribute information;Based on the task model and the preset risk adjustment parameter, the resource requirement distribution range of each task in the preset time window is calculated, and the corresponding dynamic resource requirement boundary value is determined;According to the dynamic resource requirement boundary value and task attribute information, the set of tasks to be scheduled is classified, the task meeting the preset binding condition is divided into fixed task set, and the remaining task is divided into migration task set;For the migration task set, the association graph between task node and heterogeneous core node is constructed, and the current computing semantic environment of core is used to correct resource occupation evaluation, realize the resource saving and interference suppression brought by "computing content reuse".

Description

Technical Field

[0001] This invention relates to the field of computer science and technology, and in particular to a multi-core chip task scheduling optimization method and system. Background Technology

[0002] With the continuous evolution of semiconductor technology and the diminishing marginal returns of Moore's Law, the performance improvement of single-core processors has reached its physical limits, and multi-core architecture has become the mainstream design paradigm for modern computing chips. In multi-core chips, task scheduling, as the core hub connecting software applications and hardware resources, aims to achieve an optimal balance between overall system performance, energy efficiency, and real-time performance by rationally allocating tasks to processor cores. Early task scheduling strategies were mainly based on simple polling or static priority allocation, while modern operating systems have gradually evolved into complex mechanisms that support load balancing, priority inheritance, and energy efficiency optimization.

[0003] Traditional scheduling methods often rely on static resource demand assumptions and lack dynamic modeling of the fluctuation characteristics of resource demand within a time window. This leads to a disconnect between resource allocation and actual demand, which can easily cause core overload or resource idleness. Existing technologies often only consider resource capacity matching in the process of task classification and core mapping, while ignoring the correlation of computational content between tasks and the current computational context of the core, resulting in frequent cache invalidation and data migration overhead. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a multi-core chip task scheduling optimization method to address the problem of dynamic modeling of the resource demand fluctuation characteristics of tasks within a time window, which leads to a disconnect between resource allocation and actual demand, easily causing core overload or resource idleness. Existing technologies often only consider resource capacity matching in the process of task classification and core mapping, while ignoring the correlation of computational content between tasks and the current computational context of the core, resulting in frequent cache failures and data migration overhead problems.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a multi-core chip task scheduling optimization method, which includes, Obtain the attribute information of each task in the set of tasks to be scheduled, and construct a task model based on the attribute information; Based on the task model and preset risk adjustment parameters, calculate the resource demand distribution range of each task within a preset time window, and determine the corresponding dynamic resource demand boundary value. Based on the dynamic resource demand boundary value and task attribute information, the task set to be scheduled is classified, and tasks that meet the preset binding conditions are divided into fixed task sets, and the remaining tasks are divided into migration task sets. For the migration task set, an association graph between task nodes and heterogeneous core nodes is constructed, and an initial heterogeneous mapping scheme is generated based on the edge weights using the maximum weight matching algorithm. Based on the initial heterogeneous mapping scheme, the migration task set is dynamically scheduled and corresponding execution resources are allocated.

[0007] As a preferred embodiment of the multi-core chip task scheduling optimization method of the present invention, the task model construction includes: Standardize and semantically encapsulate the multi-dimensional heterogeneous attribute information of each task in the task set to be scheduled; A unified computing object is constructed based on the mapped attribute information. The unified computing object is used to represent the vector features of the task in the scheduling space. Based on the vector features, a task model is formed for subsequent scheduling algorithm parsing and computation.

[0008] As a preferred embodiment of the multi-core chip task scheduling optimization method of the present invention, the step of determining the dynamic resource requirement boundary value includes: Model the resource usage data during task execution, construct a task resource demand model based on historical task execution information and changes in resource usage, extract resource demand fluctuation characteristics and predict peak resource demand to form a task resource demand feature vector; Based on behavioral feature data and runtime context information during task execution, a user intent recognition model is constructed to classify the intent of the task and obtain user intent labels. Based on the user intent tags, a task service tolerance quantification model is constructed, mapping the task's response latency requirements, execution stability requirements, and resource consumption sensitivity to tolerance indicators. The resource demand feature vector is coupled with the tolerance index to construct a resource demand adjustment function, and the dynamic distribution range and corresponding upper and lower bounds of task resource demand are determined within a preset time window to form dynamic resource demand boundary values. Based on the user intent tags, tolerance indicators, and dynamic resource demand boundary values, the interference tolerance level of the task on the resource competition between cores is determined. Based on the interference tolerance level, task mapping constraint parameters are generated.

[0009] As a preferred embodiment of the multi-core chip task scheduling optimization method of the present invention, the classification of the set of tasks to be scheduled includes: Dynamically generate task binding judgment conditions, perform traversal matching judgment on the task set based on the judgment conditions, and output task classification results; Based on the classification results, a set partitioning operation is performed, dividing tasks that meet the binding conditions into a fixed task set, and dividing the remaining tasks into a migration task set.

[0010] As a preferred embodiment of the multi-core chip task scheduling optimization method of the present invention, the construction of the association graph includes: The migration task set and heterogeneous core resources are modeled in a node-based manner to obtain the task node set and the core node set, respectively. Based on the coupling relationship between task execution requirements and core capability characteristics, a connection topology between nodes is established and corresponding adaptation weights are generated. Based on the node set, connection topology, and adaptation weights, a weighted association graph structure representing the relationship between the task and the core matching is constructed.

