Computer resource intelligent scheduling and optimization device for multitask concurrency

Abstract

The application discloses a computer resource intelligent scheduling and optimization device for multitask concurrency, comprising a task access unit, a resource pool management unit and a scheduling execution unit, further comprising a task multi-dimensional portrait unit, a heterogeneous resource full-state perception unit, a double-layer intelligent scheduling decision unit, a resource optimization and conflict resolution unit and a closed-loop self-iteration optimization unit; the application adopts modular design, is compatible with mainstream CPU, GPU, NPU, FPGA and other heterogeneous computing power resources, supports all types of task loads such as real-time computing, offline batch processing, AI training / inference and transaction processing, can seamlessly connect mainstream scheduling frameworks such as Kubernetes, YARN and Slurm, and is suitable for various multitask concurrency scenes such as public cloud, private cloud, edge computing and high-performance computing. Meanwhile, the application supports multi-tenant isolation and hierarchical QoS guarantee, can meet the scheduling demands of different industries such as finance, internet, scientific research and industrial control, and has strong universality, practicality and popularization value.

Description

Technical Field

[0001] This invention relates to the fields of computer system architecture, cloud computing and heterogeneous computing resource scheduling technology, and in particular to an intelligent scheduling and optimization device for computer resources oriented towards multi-task concurrency. Background Technology

[0002] With the rapid development of cloud computing, big data, artificial intelligence big data models, and edge computing industries, the architecture of data centers and computing clusters is showing two major evolutionary trends: First, computing power is becoming heterogeneous, with deep integration of heterogeneous computing units such as general-purpose CPUs, GPUs, NPUs, and FPGAs. The cluster resource dimension has expanded from the traditional binary structure of CPU and memory to a multi-dimensional heterogeneous resource system including computing power, video memory, storage I / O, network bandwidth, NUMA topology, and energy consumption. Second, the workload is becoming more complex, with clusters simultaneously carrying multiple tenants and various types of concurrent tasks, including hard real-time control tasks, soft real-time online inference tasks, high-throughput offline training tasks, batch big data processing tasks, and transactional database tasks. The resource requirements, timing constraints, QoS (Quality of Service) requirements, and execution characteristics of different tasks vary greatly, placing unprecedentedly high demands on the fineness, real-time performance, reliability, and global optimization capabilities of resource scheduling.

[0003] Resource scheduling is the core of a computing cluster, directly determining the cluster's service quality, resource utilization efficiency, energy consumption, and business stability. Current mainstream resource scheduling technologies and commercial scheduling systems, including the Kubernetes default scheduler, the Hadoop YARN scheduler, the Slurm high-performance computing scheduler, self-developed scheduling systems from public cloud vendors such as Alibaba Cloud and Huawei Cloud, and intelligent scheduling schemes based on heuristic algorithms and reinforcement learning proposed by academia, all suffer from insurmountable technical shortcomings in heterogeneous cluster scenarios with multiple concurrent tasks. These shortcomings are detailed below: 1. Most existing scheduling schemes are based on static resource quota declarations at the time of task submission. They can only identify basic resource requirements such as CPU and memory, and cannot fully characterize the dynamic resource requirements, execution timing constraints, QoS levels, computing power dependence characteristics, and data affinity of a task throughout its entire lifecycle. They lack standardized multi-dimensional task profiling capabilities. On the one hand, to ensure the stability of task execution, users often over-declare resource quotas, resulting in a large amount of resources being idle for a long time and extremely low resource utilization. On the other hand, the peak resource demand during task execution cannot be predicted in advance, and static quotas cannot cope with dynamic fluctuations. This easily leads to task timeouts and QoS failures due to insufficient resource supply, especially in scenarios with tens of thousands of concurrent tasks, where this problem is amplified dramatically.

[0004] 2. Existing scheduling architectures fall into two core modes: First, a centralized global scheduling architecture, which provides a global resource view and enables cluster-level optimization. However, it suffers from long scheduling decision chains and high latency, typically with second-level scheduling cycles, failing to meet the millisecond-level scheduling requirements of hard real-time tasks and unable to handle sudden load fluctuations. Second, a distributed local scheduling architecture, where scheduling decisions are executed locally on each node, resulting in low latency. However, it lacks a global resource view and is prone to uneven load distribution among nodes, cross-node resource contention, and increased fragmentation of global resources, making it unable to achieve cluster-level multi-objective optimization. Current technologies consistently fail to resolve the core contradiction between "global optimality" and "real-time responsiveness," and cannot simultaneously meet the dual requirements of ensuring hard real-time task performance and optimizing global cluster resources.

[0005] 3. Existing technologies for collecting cluster resource status data are mostly coarse-grained, at the second level, only able to obtain basic indicators such as CPU utilization and memory usage. They cannot achieve microsecond-level, full-dimensional status awareness, nor can they capture instantaneous performance fluctuations of heterogeneous resources. Furthermore, they lack in-depth identification and quantitative evaluation of resource performance bottlenecks, NUMA topology access latency, shared cache contention, I / O bandwidth bottlenecks, energy consumption characteristics, and hardware stability, making it impossible to construct a comprehensive and accurate resource status matrix. This coarse-grained and low-dimensional resource awareness directly leads to a disconnect between scheduling decisions and the actual operating state of the cluster. This easily results in situations where the node resource status has changed after a scheduling decision is issued, causing scheduling failures, resource conflicts, and abnormal task execution.

[0006] 4. Cluster scheduling is essentially a multi-objective optimization problem, requiring simultaneous consideration of five interrelated and even conflicting core objectives: task completion rate, QoS compliance rate, resource utilization, system energy consumption, and multi-tenant fairness. Existing heuristic scheduling algorithms (such as fair scheduling, capacity scheduling, and first-come, first-served) can only achieve simple optimization of a single or dual objective, failing to achieve a coordinated balance among multiple objectives. Existing intelligent scheduling schemes based on reinforcement learning mostly suffer from problems such as action space explosion, high model convergence difficulty, and high decision latency. Furthermore, they often employ a single reward function, focusing only on improving resource utilization and failing to simultaneously guarantee QoS for hard real-time tasks, tenant fairness, and energy consumption optimization. This easily leads to problems such as high-priority tasks preempting all resources, low-priority tenants "starving," and persistently high cluster energy consumption.