[0011] As a preferred embodiment of the multi-core chip task scheduling optimization method of the present invention, the method for generating the initial heterogeneous mapping scheme includes: Obtain the operating status data of heterogeneous cores, including core load rate, resource utilization rate and execution status, and construct a core status vector; Extract the computational features of the tasks currently being executed by the heterogeneous cores, and construct node computational content feature vectors to characterize the current computational semantic environment of the cores; Based on the core's shared resource access behavior, a resource competition model between cores is established, and an interference matrix between cores is constructed to characterize the intensity of resource competition and neighborhood interference relationships between different cores. Based on the task resource requirement feature vector and the core state vector, a task-core matching degree function is constructed to perform a preliminary calculation on the adaptation relationship between the task and the core. Calculate the content correlation index between the task to be scheduled and the feature vector of the node's computational content. The content correlation index is used to characterize the computational reuse potential and cache locality advantage between tasks. By jointly analyzing the interference matrix, task scheduling feature vector, and content relevance index, the interference risk value of the task on different heterogeneous cores can be determined. Based on the task mapping constraint parameters, candidate core nodes are screened, and core nodes with interference risk values ​​higher than the interference tolerance level corresponding to the task are eliminated. The matching degree function is weighted and corrected based on the content relevance index to increase the matching weight of tasks that are highly relevant to the current core computing content, and an interference suppression factor is introduced to apply differentiated interference constraints to different tasks. Based on the modified matching degree function, the task node and the core node are optimized and matched to determine the mapping relationship between the task and the core. An initial heterogeneous mapping scheme is generated based on the mapping relationship.

[0012] As a preferred embodiment of the multi-core chip task scheduling optimization method of the present invention, the dynamic scheduling includes: Task scheduling is performed based on the initial heterogeneous mapping scheme, and the task execution environment and resource usage are monitored in real time to obtain a dynamic scheduling status set; Anomaly detection and deviation assessment are performed on the dynamic scheduling state set to determine the degree of deviation of the current execution state from the initial plan; Based on the degree of deviation, the task scheduling strategy is corrected online, the mapping relationship between tasks and cores is updated, and computing resources are reallocated to complete the dynamic scheduling process.

[0013] In a second aspect, the present invention provides a multi-core chip task scheduling optimization system, including a task modeling module, a boundary calculation module, a task classification module, a mapping generation module, and a dynamic scheduling module; The task modeling module is used to obtain the attribute information of each task in the task set to be scheduled, and to construct a task model based on the attribute information. The boundary calculation module is used to calculate the resource demand distribution range of each task within a preset time window based on the task model and preset risk adjustment parameters, and to determine the corresponding dynamic resource demand boundary value. The task classification module is used to classify the set of tasks to be scheduled according to the dynamic resource demand boundary value and task attribute information, and to classify the tasks that meet the preset binding conditions into fixed task sets, and the remaining tasks into migration task sets. The mapping generation module is used to construct an association graph between task nodes and heterogeneous core nodes for the migration task set, and generate an initial heterogeneous mapping scheme based on the edge weights using the maximum weight matching algorithm. The dynamic scheduling module is used to dynamically schedule the migration task set based on the initial heterogeneous mapping scheme and allocate corresponding execution resources.

[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the multi-core chip task scheduling optimization method as described in the first aspect of the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the multi-core chip task scheduling optimization method as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by classifying tasks based on dynamic boundaries, high-stability tasks are fixed and suitable optimization tasks are migrated, which effectively reduces the global matching computation scale and scheduling overhead. For the migration task set, maximum weight matching is performed by constructing an association graph that integrates content relevance and interference risk. The current computational semantic environment of the core is used to correct the resource occupation assessment, thereby realizing resource saving and interference suppression brought about by "computation content reuse". Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart for a multi-core chip task scheduling optimization method.

[0019] Figure 2 A schematic diagram of a multi-core chip task scheduling optimization system. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Reference Figures 1-2 As one embodiment of the present invention, this embodiment provides a multi-core chip task scheduling optimization method, including the following steps: S1. Obtain the attribute information of each task in the task set to be scheduled, and build a task model based on the attribute information.

[0024] Furthermore, the multi-dimensional heterogeneous attribute information of each task in the task set to be scheduled is standardized, mapped, and semantically encapsulated.

[0025] A unified computational object is constructed based on the mapped attribute information. This unified computational object is used to represent the vector characteristics of tasks in the scheduling space.

[0026] Based on vector features, a task model is formed for subsequent scheduling algorithm parsing and computation.

[0027] It should be noted that in the existing technology, different subsets of task attributes are often used in different scheduling stages. For example, the resource prediction stage uses historical resource data, while the scheduling stage uses priority or deadline information. This decentralized processing method leads to inconsistent expression of tasks in different stages, making it difficult to form a unified decision-making basis in the scheduling process.

[0028] To address the problem of fragmented representation caused by tasks using different attribute subsets at different scheduling stages in existing technologies, the first step is to standardize and map the multi-dimensional heterogeneous attribute information of each task in the task set to be scheduled, and to encapsulate it semantically. By unifying attributes from different sources and with different semantic granularities into a computable representation, tasks can have a unified structural representation when they enter the scheduling system.

[0029] By constructing a unified computational object and mapping it to vector features, tasks can have a consistent mathematical expression in the scheduling space.

[0030] This ensures that tasks are processed based on the same feature representation in subsequent resource requirement analysis, task classification, and task-core mapping processes.

[0031] By using a unified expression, the problem of information fragmentation between different stages is eliminated, thereby ensuring the consistency of the entire scheduling chain.

[0032] This avoids the scheduling deviation problem caused by inconsistencies in multi-stage processing in existing technologies.

[0033] S2. Based on the task model and preset risk adjustment parameters, calculate the distribution range of resource demand for each task within a preset time window, and determine the corresponding dynamic resource demand boundary value.

[0034] Furthermore, we model the resource usage data during task execution, construct a task resource demand model based on historical task execution information and changes in resource usage, extract resource demand fluctuation characteristics, predict peak resource demand, and form a task resource demand feature vector.

[0035] Based on behavioral feature data and runtime context information during task execution, a user intent recognition model is constructed to classify the intent of the task and obtain user intent labels.

[0036] Based on user intent tags, a task service tolerance quantification model is constructed, mapping the task's response latency requirements, execution stability requirements, and resource consumption sensitivity to tolerance indicators.

[0037] By coupling the resource demand feature vector with the tolerance index, a resource demand adjustment function is constructed, and the dynamic distribution range and corresponding upper and lower bounds of task resource demand are determined within a preset time window, forming dynamic resource demand boundary values.

[0038] Based on user intent tags, tolerance indicators, and dynamic resource demand boundary values, the interference tolerance level of tasks on inter-core resource competition is determined.

[0039] Based on the interference tolerance level, generate task mapping constraint parameters.

[0040] It should be noted that in existing technologies, resource requirements are usually output in the form of average value, peak value or fixed upper limit, and their main function is to determine whether the system has the capacity to bear the load.

[0041] The results usually remain at the level of resource prediction, only indicating "how much resources the task roughly needs", but it is difficult to directly answer "what kind of resource competition environment the task can tolerate" and "which core environments the task should exclude when mapping".

[0042] This results in a break between the resource prediction stage and the scheduling mapping stage; that is, the previous step only obtains the resource consumption estimate, while the next step still requires redefining the scheduling constraints.

[0043] Construct a task resource requirement feature vector to uniformly represent the task's requirement characteristics across multiple resource dimensions.

[0044] In existing technologies, the description of task resource requirements is usually in the form of a single scalar, such as using CPU utilization or memory usage as representative values ​​of task resource requirements. In actual execution, tasks usually occupy multiple types of resources simultaneously, including computing resources, cache resources, storage resources, and bandwidth resources. If only a single resource indicator is used for description, it will lead to the loss of multi-dimensional information about resources, making it impossible for the scheduling system to accurately characterize the true resource structure characteristics of the task. Different resource dimensions have different roles in scheduling decisions, and a single indicator is difficult to reflect the coupling relationship between resources, thus affecting the accuracy of subsequent resource allocation and core mapping.

[0045] Constructing a multi-dimensional unified expression form enables tasks to possess a complete resource structure description capability when entering the scheduling model. Let the task... The expression for the resource demand feature vector is: ; in, Indicates task In the Demand representation quantity in the category of resources.

[0046] By incorporating behavioral feature data and runtime context information during task execution, a user intent recognition model is constructed, and a task service tolerance index is generated.

[0047] In existing technologies, task scheduling decisions are usually based solely on resource requirements, lacking a unified quantitative expression of task service characteristics. For example, different tasks may differ in response latency requirements, execution stability, and tolerance for resource fluctuations. However, these differences are usually represented by discrete rules or labels, making it difficult to directly participate in the calculation. This results in tasks with similar resource requirements exhibiting completely different execution effects during scheduling, thus affecting the rationality of scheduling decisions.

[0048] By uniformly quantifying and modeling the service semantics of tasks, they can be transformed into computable parameters that participate in the subsequent scheduling process. To enable these differences to participate in subsequent boundary calculations, the expression for uniformly mapping them to tolerance indices is as follows: ; in, Indicates task latency sensitivity This indicates the requirement for execution stability. Indicates the degree of sensitivity to resource consumption. , , These are weighting coefficients, enabling previously dispersed service requirements to be uniformly incorporated into the scheduling model.

[0049] By using a linear combination method, various tolerance factors can be comprehensively included in the calculation according to preset weights, thereby transforming high-level service semantics into computable parameters for the subsequent generation of resource boundaries and constraints.

[0050] In existing technologies, resource requirements are usually used directly for scheduling decisions in the form of average or peak values, which only reflects the level of resource consumption by tasks without considering the task's ability to adapt to resource fluctuations. If service tolerance is introduced as an independent constraint into the scheduling process, it is easy to cause a disconnect between the expression of resource requirements and service attributes, making it difficult to form a unified decision basis. It is necessary to construct an adjustment mechanism that can couple resource requirements with task tolerance, so that the boundary of resource requirements can be dynamically adjusted according to the service characteristics of tasks.

[0051] After obtaining the resource demand feature vector and tolerance index, construct the resource demand adjustment function: ; in, Indicates task At any moment Adjusted resource demand vector, This is the adjustment coefficient.