[0007] 5. In high-concurrency, multi-tasking scenarios, a large number of tasks simultaneously compete for shared resources, which can easily lead to problems such as CPU shared cache contention, I / O bandwidth contention, GPU computing power / memory conflicts, and network port conflicts. Most existing technologies are reactive and reactive, lacking a full-process conflict handling mechanism that includes pre-detection and in-process hierarchical resolution. Resource conflicts often cause a surge in task execution latency or even execution failure. At the same time, the distributed scheduling of discrete tasks leads to severe resource fragmentation, especially the fragmentation of GPU memory, large page memory, and CPU cores. This results in the industry pain point of "sufficient total idle resources in the cluster, but no continuous available resource blocks to complete task scheduling." Existing technologies lack targeted fragmentation governance mechanisms, which further limits the improvement of cluster resource utilization.

[0008] 6. Most existing scheduling systems' scheduling strategies and algorithm models are trained on historical data for fixed scenarios and then deployed, lacking closed-loop feedback and autonomous iteration capabilities throughout the entire lifecycle. When cluster hardware architecture is upgraded, task load types change, or business peak characteristics change, the effectiveness of the original scheduling strategies will drop sharply, requiring manual data collection, model training, and strategy adjustments, resulting in extremely high operation and maintenance costs and making it impossible to adapt to rapidly iterating business needs and changes in the cluster environment.

[0009] In summary, existing resource scheduling technologies suffer from numerous core defects in heterogeneous computing scenarios with multi-tenant and multi-task concurrency. These defects include inaccurate task characterization, coarse resource awareness, inability of the architecture to balance global optimization and real-time response, weak multi-objective collaborative optimization capabilities, prominent resource conflicts and fragmentation issues, and lack of closed-loop self-iterative capabilities. These technologies cannot meet the current computing clusters' scheduling requirements for high reliability, high performance, high utilization, low energy consumption, and fairness. Therefore, a new intelligent scheduling and optimization device is urgently needed to systematically solve the above-mentioned technical pain points. Summary of the Invention

[0010] In order to overcome the shortcomings of the prior art, one of the objectives of this invention is to provide a computer resource intelligent scheduling and optimization device for multi-task concurrency.

[0011] One of the objectives of this invention is achieved through the following technical solution: A computer resource intelligent scheduling and optimization device for multi-task concurrency includes a task access unit, a resource pool management unit, and a scheduling execution unit. It also includes a task multi-dimensional profiling unit, a heterogeneous resource full-state perception unit, a two-layer intelligent scheduling decision unit, a resource optimization and conflict resolution unit, and a closed-loop self-iterative optimization unit. The input end of the task multi-dimensional profiling unit is connected to the output end of the task access unit, and is used to extract full life cycle features of concurrently accessed scheduled tasks to generate a multi-dimensional task profile that includes task static attributes, dynamic resource requirements, execution timing constraints, QoS level requirements and computing power dependence features. The input end of the heterogeneous resource full-state perception unit is connected to the output end of the resource pool management unit, and is used to collect the real-time full-dimensional state of the computer heterogeneous resource pool, and generate a resource state matrix that includes the real-time load, available margin, performance bottleneck and energy consumption characteristics of CPU, memory, GPU, storage IO, network bandwidth resources. The first input of the dual-layer intelligent scheduling decision unit is connected to the output of the task multi-dimensional profiling unit, and the second input is connected to the output of the heterogeneous resource full-state perception unit. It is used to perform dual-layer scheduling decisions of global pre-scheduling and local real-time scheduling based on multi-dimensional task profiling and resource state matrix, and generate initial resource allocation scheme and real-time scheduling strategy. The input end of the resource optimization and conflict resolution unit is connected to the output end of the dual-layer intelligent scheduling decision unit, and the output end is connected to the input end of the scheduling execution unit. It is used to perform fragmentation management and optimization on the initial resource allocation scheme, and at the same time to detect and resolve resource competition conflicts in the process of concurrent execution of multiple tasks in real time, and generate the final executable resource scheduling instructions. The input of the closed-loop self-iterative optimization unit is connected to the output of the scheduling execution unit, and the output is connected to the optimization input of the dual-layer intelligent scheduling decision unit. It is used to collect actual task execution data and resource operation data, generate scheduling effect evaluation indicators, and iteratively optimize the scheduling model parameters of the dual-layer intelligent scheduling decision unit based on the evaluation indicators.

[0012] As a further improvement to the above technical solution: The multi-dimensional task profiling unit includes a task feature extraction subunit, a resource demand prediction subunit, and a profiling classification subunit. The task feature extraction subunit is used to extract the static attribute features of the task to be scheduled, including task type, tenant, priority, execution deadline, parallelizability and dependency. The resource demand prediction subunit is used to predict the dynamic CPU, memory, GPU, IO and network bandwidth resource demand curves of a task throughout its entire execution cycle based on the task's historical execution data and the computational logic characteristics of the current task, using a lightweight time-series prediction model. The profile classification subunit is used to classify tasks into three categories—hard real-time guarantee tasks, soft real-time guarantee tasks, and non-real-time throughput tasks—based on the task's QoS level requirements, timing constraint strictness, and resource demand characteristics, and to generate corresponding profile tags for each category of tasks.