[0052] In existing technologies, resource requirements are usually based on the predicted value at a single point in time as the basis for scheduling. However, tasks often exhibit obvious time fluctuations during operation. If scheduling is based solely on resource requirements at a certain moment, it is easy to cause insufficient resource allocation or redundant allocation problems when the load fluctuates, thereby affecting system stability.

[0053] Resource demand needs to be modeled from a time window perspective so that scheduling decisions can reflect the range of task changes over a future period, within a preset time window. Within this framework, the upper and lower bounds of dynamic resource requirements for each resource dimension are defined: ; ; in, and Representing tasks In the Lower and upper bounds of requirements in the resource category dimension.

[0054] In existing technologies, the adaptability of tasks to resource competition environments is usually not explicitly modeled, or is only coarsely classified based on task type. This fails to reflect the combined impact of task service requirements and resource demand fluctuations on scheduling stability. If judgment is based solely on service tolerance, the uncertainty brought about by resource fluctuations is ignored. If judgment is based solely on resource fluctuations, the characteristics of task services cannot be reflected.

[0055] A unified metric is constructed to ensure that task interference tolerance is simultaneously constrained by service characteristics and resource fluctuations. To enable resource analysis results to directly serve subsequent mapping decisions, the task interference tolerance level is further defined: ; This allows the task's tolerance to interference to simultaneously reflect service characteristics and resource fluctuation features, thus providing a unified basis for the generation of subsequent mapping constraint parameters.

[0056] After obtaining the interference tolerance level, generate the task mapping constraint parameters: ; in, For the task The mapping constraint parameters, This is a mapping function from the interference tolerance level to the mapping constraint parameters.

[0057] Since the interference tolerance level is still an expression at the task attribute level, while the subsequent core screening process requires constraints that can be directly compared and eliminated, the interference tolerance level is converted into task mapping constraint parameters through a mapping function. This allows the results of the preceding resource analysis to be directly passed to the subsequent core screening steps, thereby realizing the continuous derivation from "resource requirements" to "scheduling constraints".

[0058] From the characteristics of task resource requirements and service tolerance to the dynamic resource requirement boundary, interference tolerance level and mapping constraint parameters, a continuous generation chain is formed. Through this continuous generation chain, task-side information can directly participate in subsequent core mapping decisions. By transforming resource analysis results into constraints, the problem of separation between resource prediction and scheduling decisions in existing technologies is eliminated, thereby providing input basis with individual task differences for subsequent mapping stages.

[0059] S3. Based on the dynamic resource demand boundary values ​​and task attribute information, classify the set of tasks to be scheduled, divide the tasks that meet the preset binding conditions into the fixed task set, and divide the remaining tasks into the migration task set.

[0060] Furthermore, the system dynamically generates task binding criteria, performs traversal matching and judgment on the task set based on these criteria, and outputs the task classification results.

[0061] Based on the classification results, a set partitioning operation is performed, dividing tasks that meet the binding conditions into a fixed task set, and dividing the remaining tasks into a migration task set.

[0062] It should be noted that in existing technologies, task division is usually based on fixed rules, such as whether it is portable, whether it is a critical task, or whether it is a real-time task.

[0063] While static partitioning is simple to implement, it does not take into account the changes in task resource requirements within a preset time window, nor does it consider the task's adaptability to complex execution environments. As a result, tasks that should maintain stable execution are easily included in the migration set, or tasks that should be scheduled through global optimization are prematurely fixed to local resources.

[0064] By dynamically generating task binding criteria, the task classification process no longer relies on pre-fixed task type labels, but instead on dynamic resource requirement boundary values ​​and task attribute information obtained in previous steps.

[0065] The task binding determination criteria may include at least one or more of the following: resource boundary fluctuation threshold, service tolerance threshold, and migration sensitivity constraint.

[0066] Whether a task is suitable for fixed execution should not only depend on static attributes, but also on whether the resource boundaries are stable, whether the operational constraints are strict, and whether it is appropriate to frequently participate in migration mapping during subsequent scheduling.

[0067] By combining resource boundaries with task attributes, task classification results can reflect the adaptability of tasks in the current scheduling environment.

[0068] Dynamic decision criteria ensure that the classification process remains consistent with the preceding resource analysis.

[0069] This avoids the coarse-grained problem caused by task division based solely on fixed rules in existing technologies.

[0070] The task set is matched and judged in a traversal manner according to the judgment conditions, and the tasks that meet the binding conditions are divided into fixed task sets based on the classification results, while the remaining tasks are divided into migration task sets.

[0071] This allows a fixed task set to handle tasks that require high execution stability, are sensitive to environmental changes, or are not suitable for frequent remapping.

[0072] Make the migration task set contain tasks suitable for entering the subsequent global optimization mapping process.

[0073] By performing structural routing in advance, the scale of tasks required for subsequent complex matching calculations is reduced. This avoids the redundant overhead caused by performing complex mapping calculations on all tasks uniformly in existing technologies.

[0074] S4. For the migration task set, construct the association graph between task nodes and heterogeneous core nodes, and generate the initial heterogeneous mapping scheme based on the edge weight using the maximum weight matching algorithm.

[0075] Furthermore, node-based modeling is performed on the migration task set and heterogeneous core resources to obtain task node set and core node set, respectively.

[0076] Based on the coupling relationship between task execution requirements and core capability characteristics, a connection topology between nodes is established and corresponding adaptation weights are generated.

[0077] Based on node sets, connection topology, and adaptation weights, a weighted association graph structure representing the relationship between the task and the core matching is constructed.

[0078] Obtain the runtime status data of heterogeneous cores, including core load rate, resource utilization rate and execution status, and construct a core status vector.

[0079] Extract the computational features of the tasks currently being executed by the heterogeneous cores, construct node computational content feature vectors, and use them to characterize the current computational semantic environment of the cores.

[0080] Based on the core's shared resource access behavior, a resource competition model between cores is established, and an interference matrix between cores is constructed to characterize the intensity of resource competition and neighborhood interference relationships between different cores.

[0081] Based on the task resource requirement feature vector and the core state vector, a task-core matching degree function is constructed to perform a preliminary calculation on the adaptation relationship between the task and the core.

[0082] Calculate the content relevance index between the task to be scheduled and the feature vector of the computation content of the node. The content relevance index is used to characterize the computation reuse potential between tasks and the advantage of cache locality.

[0083] By jointly analyzing the interference matrix, task scheduling feature vector, and content relevance index, the interference risk value of the task on different heterogeneous cores can be determined.

[0084] Based on the task mapping constraint parameters, candidate core nodes are screened, and core nodes with interference risk values ​​higher than the interference tolerance level corresponding to the task are removed.

[0085] The matching function is weighted and corrected based on the content relevance index to increase the matching weight of tasks that are highly relevant to the current core computing content. An interference suppression factor is introduced to apply differentiated interference constraints to different tasks.

[0086] Based on the modified matching degree function, the task nodes and core nodes are optimized and matched to determine the mapping relationship between the task and the core.

[0087] Generate an initial heterogeneous mapping scheme based on the mapping relationship.

[0088] It should be noted that in the existing technology, the mapping between tasks and heterogeneous cores is usually based on the static matching relationship between task requirements and core capabilities to construct edge weights. That is, as long as the core has sufficient resources, it is considered suitable to carry out the task.

[0089] However, it does not take into account what computations the core is currently performing, nor does it consider the competition for shared resources in the surrounding environment where the core is located, thus making it difficult to reflect the true suitability of task execution on different cores.

[0090] Node-based modeling is performed on the migration task set and heterogeneous core resources to obtain task node sets and core node sets respectively, and a weighted association graph structure representing the matching relationship between tasks and core resources is constructed.

[0091] It provides a unified carrying structure for subsequent matching factors of different dimensions, so that resource status, content reuse capability and interference constraints can work together to form the task-core edge weight.

[0092] Obtain the running status data of the heterogeneous core, construct the core status vector, and extract the computational features of the tasks currently being executed by the heterogeneous core to construct the node computation content feature vector.

[0093] In existing technologies, the matching of tasks and cores is usually based solely on resource capacity without considering the impact of the core's current execution content on new tasks. However, the currently executing task on the core will form a specific cache state, intermediate results, and data access path, which will affect the execution efficiency of new tasks. If ignored, it may lead to the task being assigned to a core with sufficient resources but mismatched computational semantics, reducing execution efficiency. Therefore, it is necessary to build a similarity measurement mechanism between tasks and the current computational content of the core.

[0094] In existing technologies, task mapping usually relies on a single matching index, which makes it difficult to simultaneously take into account the three types of factors: resource matching, content reuse, and interference suppression. Simple weighting methods are also prone to causing different factors to cancel each other out, failing to reflect the independent role of each factor. Therefore, it is necessary to construct a matching model with the synergistic effect of multiple factors.

[0095] To quantify the correlation between the tasks to be scheduled and the core current computational content, feature vectors of task content are constructed respectively. Calculate the content feature vector with the node And define content relevance metrics: ; In existing technologies, inter-core interference is usually described using a fixed model, but it is assumed that different tasks are affected in a consistent manner under the same environment.

[0096] In reality, different tasks have significantly different sensitivities to interference due to differences in computational content. It is necessary to incorporate the characteristics of the task content into the interference modeling process so that the interference evaluation results reflect task differences.

[0097] In existing technologies, content relevance is usually only used to reflect the general semantic similarity between tasks or between a task and the core. However, it is difficult to distinguish between the two types of scenarios: "similar only" and "can directly inherit existing computational context". When there is a subtask relationship, local continuation relationship or computational inclusion relationship between the task to be scheduled and the currently executing task of the core, the reusable cached state, intermediate results and execution path are usually stronger. Therefore, it is necessary to introduce an enhancement mechanism on the basis of general content relevance to highlight the priority of strong reuse scenarios in mapping decisions.

[0098] Introducing a subtask enhancement mechanism: ; in, Indicates task With core Subtask relationship indicator between currently executing tasks, if tasks With core If the current task's subtasks or local computational content are involved, then... ,otherwise , Enhancement factor for subtasks.

[0099] In existing technologies, task resource consumption is usually regarded as a fixed value independent of the running location. That is, it is assumed that the resource consumption of the same task is basically the same on different cores. If the task has a high degree of reuse relationship with the current computing content of the core, the cache hit rate, intermediate result reuse rate and access path inheritance degree of the task on the core may be higher, thereby reducing the actual resource consumption. Therefore, it is necessary to build an evaluation mechanism that can reflect the impact of content reuse on resource consumption.