[0013] The heterogeneous resource full-state perception unit includes a resource state acquisition subunit, a performance bottleneck identification subunit, and a resource availability assessment subunit. The resource status acquisition subunit adopts a distributed probe architecture and collects data on CPU core utilization, memory usage, GPU computing power utilization, video memory usage, storage IOPS, network throughput and node energy consumption of each node in the computer heterogeneous resource pool in real time with a preset microsecond-level acquisition cycle. The performance bottleneck identification subunit is used to identify performance bottleneck nodes, overloaded resource dimensions, and idle resource fragments within the resource pool based on the collected resource status data and the isolated forest algorithm. The resource availability assessment subunit is used to combine real-time load, historical stability and performance bottleneck data of resources to generate a multi-dimensional availability score for each resource node and construct a real-time updated resource status matrix.

[0014] The dual-layer intelligent scheduling decision unit includes a global intelligent pre-scheduling subunit and a local real-time scheduling subunit. The global intelligent pre-scheduling subunit has a built-in multi-objective deep reinforcement learning scheduling model. It takes multi-dimensional task profiles and resource status matrices as inputs, and takes task completion rate, QoS compliance rate, global resource utilization rate, overall system energy consumption and scheduling fairness among tenants as optimization objectives. It outputs a global task queuing sequence, node-level resource pre-allocation scheme and task scheduling time window planning. The local real-time scheduling subunit is deployed locally on each resource node. It is used to dynamically adjust the execution order of tasks, core-level resource binding, and video memory allocation in real time within the globally pre-allocated resource quota based on the real-time resource status of the node, changes in task execution progress, and sudden task access events, thereby generating a real-time scheduling strategy.

[0015] The multi-objective deep reinforcement learning scheduling model of the global intelligent pre-scheduling subunit adopts a hierarchical action space design and a multi-scale reward function. The hierarchical action space includes a task allocation action layer and a resource quota action layer. The task allocation action layer is used to determine the target resource node corresponding to the task, and the resource quota action layer is used to allocate the upper and lower limits of the corresponding resource dimension quota to the task. The multi-scale reward function includes immediate rewards and long-term rewards. The immediate rewards are generated based on the QoS compliance and resource utilization efficiency of a single task, while the long-term rewards are generated based on the average resource utilization, average energy consumption, and tenant fairness index within the global scheduling cycle. The model iteratively optimizes the strategy network parameters through a temporal difference learning algorithm.

[0016] The resource optimization and conflict resolution unit includes a resource fragmentation governance subunit and a concurrent conflict resolution subunit; The resource fragmentation management subunit is used to optimize the global resource pre-allocation scheme. By resource aggregation scheduling based on task profiles, resource affinity binding of similar tasks, merging and reserving idle resource fragments, it reduces the fragmentation rate of memory, CPU cores and video memory in the resource pool and increases the proportion of continuously available resource blocks. The concurrent conflict resolution subunit is used to detect resource contention conflicts during the concurrent execution of multiple tasks in real time, including shared cache contention, IO bandwidth contention, and network port conflicts. Based on the priority level and profile tags of the tasks, it adopts time-sharing scheduling of conflicting tasks, dynamic adjustment of resource quotas, and isolation allocation of conflicting resources to complete the real-time resolution of conflicts.

[0017] The concurrent conflict resolution subunit has a built-in conflict classification processing mechanism. For detected resource contention conflicts, the conflicts are classified into three levels: minor conflict, general conflict and severe conflict according to the scope of conflict impact, task QoS level and conflict duration. For minor conflicts, a dynamic fine-tuning of resource quotas is used to resolve them; for general conflicts, a staggered execution of conflicting tasks is used to resolve them; for severe conflicts, a resource reservation for high-priority tasks and a suspension and migration of low-priority tasks are used to resolve them.

[0018] The closed-loop self-iterative optimization unit includes an execution data acquisition subunit, a scheduling effect evaluation subunit, and a model iterative optimization subunit; The execution data acquisition subunit is used to collect the task lifecycle execution data output by the scheduling execution unit, including the actual execution time of the task, the actual resource usage curve, QoS compliance status, abnormal events in task execution, and the actual operation data of the resource pool. The scheduling effect evaluation subunit is used to calculate scheduling effect evaluation indicators based on the collected execution data, including task completion rate, hard real-time task QoS compliance rate, average scheduling latency, global resource average utilization rate, resource fragmentation rate, system unit computing power energy consumption, and tenant scheduling fairness index. The model iterative optimization subunit is used to compare the scheduling effect evaluation index with the preset optimization target threshold. When the index deviation exceeds the preset range, an incremental training sample set is constructed based on the collected execution data to perform incremental training and parameter iterative update of the scheduling model of the two-layer intelligent scheduling decision unit.

[0019] The resource optimization and conflict resolution unit is also connected to a resource elastic scaling unit. The resource elastic scaling unit is used to perform elastic scaling operations on container instances and virtual machine nodes in the resource pool according to the resource status matrix and the overall resource demand trend of multi-task concurrency. At the same time, it dynamically and elastically adjusts the resource quota of tasks, so as to achieve on-demand allocation of resources and energy consumption optimization while ensuring the QoS of tasks.

[0020] The task access unit supports concurrent access of multi-tenant and multi-type tasks, including real-time computing tasks, batch offline computing tasks, AI inference tasks, AI training tasks, and database transaction processing tasks. The task multi-dimensional profiling unit configures independent tenant weights and QoS guarantee policies for tasks of different tenants. The dual-layer intelligent scheduling decision unit ensures scheduling fairness among multiple tenants based on tenant weights during the scheduling decision process.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The invention utilizes a multi-dimensional task profiling unit to achieve end-to-end processing of task static attribute extraction, dynamic resource demand prediction throughout the entire lifecycle, and three-level task classification, generating standardized multi-dimensional task profiles. This completely changes the shortcomings of traditional static quota scheduling, which suffers from "unpredictable demand." Simultaneously, through a heterogeneous resource full-state perception unit, based on distributed eBPF probes, it achieves microsecond-level full-dimensional collection of heterogeneous resources. Combined with performance bottleneck identification and multi-dimensional availability scoring, it constructs a real-time updated resource state matrix, solving the problem of "unclear resource visibility" in traditional technologies. The bidirectional precise profiling provides high-fidelity input support for scheduling decisions, enabling refined matching of task demand and resource supply. This fundamentally avoids waste caused by excessive resource reservation and QoS non-compliance caused by insufficient resource supply.