[0100] Construct a resource usage evaluation function: ; in, Indicates task Total original demand across all resource dimensions.

[0101] By introducing a content enhancement factor into the denominator The higher the reusability of tasks and core current content, the lower the equivalent resource consumption assessment value. This transforms "content reuse potential" into a quantitative expression of "reduced resource consumption," thereby breaking through the assumption of fixed resource consumption in existing technologies and making the resource assessment results related to the core current computing semantic environment.

[0102] Based on the core's shared resource access behavior, a resource competition model between cores is established, an interference matrix between cores is constructed, and the interference risk value of tasks on different heterogeneous cores is determined.

[0103] To ensure that interference risk is affected not only by competition in the core neighborhood but also by the task's and current core content's reusability capabilities, the interference risk value is defined as: ; in, Indicates core The neighborhood core set, Indicates core With neighboring core The intensity of interference between them.

[0104] Core neighborhood interference does not have the same impact on all tasks: if a task has a high content reuse capability with the current core, it may reduce repeated loading and shorten the cold start phase during execution, thereby mitigating the negative impact of neighborhood competition to some extent. This can be achieved by introducing interference matrices... This approach ensures that the higher the content reuse capability, the lower the equivalent interference risk. It unifies "content relevance" and "neighborhood competitive environment" into the same risk expression, thereby forming an interference assessment model that is more in line with actual execution scenarios.

[0105] Candidate core nodes are filtered based on task mapping constraint parameters: ; The reason for screening before matching calculation is that not all cores with resource capacity are suitable as candidate cores. Only when the equivalent interference risk of a core is not higher than the constraint level that the task itself can tolerate is the core suitable to enter the subsequent optimization matching process. This allows the constraint parameters generated on the task side to directly affect the formation of the candidate set on the core side. By screening before calculating, unsuitable cores are reduced from entering the matching calculation, thereby substantially transmitting the analysis results to the mapping decision.

[0106] In existing technologies, task mapping usually relies on a single matching index, which makes it difficult to simultaneously take into account the three types of factors: resource matching, content reuse, and interference suppression. Simple weighting methods are also prone to causing different factors to cancel each other out, failing to reflect the independent role of each factor. Therefore, it is necessary to construct a matching model with the synergistic effect of multiple factors.

[0107] After completing the candidate core screening, construct the revised matching function: ; in, With core The state vector, For content reuse gain factor, This represents the interference suppression coefficient.

[0108] The matching function enables resource matching, content reuse, and interference suppression capabilities to work synergistically in the task-core mapping process.

[0109] Based on the modified matching degree function, the task node and core node are optimized and matched to determine the mapping relationship between the task and the core, and an initial heterogeneous mapping scheme is generated. This realizes the upgrade of mapping decision from "whether the resources are sufficient" to "whether the resources are suitable, whether the content is reusable, and whether the environment is acceptable", thereby breaking through the processing limitations of the existing technology that simply relies on resource capacity matching for task scheduling.

[0110] S5. Based on the initial heterogeneous mapping scheme, dynamically schedule the migration task set and allocate corresponding execution resources.

[0111] Furthermore, task scheduling is performed based on the initial heterogeneous mapping scheme, and the task execution environment and resource usage are monitored in real time to obtain a dynamic scheduling state set.

[0112] Anomaly detection and deviation assessment are performed on the dynamic scheduling state set to determine the degree of deviation of the current execution state from the initial plan.

[0113] Based on the degree of deviation, the task scheduling strategy is corrected online, the mapping relationship between tasks and cores is updated, and computing resources are reallocated to complete the dynamic scheduling process.

[0114] It should be noted that, in the existing technology, although many scheduling schemes can generate mapping results in the initial stage, the mapping results are usually based on the static state of the task when it enters the scheduling system.

[0115] When core load changes, shared resource contention changes, task execution progress shifts, or node computation content changes during system operation, the original mapping relationship may no longer be applicable.

[0116] If the initial plan is still mechanically executed, it is easy for the complex modeling and mapping decisions in the preceding stages to fail during runtime, resulting in a processing loophole of "refined front-end decision-making and rigid back-end execution".

[0117] After obtaining the initial heterogeneous mapping scheme, instead of executing it directly, the migration task set is dynamically scheduled according to the initial heterogeneous mapping scheme, and the task execution environment and resource usage are monitored in real time to obtain the dynamic scheduling status set.

[0118] Although the initial mapping scheme has taken into account resource matching, content reuse, and interference suppression, these factors may change during operation. For example, interference in the core neighborhood may decrease when other tasks finish or increase when new parallel tasks enter. The current computation content of the core may also change after a period of execution, thus affecting the content reuse capability between new tasks and the core. Therefore, it is necessary to continuously monitor the dynamic scheduling state set during operation in order to determine whether the current execution state still conforms to the design basis of the initial mapping scheme.