[0022] 2. This invention breaks through the limitations of traditional centralized / distributed scheduling architecture and designs a two-layer collaborative scheduling architecture of global intelligent pre-scheduling + local real-time scheduling: the global intelligent pre-scheduling subunit is based on a multi-objective deep reinforcement learning model to achieve cluster-level global multi-objective optimization scheduling and has a complete global resource and task vision; the local real-time scheduling subunit is deployed locally on the node and achieves millisecond-level dynamic scheduling adjustment within the globally pre-allocated quota to cope with sudden load and local resource fluctuations.

[0023] 3. Addressing the core pain points of multi-task concurrency scenarios, this invention designs a three-level resource fragmentation governance mechanism. Through resource aggregation scheduling, affinity binding, and idle resource merging and reservation, the cluster resource fragmentation rate is controlled within 5%, completely solving the industry problem of "having idle resources but being unable to schedule them," and significantly increasing the proportion of continuously available resources in the cluster. At the same time, a three-level resource conflict resolution mechanism is designed to achieve pre-detection of resource competition conflicts and real-time hierarchical processing during the event. The average conflict resolution latency is ≤50ms, enabling seamless conflict handling and significantly reducing issues such as soaring task execution latency and execution anomalies caused by multi-task resource contention, thus ensuring the stability of task execution in high-concurrency scenarios.

[0024] 4. To address the inherent shortcomings of reinforcement learning scheduling models, this invention employs a hierarchical action space design and a multi-scale reward function, compressing the original million-level action space to the thousand-level, significantly reducing the model's convergence difficulty and improving decision-making speed and accuracy. Simultaneously, it achieves a collaborative balance of multi-dimensional mutually exclusive objectives by simultaneously optimizing task completion rate, QoS compliance rate, resource utilization, system energy consumption, and tenant fairness.

[0025] 5. This invention designs a closed-loop self-iterative optimization unit for the entire process. By collecting all task execution data and resource operation data, a standardized scheduling effect evaluation system is constructed. It can automatically build an incremental training sample set based on indicator deviations, completing incremental training and parameter iterative updates of the scheduling model. This architecture enables the scheduling system to learn and optimize autonomously, continuously adapting to cluster hardware upgrades, load type changes, and business peak adjustments without manual intervention. This significantly reduces the operation and maintenance costs of the scheduling system and substantially improves its long-term adaptability, robustness, and lifespan.

[0026] 6. This invention adopts a modular design, compatible with mainstream heterogeneous computing resources such as CPUs, GPUs, NPUs, and FPGAs. It supports all types of task workloads, including real-time computing, offline batch processing, AI training / inference, and transaction processing. It can seamlessly connect to mainstream scheduling frameworks such as Kubernetes, YARN, and Slurm, and is adaptable to various multi-task concurrent scenarios such as public clouds, private clouds, edge computing, and high-performance computing. It also supports multi-tenant isolation and tiered QoS guarantees, meeting the scheduling needs of different industries such as finance, internet, scientific research, and industrial control. It possesses strong versatility, practicality, and promotional value.

[0027] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0028] Figure 1 This is a flowchart of this embodiment. Detailed Implementation

[0029] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0030] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0032] 1. Task Access Unit In this embodiment, the task access unit adopts a multi-protocol compatible gateway architecture, which realizes concurrent access of multi-tenant and multi-type tasks based on gRPC and RESTful dual protocol stacks. It is compatible with Kubernetes CRD, YARN job submission protocol, Slurm job submission interface, and Flink / Spark task submission specification, and can seamlessly connect to mainstream big data, AI, and real-time computing frameworks.

[0033] The core execution logic of the task access unit is as follows: Tenant authentication and permission verification: Integrates with the cluster tenant management system to complete tenant authentication, resource quota verification, and intercept illegal task submission requests; Task validity verification: Verify the integrity of the task description file, the availability of the execution image, and the compliance of the dependency configuration, and filter out invalid tasks; Task tiered queuing: Based on the priority identifier carried by the task, the verified tasks are stored in the hard real-time task priority queue, the soft real-time task queue, and the non-real-time throughput task queue respectively. The queue adopts a lock-free design and supports tens of thousands of concurrent writes per second. At the same time, the original task data is synchronously sent to the task multi-dimensional profile unit.

[0034] In this embodiment, the task access unit supports concurrent access to the following task types: real-time computing tasks (Flink / Spark Streaming stream processing tasks), batch offline computing tasks (Hadoop / Spark batch processing tasks), AI inference tasks (TensorRT / TorchServe online inference service), AI training tasks (PyTorch / TensorFlow distributed training tasks), and database OLTP transaction processing tasks, fully covering the mainstream task load types in the data center.

[0035] 2. Multi-dimensional task profiling unit The input end of the task multi-dimensional profiling unit is connected to the output end of the task access unit. It includes a task feature extraction sub-unit, a resource demand prediction sub-unit, and a profile grading sub-unit. Its core function is to generate a multi-dimensional feature profile of the entire task lifecycle, providing accurate task demand basis for scheduling decisions.

[0036] 2.1 Task Feature Extraction Subunit In this embodiment, the task feature extraction subunit adopts a dual-mode feature extraction approach of static parsing + dynamic pre-scanning: Static parsing: The description file, execution image, and dependency configuration of the task are parsed to extract the static attribute features of the task, including task type, tenant ID, global priority (level 0-9, with level 0 being the lowest and level 9 being the highest), execution deadline, maximum parallelism, upstream and downstream task dependencies, data affinity requirements (such as specifying the storage node to be accessed and the location of the dataset), and fault retry policy. Dynamic pre-scanning: For new tasks without historical execution data, a lightweight sandbox is used to perform pre-scanning, extracting the task's computational logic characteristics, memory access patterns, IO read / write characteristics, and computing power dependency types, supplementing implicit features that cannot be covered by static analysis.