[0119] In existing technologies, dynamic scheduling either relies on fixed-period rescheduling or triggers adjustments only after a significant state anomaly, making it difficult to balance timely response with migration overhead control. If remapping is triggered for all state changes, excessive migration and additional resource consumption can easily occur. If correction is not triggered for a long period, the initial mapping scheme may gradually become invalid during runtime. Therefore, it is necessary to construct an evaluation mechanism that can comprehensively reflect the degree of deviation of the current execution state from the initial mapping basis, so as to provide quantitative triggering conditions for online correction.

[0120] Anomaly detection and deviation assessment are performed on the dynamic scheduling state set, and a deviation degree function is constructed: ; in, Indicates task The actual resource demand vector at the current moment. This represents the resource requirement vector during the initial mapping phase. This indicates the current interference risk value. This represents the interference risk value during the initial mapping. This indicates the current estimated resource usage. This represents the estimated resource usage value at the time of initial mapping.

[0121] Task deviation should not be judged from a single dimension, but should comprehensively consider three types of factors: changes in resource demand, changes in the interfering environment, and changes in resource occupancy. The degree of deviation in resource demand should be reflected by the vector difference term, the changes in environmental competition should be reflected by the risk difference term, and the changes in actual execution costs should be reflected by the occupancy difference term, thereby forming a unified indicator that can comprehensively characterize the degree of deviation in task operation status.

[0122] When the deviation exceeds a preset threshold, the task scheduling strategy is corrected online, the mapping relationship between tasks and cores is updated, and computing resources are reallocated.

[0123] When the deviation exceeds a preset threshold, an online correction process is triggered to update the mapping relationship between tasks and cores and reallocate computing resources.

[0124] By setting threshold trigger boundaries, dynamic correction is only initiated when the current execution state deviates from the initial plan to a predetermined extent, thus balancing operational stability and scheduling flexibility.

[0125] During the online correction process, priority can be given to recalculating the candidate core set for migration tasks that are more affected, and the matching relationship can be locally updated based on the updated content relevance, interference risk value and resource consumption assessment value.

[0126] Not all tasks need to be remapped during runtime. Only tasks with significant deviations or changes in the execution environment require localized adjustments. This reduces the overall system disturbance caused by rescheduling and maintains the stability of most mapping relationships through local updates. Incremental adjustments to key tasks improve dynamic scheduling efficiency, thus balancing operational stability and scheduling flexibility.

[0127] This embodiment also provides a multi-core chip task scheduling optimization system, including: The module includes a task modeling module, a boundary calculation module, a task classification module, a mapping generation module, and a dynamic scheduling module.

[0128] The task modeling module is used to obtain the attribute information of each task in the task set to be scheduled, and to build a task model based on the attribute information.

[0129] The boundary calculation module is used to calculate the distribution range of resource demand for each task within a preset time window based on the task model and preset risk adjustment parameters, and to determine the corresponding dynamic resource demand boundary value.

[0130] The task classification module is used to classify the set of tasks to be scheduled based on the dynamic resource demand boundary value and task attribute information. Tasks that meet the preset binding conditions are divided into fixed task sets, and the remaining tasks are divided into migration task sets.

[0131] The mapping generation module is used to construct an association graph between task nodes and heterogeneous core nodes for the migration task set, and generate an initial heterogeneous mapping scheme based on the edge weight using the maximum weight matching algorithm.

[0132] The dynamic scheduling module is used to dynamically schedule the migration task set based on the initial heterogeneous mapping scheme and allocate the corresponding execution resources.

[0133] This embodiment also provides a computer device applicable to the multi-core chip task scheduling optimization method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the multi-core chip task scheduling optimization method proposed in the above embodiment.

[0134] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0135] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the multi-core chip task scheduling optimization method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0136] In summary, this invention effectively reduces the global matching computation scale and scheduling overhead by classifying tasks based on dynamic boundaries, fixing highly stable tasks, and migrating appropriately optimized tasks. For the migration task set, it constructs a correlation graph that integrates content relevance and interference risk for maximum weight matching, and uses the current computational semantic environment of the core to correct resource occupancy assessment, thus achieving resource savings and interference suppression brought about by "computation content reuse".

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-core chip task scheduling optimization method, characterized in that: include: Obtain the attribute information of each task in the set of tasks to be scheduled, and construct a task model based on the attribute information; Based on the task model and preset risk adjustment parameters, calculate the resource demand distribution range of each task within a preset time window, and determine the corresponding dynamic resource demand boundary value. Based on the dynamic resource demand boundary value and task attribute information, the task set to be scheduled is classified, and tasks that meet the preset binding conditions are divided into fixed task sets, and the remaining tasks are divided into migration task sets. For the migration task set, an association graph between task nodes and heterogeneous core nodes is constructed, and an initial heterogeneous mapping scheme is generated based on the edge weights using the maximum weight matching algorithm. Based on the initial heterogeneous mapping scheme, the migration task set is dynamically scheduled and corresponding execution resources are allocated.

2. The multi-core chip task scheduling optimization method as described in claim 1, characterized in that: The construction task model includes: Standardize and semantically encapsulate the multi-dimensional heterogeneous attribute information of each task in the task set to be scheduled; A unified computing object is constructed based on the mapped attribute information. The unified computing object is used to represent the vector features of the task in the scheduling space. Based on the vector features, a task model is formed for subsequent scheduling algorithm parsing and computation.