[0037] The extracted static and latent features are uniformly stored in the task feature library as the basic data for resource demand prediction and profile classification.

[0038] 2.2 Resource Demand Forecasting Subunit In this embodiment, the resource demand prediction subunit has a built-in prediction model with a lightweight temporal convolutional network (TCN) and attention mechanism. The model is pre-trained based on a dataset of more than 10 million tasks throughout their entire lifecycle, with an inference latency of ≤10ms, which meets the performance requirements of real-time scheduling.

[0039] The core execution logic of the model is as follows: taking the feature extraction results of the task and the historical execution time series data of the same type of task in the same tenant as input, the TCN network captures the time series change pattern of task resource demand, and the attention mechanism strengthens the demand characteristics of key execution stages. Finally, it outputs the dynamic resource demand curve at the second level throughout the entire execution cycle of the task, including the number of CPU cores, memory capacity, GPU memory, GPU computing power ratio, storage IOPS, minimum guaranteed value of network bandwidth, peak demand value and duration. It accurately depicts the changes in resource demand throughout the entire cycle of the task from start-up, execution to end, and solves the problem of resource redundancy or insufficiency caused by static quota allocation in traditional scheduling.

[0040] 2.3 Image Classification Subunit In this embodiment, the profile classification subunit establishes a three-level profile classification rule based on the task's feature data and resource demand prediction results, and generates a unique profile ID and standardized profile label for each task. Hard real-time guarantee tasks: Priority level ≥ 7, with strict execution deadlines, QoS requirements of 100% task completion rate and end-to-end latency ≤ preset threshold. Typical scenarios include financial real-time transaction risk control, autonomous driving edge inference, and industrial real-time control tasks. The profile is tagged as "Hard Real-time - High Guarantee". Soft real-time assurance tasks: priority level 3-6, with execution deadline but allowing minor timeouts of ≤100ms, QoS compliance rate requirement ≥99.9%, typical scenarios include video transcoding, real-time monitoring data analysis, online recommendation inference tasks, and the profile tag is marked as

Soft Real-time - Medium Assurance

[0041] The generated multi-dimensional task profiles are synchronized in real time to the two-layer intelligent scheduling decision unit and stored in the profile database for subsequent model iteration and demand prediction optimization. For tasks from different tenants, the profile hierarchical sub-unit is configured with independent tenant weights and QoS guarantee policies to support fair scheduling for multiple tenants.

[0042] 3. Heterogeneous resource full-state sensing unit The input end of the heterogeneous resource full-state perception unit is connected to the output end of the resource pool management unit. It includes a resource status acquisition subunit, a performance bottleneck identification subunit, and a resource availability assessment subunit. Its core function is to realize full-dimensional, low-overhead, and real-time status perception of the heterogeneous resource pool, providing accurate resource supply basis for scheduling decisions.

[0043] 3.1 Resource Status Acquisition Subunit In this embodiment, the resource status acquisition subunit adopts a distributed eBPF probe architecture. A lightweight acquisition probe is deployed on each computing node in the cluster. The acquisition period is configurable, with a default acquisition level of microseconds (100μs). The probe CPU utilization rate is ≤1% of the total CPU of the node. It is a non-intrusive acquisition method that does not require modification of task code or operating system kernel.

[0044] The core data dimensions collected include: CPU resources: User-mode / kernel-mode utilization per physical core, context switch count, L3 cache hit rate, NUMA node topology and access latency; Memory resources: used / available capacity, memory bandwidth usage, large page memory usage, memory fragmentation rate, and swap partition usage; Heterogeneous computing resources: GPU CUDA core utilization, memory usage, NVLink bandwidth, temperature and power consumption; FPGA logic resource utilization and computing throughput; Storage resources: disk IOPS, read / write bandwidth, average access latency, and free storage space; Network resources: network card inbound / outbound bandwidth, number of TCP connections, packet loss rate, network round-trip time (RTT); Overall node data: total power consumption, runtime, and hardware fault alarm information.

[0045] The collected real-time data is reported to the control plane in real time through a high-throughput message queue (Kafka), stored in a time-series database, and simultaneously synchronized to the performance bottleneck identification subunit.

[0046] 3.2 Performance Bottleneck Identification Subunit In this embodiment, the performance bottleneck identification subunit incorporates an isolated forest anomaly detection algorithm. The algorithm is pre-trained based on a historical resource status anomaly dataset of the cluster, with an identification latency of ≤20ms, and can identify anomalies and performance bottlenecks within the resource pool in real time. Overload node identification: Identify overload nodes with CPU / memory / GPU utilization consistently ≥90% for more than 500ms, as well as nodes that have reached a bottleneck in a single resource dimension (such as IO bottleneck nodes with storage IO latency ≥100ms). Idle resource fragment identification: Identify discrete, unused idle resource fragments within the cluster, including fragmented CPU cores, memory pages, and GPU memory blocks, and calculate the total amount and distribution of fragmented resources; Abnormal Node Identification: Identify abnormal nodes with hardware failures, network jitter, or sudden performance drops, mark them as unschedulable nodes, and prevent tasks from being scheduled to abnormal nodes, which could lead to execution failure.

[0047] Once bottleneck data and anomaly markers are identified, they are synchronized to the resource availability assessment subunit in real time.

[0048] 3.3 Resource Availability Assessment Subunit In this embodiment, the resource availability assessment subunit adopts a multi-dimensional weighted scoring model to generate an availability score of 0-100 for each resource node. The scoring dimensions and weights are: real-time load status (30%), historical operational stability (fault-free operation time in the past 7 days, 25%), performance bottleneck status (20%), energy efficiency (15%), and network connectivity (10%). The higher the score, the higher the availability and scheduling priority of the node.

[0049] Based on node availability scores and real-time resource status data, the sub-unit constructs a real-time updated resource status matrix. The rows of the matrix correspond to the resource node IDs within the cluster, and the columns correspond to the real-time values, available resources, availability scores, and anomaly markers for each resource dimension. The matrix's update frequency is synchronized with the data collection cycle, defaulting to once every 100μs, ensuring that the resource data used for scheduling decisions is completely consistent with the actual cluster status. The resource status matrix is ​​synchronized in real-time to the two-layer intelligent scheduling decision unit.

[0050] 4. Two-layer intelligent scheduling and decision-making unit The dual-layer intelligent scheduling decision unit is the core decision unit of this device. The first input end is connected to the output end of the task multi-dimensional profiling unit, and the second input end is connected to the output end of the heterogeneous resource full-state perception unit. It includes a global intelligent pre-scheduling subunit and a local real-time scheduling subunit. The core is used to realize dual-layer collaborative scheduling of global multi-objective optimization and local low-latency response, solving the core pain points of "high latency of global scheduling and no global view of local scheduling" in traditional scheduling architecture.

[0051] 4.1 Global Intelligent Pre-Scheduling Subunit In this embodiment, the global intelligent pre-scheduling subunit incorporates a multi-objective dual-delay deep Q-network (D3QN) deep reinforcement learning scheduling model, which is deployed in the control plane with a scheduling period of 10ms. Its core function is to realize cluster-level global task scheduling and resource pre-allocation.

[0052] The model's inputs are: batch multi-dimensional task profiles, real-time resource status matrix, tenant weight configuration, and cluster scheduling optimization target threshold; the model's outputs are: global task queuing execution sequence, target resource nodes for each task, node-level resource pre-allocation scheme (minimum guaranteed quota and maximum usage quota for each resource dimension), and task scheduling time window planning.

[0053] The core optimization objectives of the model are five interrelated multi-objective optimizations: maximizing task completion rate, maximizing hard real-time task QoS compliance rate, maximizing global resource utilization, minimizing overall system energy consumption, and maximizing scheduling fairness among multi-tenants (measured by the Jain fairness index).

[0054] To address the issues of action space explosion and high convergence difficulty in multi-objective reinforcement learning, this embodiment performs targeted optimizations on the model, corresponding to the core features of claim 5: Layered Action Space Design: The model's action space is divided into two layers. The first layer is the task allocation action layer, where the action space consists of available resource node IDs within the cluster and the target scheduling node corresponding to the core decision task. The second layer is the resource quota action layer, where the action space consists of standardized quota levels for each resource dimension (CPU in 0.5-core increments, memory in 1GB increments, GPU memory in 512MB increments), and the upper and lower limits of resource quotas for the core decision task. This layered design compresses the original million-level action space to the thousand-level, significantly reducing the model's convergence difficulty and improving decision-making speed.

[0055] Multi-scale reward function design: A dual-scale reward mechanism combining immediate and long-term rewards is set up, wherein: Instant reward r_t: Instant feedback after a single task is scheduled. Positive rewards are given if the task QoS is guaranteed and the resource allocation redundancy is ≤30%. Negative rewards are given if there are resource allocation conflicts or QoS cannot be guaranteed. High negative penalties are given if the QoS of a hard real-time task does not meet the standard. Long-term reward R: Global feedback after a complete scheduling cycle (10s) ends. It is calculated based on global scheduling indicators. Positive rewards are given for every 1% increase in global resource utilization, every 1% decrease in unit computing power energy consumption, and every 0.01 increase in fairness index. High negative penalties are given for task completion rate and QoS compliance rate that are lower than the threshold.

[0056] During the model pre-training phase, 1 million rounds of iterative training were completed based on the cluster simulation scheduling dataset. After convergence, the model was deployed online. Subsequently, incremental training and parameter updates were completed through a closed-loop self-iterative optimization unit to continuously adapt to changes in cluster load.

[0057] 4.2 Local Real-Time Scheduling Subunit In this embodiment, the local real-time scheduling subunit is deployed on the local user-space agent of each computing node in the cluster. The scheduling latency is ≤1ms. Its core function is to realize real-time dynamic scheduling within the node within the globally pre-allocated resource quota.

[0058] The inputs to the sub-unit are: globally pre-allocated node resource quotas, real-time node resource status, task execution progress feedback, and sudden task access events; the outputs are: task execution queue sorting within the node, CPU core and NUMA node binding strategy, GPU memory sharding and computing power isolation configuration, and IO / network bandwidth limiting strategy.

[0059] The core execution logic is as follows: when a sudden high-priority task, resource demand fluctuation, or local resource conflict occurs within a node, there is no need to report to the global scheduling unit. Instead, the execution order and resource binding of the task are dynamically adjusted in real time within the pre-allocated global resource quota. This prioritizes the resource supply and QoS requirements of hard real-time tasks, resolves the latency bottleneck of global scheduling, and copes with sudden load fluctuations.

[0060] 5. Resource Optimization and Conflict Resolution Unit The input end of the resource optimization and conflict resolution unit is connected to the output end of the two-layer intelligent scheduling decision unit, and the output end is connected to the input end of the scheduling execution unit. It includes a resource fragmentation governance subunit and a concurrent conflict resolution subunit. Its core function is to optimize resource allocation schemes, resolve resource competition conflicts of multiple concurrent tasks in real time, and improve resource utilization efficiency and task execution stability.

[0061] 5.1 Resource Fragmentation Governance Subunit In this embodiment, the resource fragmentation management subunit performs three levels of fragmentation management optimization on the global resource pre-allocation scheme to control the cluster resource fragmentation rate to within 5%. Resource aggregation and scheduling optimization: Tasks of the same type, with similar resource requirements and overlapping execution time windows are centrally scheduled to the same batch of resource nodes to avoid resource fragmentation caused by the scattered deployment of small tasks, while improving the utilization rate of node resources. Affinity binding optimization: Bind tasks with data dependencies and computing power dependencies to the same NUMA node, the same GPU device, and the same storage node to reduce data transfer overhead across nodes and across NUMA nodes, while merging discrete idle resource fragments to generate continuously available resource blocks; Idle resource merging and reservation: By using memory consolidation, CPU core isolation, and GPU memory sharding and merging, discrete idle resources within a node are merged. At the same time, fixed dedicated resource blocks are reserved for hard real-time tasks to avoid resource shortages caused by sudden task occupancy.