3. The multi-core chip task scheduling optimization method as described in claim 2, characterized in that: The determination of dynamic resource demand boundary values ​​includes: Model the resource usage data during task execution, construct a task resource demand model based on historical task execution information and changes in resource usage, extract resource demand fluctuation characteristics and predict peak resource demand to form a task resource demand feature vector; Based on behavioral feature data and runtime context information during task execution, a user intent recognition model is constructed to classify the intent of the task and obtain user intent labels. Based on the user intent tags, a task service tolerance quantification model is constructed, mapping the task's response latency requirements, execution stability requirements, and resource consumption sensitivity to tolerance indicators. The resource demand feature vector is coupled with the tolerance index to construct a resource demand adjustment function, and the dynamic distribution range and corresponding upper and lower bounds of task resource demand are determined within a preset time window to form dynamic resource demand boundary values. Based on the user intent tags, tolerance indicators, and dynamic resource demand boundary values, the interference tolerance level of the task on the resource competition between cores is determined. Based on the interference tolerance level, task mapping constraint parameters are generated.

4. The multi-core chip task scheduling optimization method as described in claim 3, characterized in that: The classification of the set of tasks to be scheduled includes: Dynamically generate task binding judgment conditions, perform traversal matching judgment on the task set based on the judgment conditions, and output task classification results; Based on the classification results, a set partitioning operation is performed, dividing tasks that meet the binding conditions into a fixed task set, and dividing the remaining tasks into a migration task set.

5. The multi-core chip task scheduling optimization method as described in claim 4, characterized in that: The construction of the association graph includes: The migration task set and heterogeneous core resources are modeled in a node-based manner to obtain the task node set and the core node set, respectively. Based on the coupling relationship between task execution requirements and core capability characteristics, a connection topology between nodes is established and corresponding adaptation weights are generated. Based on the node set, connection topology, and adaptation weights, a weighted association graph structure representing the relationship between the task and the core matching is constructed.

6. The multi-core chip task scheduling optimization method as described in claim 5, characterized in that: The initial heterogeneous mapping generation scheme includes: Obtain the operating status data of heterogeneous cores, including core load rate, resource utilization rate and execution status, and construct a core status vector; Extract the computational features of the tasks currently being executed by the heterogeneous cores, and construct node computational content feature vectors to characterize the current computational semantic environment of the cores; Based on the core's shared resource access behavior, a resource competition model between cores is established, and an interference matrix between cores is constructed to characterize the intensity of resource competition and neighborhood interference relationships between different cores. Based on the task resource requirement feature vector and the core state vector, a task-core matching degree function is constructed to perform a preliminary calculation on the adaptation relationship between the task and the core. Calculate the content correlation index between the task to be scheduled and the feature vector of the node's computational content. The content correlation index is used to characterize the computational reuse potential and cache locality advantage between tasks. By jointly analyzing the interference matrix, task scheduling feature vector, and content relevance index, the interference risk value of the task on different heterogeneous cores can be determined. Based on the task mapping constraint parameters, candidate core nodes are screened, and core nodes with interference risk values ​​higher than the interference tolerance level corresponding to the task are eliminated. The matching degree function is weighted and corrected based on the content relevance index to increase the matching weight of tasks that are highly relevant to the current core computing content, and an interference suppression factor is introduced to apply differentiated interference constraints to different tasks. Based on the modified matching degree function, the task node and the core node are optimized and matched to determine the mapping relationship between the task and the core. An initial heterogeneous mapping scheme is generated based on the mapping relationship.

7. The multi-core chip task scheduling optimization method as described in claim 6, characterized in that: The dynamic scheduling includes: Task scheduling is performed based on the initial heterogeneous mapping scheme, and the task execution environment and resource usage are monitored in real time to obtain a dynamic scheduling status set; Anomaly detection and deviation assessment are performed on the dynamic scheduling state set to determine the degree of deviation of the current execution state from the initial plan; Based on the degree of deviation, the task scheduling strategy is corrected online, the mapping relationship between tasks and cores is updated, and computing resources are reallocated to complete the dynamic scheduling process.

8. A multi-core chip task scheduling optimization system, based on the multi-core chip task scheduling optimization method according to any one of claims 1 to 7, characterized in that: include: The module includes a task modeling module, a boundary calculation module, a task classification module, a mapping generation module, and a dynamic scheduling module. The task modeling module is used to obtain the attribute information of each task in the task set to be scheduled, and to construct a task model based on the attribute information. The boundary calculation module is used to calculate the resource demand distribution range of each task within a preset time window based on the task model and preset risk adjustment parameters, and to determine the corresponding dynamic resource demand boundary value. The task classification module is used to classify the set of tasks to be scheduled according to the dynamic resource demand boundary value and task attribute information, and to classify the tasks that meet the preset binding conditions into fixed task sets, and the remaining tasks into migration task sets. The mapping generation module is used to construct an association graph between task nodes and heterogeneous core nodes for the migration task set, and generate an initial heterogeneous mapping scheme based on the edge weights using the maximum weight matching algorithm. The dynamic scheduling module is used to dynamically schedule the migration task set based on the initial heterogeneous mapping scheme and allocate corresponding execution resources.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the multi-core chip task scheduling optimization method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the multi-core chip task scheduling optimization method according to any one of claims 1 to 7.

Images