[0062] 5.2 Concurrent Conflict Resolution Subunit In this embodiment, the concurrent conflict resolution subunit has a built-in real-time conflict detection engine and a hierarchical resolution mechanism. The resource contention conflicts that can be detected in real time include: CPU shared L3 cache contention, storage I / O bandwidth contention, network port conflicts, GPU computing power / video memory contention, and NUMA node cross-domain access conflicts.

[0063] The subunit adopts a three-level conflict classification and handling mechanism, corresponding to the core feature of claim 7. Based on the scope of the conflict's impact, task level, and duration, the conflict is divided into three levels and corresponding resolution strategies are executed: Minor conflicts: The impact of a conflict is limited to ≤2 tasks within a single node, with no hard real-time tasks. The expected duration of the conflict is ≤100ms. It is resolved by dynamically adjusting resource quotas and allocating IO / network bandwidth quotas for conflicting tasks in a time-sharing manner to smooth out peak resource demand. General conflicts: The scope of the conflict is a single node, involving ≥3 tasks or including soft real-time tasks. The conflict is expected to last for 100ms-1s. It is resolved by using time-sharing and staggered execution of conflicting tasks, adjusting the execution time window of the tasks, and avoiding simultaneous contention for the same bottleneck resource. Severe conflict: The conflict affects multiple nodes, involves hard real-time tasks, or the conflict is expected to last for ≥1 second. It is resolved by reserving resources for high-priority tasks and suspending and migrating low-priority tasks. Dedicated resources are immediately locked for hard real-time tasks, and low-priority non-real-time tasks are suspended and migrated to other idle nodes to ensure that the QoS of hard real-time tasks is not affected.

[0064] In this embodiment, the average latency of conflict detection and resolution is ≤50ms, which enables seamless conflict handling and avoids spikes in task execution latency or execution failures caused by resource contention.

[0065] 6. Resource Elastic Scaling Unit In this embodiment, the resource optimization and conflict resolution unit is also connected to a resource elastic scaling unit, which corresponds to the core feature of claim 9. The core is used to realize the on-demand elastic scaling of cluster resources, taking into account both task QoS guarantee and energy consumption optimization.

[0066] The resource elastic scaling unit incorporates an ARIMA time-series prediction model. Based on the resource state matrix and the overall resource demand trend of multi-task concurrency, it predicts the overall cluster resource demand within the next 5 minutes and executes corresponding elastic operations. Elastic scaling: When the predicted future resource demand exceeds 80% of the current available resources of the cluster, scaling is automatically triggered. New virtual machine / container instances are created through the IaaS interface, and after the environment initialization is completed, they are added to the resource pool to supplement the available resources of the cluster. Elastic scaling down: When the predicted future resource demand is less than 30% of the current available resources of the cluster, and the duration is ≥10 minutes, scaling down is automatically triggered. After migrating the tasks on idle nodes to other nodes, the idle nodes are released, reducing the overall energy consumption of the cluster. Dynamic quota adjustment: For a single task, when it is detected that the actual resource demand of the task is consistently lower than 50% of the allocated quota and the duration is ≥1 minute, the resource quota of the task will be automatically reduced to release idle resources to other tasks and improve resource utilization.

[0067] 7. Scheduling Execution Unit The input end of the scheduling execution unit is connected to the output end of the resource optimization and conflict resolution unit. It is deployed locally on each computing node and is used to execute the final resource scheduling instructions to complete the full life cycle management of the task.

[0068] In this embodiment, the scheduling execution unit implements resource isolation and management based on Linux Cgroup and Namespace technologies. The core execution logic is as follows: Receive resource scheduling instructions and complete the creation of containers / virtual machines, image retrieval, and loading of dependency configurations for the task; According to the scheduling instructions, the CPU core binding, memory quota limit, GPU memory / computing power isolation, and IO / network bandwidth rate limiting of the task are completed to achieve precise resource management; Initiate task execution, monitor the task's execution status and resource usage in real time, and report task execution progress and abnormal events to the closed-loop self-iterative optimization unit; After the task is completed, clean up the task's runtime environment, release the occupied resources, and update the resource pool status.

[0069] 8. Closed-loop self-iterative optimization unit The input of the closed-loop self-iterative optimization unit is connected to the output of the scheduling execution unit, and the output is connected to the optimization input of the two-layer intelligent scheduling decision unit. It includes an execution data acquisition subunit, a scheduling effect evaluation subunit, and a model iterative optimization subunit. The core is used to realize the autonomous learning and closed-loop iterative optimization of the scheduling model, which can continuously adapt to changes in cluster load without manual intervention.

[0070] 8.1 Execute the data acquisition subunit In this embodiment, the execution data acquisition subunit collects all lifecycle execution data of the task and cluster resource operation data through the feedback interface of the scheduling execution unit. This includes: task submission time, scheduling completion time, start execution time, end time, actual execution duration, actual resource usage curve every 100ms, QoS compliance status, timeout / abnormal interruption events, and resource utilization, energy consumption data, and fragmentation rate data of each node in the cluster. All collected data is uniformly stored in the InfluxDB time-series database as the basis for evaluation and model optimization.

[0071] 8.2 Scheduling Effectiveness Evaluation Subunit In this embodiment, the scheduling effect evaluation subunit calculates a complete scheduling effect evaluation index system with a 10-second evaluation cycle. The core indicators include: Task completion rate: Number of tasks completed within the period / Total number of tasks submitted; Hard real-time task QoS compliance rate: Number of hard real-time tasks completed within the deadline of the period / Total number of hard real-time tasks; Average scheduling latency: The average time from task submission to completion of scheduling; Global resource average utilization: a weighted average of CPU, memory, and GPU utilization; Resource fragmentation rate: The proportion of discrete idle resources to total idle resources within a cluster; Energy consumption per unit of computing power: Total power consumption of the cluster / Total computing power output value; Tenant scheduling fairness index: Jain fairness index for multi-tenant resource allocation.

[0072] The calculated evaluation metrics are synchronized in real time to the model iteration and optimization sub-unit, and simultaneously visualized on the cluster management platform.

[0073] 8.3 Model Iterative Optimization Subunit In this embodiment, the core execution logic of the model iterative optimization subunit is as follows: The scheduling performance evaluation indicators within the period are compared with the preset optimization target thresholds (hard real-time task QoS compliance rate ≥ 99.99%, global average resource utilization rate ≥ 80%, resource fragmentation rate ≤ 5%). When the deviation of the core indicators exceeds the preset range and continues for more than 2 evaluation periods, the most recent scheduling execution data is extracted from the time series database to construct an incremental training sample set. The sample set includes: the input state during scheduling (task profile, resource state matrix), the executed scheduling actions, and the final reward value. Based on the incremental training sample set, the deep reinforcement learning model of the global intelligent pre-scheduling subunit is incrementally trained, and the parameters of the policy network are iteratively updated. The updated model is then subjected to gray-scale verification. Once the verification is successful, the online production model is replaced, completing a closed-loop iterative optimization and continuously improving the decision-making performance of the scheduling model.

[0074] Complete operation steps of the device The complete operation process of the device described in this embodiment includes the following steps: Task access and verification: Multi-tenants submit tasks concurrently through the task access unit. The task access unit completes identity authentication, permission verification and task legality verification, stores the task in the corresponding priority queue, and sends it synchronously to the task multi-dimensional profile unit. Multi-dimensional task profile generation: The multi-dimensional task profile unit extracts full-dimensional features of the task, predicts dynamic resource demand curves, completes task classification, generates standardized multi-dimensional task profiles, and sends them to the two-layer intelligent scheduling decision unit. Heterogeneous resource status awareness: The heterogeneous resource full-state awareness unit collects full-dimensional status data of heterogeneous resources in the cluster in real time, identifies performance bottlenecks, evaluates node availability, generates a real-time updated resource status matrix, and sends it to the two-layer intelligent scheduling decision unit. Two-layer intelligent scheduling decision-making: The global intelligent pre-scheduling subunit performs global multi-objective optimization scheduling based on task profiles and resource status matrices, generating node-level resource pre-allocation schemes; the local real-time scheduling subunit generates local real-time scheduling strategies within the quota based on the real-time status of nodes. Resource optimization and conflict resolution: The resource optimization and conflict resolution unit performs fragmentation management and optimization on the pre-allocation scheme, while detecting and pre-resolving potential resource competition conflicts, generating the final executable resource scheduling instructions, and sending them to the scheduling execution unit; Task scheduling and execution control: The scheduling and execution unit receives scheduling instructions, completes task deployment, resource isolation and binding, starts task execution, and monitors task execution status and resource usage in real time; Closed-loop self-iterative optimization: The closed-loop self-iterative optimization unit collects task execution data and resource operation data, calculates scheduling effect evaluation indicators, completes incremental training and parameter iteration of the scheduling model based on indicator deviation, and optimizes subsequent scheduling decisions; Resource elastic scaling: Based on resource demand trend prediction and real-time resource status, the resource elastic scaling unit performs elastic scaling of cluster resources to achieve on-demand resource allocation and energy consumption optimization.

[0075] Specific application scenarios and examples This embodiment is applied to a heterogeneous computing cluster of a public cloud provider. The cluster contains 200 computing nodes, including 100 general-purpose CPU computing nodes (2-way Intel Xeon 8375C CPU, 512G memory), 50 GPU training nodes (8-card NVIDIA A100 GPU, 1TB memory), and 50 GPU inference nodes (4-card NVIDIA T4 GPU, 256G memory). The cluster is simultaneously connected to 10 enterprise tenants, and the concurrently submitted tasks cover four major categories: real-time financial risk control inference, internet video transcoding, scientific research AI model training, and enterprise offline big data analysis.

[0076] The specific operation process is as follows: 1. A financial tenant (tenant weight 3) submits a real-time transaction risk control AI inference task with 100 concurrent connections. The task has a priority level of 8 and is a hard real-time guarantee task with an execution deadline of 50ms. After the task access unit completes the authentication and verification, the task is stored in the hard real-time priority queue and sent to the task multi-dimensional profile unit simultaneously. 2. The multi-dimensional task profiling unit extracts task features and, based on the tenant's historical risk control task execution data, predicts the resource requirements of each inference task as 0.5 core CPU, 2G memory, 1G GPU memory, and 5% GPU computing power, generating a hard real-time guarantee task profile and synchronizing it to the dual-layer intelligent scheduling decision unit. 3. The heterogeneous resource full-state perception unit collects the status of cluster nodes in real time through eBPF probes, generates a resource status matrix, and identifies that the availability score of 4 GPU inference nodes is ≥90 points, indicating that they have sufficient available video memory and computing power. 4. The DRL model of the global intelligent pre-scheduling subunit, based on task profile and resource status matrix, evenly distributes 100 inference tasks to 4 highly available GPU inference nodes. Each node is allocated 25 tasks, and each task is pre-allocated with a quota of 0.5 CPU cores, 2G memory, 1G video memory, and 5% computing power. At the same time, 10% redundant resources are reserved for each node to cope with sudden traffic peaks. 5. The local real-time scheduling subunit binds tasks to the corresponding NUMA node's CPU core and GPUSM stream processor locally on each GPU node, sets memory isolation and computing power limiting to ensure stable inference latency for each task. 6. The resource optimization and conflict resolution unit performs affinity optimization on the allocation scheme, binds tasks of the same tenant to the same NUMA domain, reduces data access latency, and pre-detects IO bandwidth contention within the node. It then resolves the conflict through time-sharing scheduling and generates the final scheduling instruction. 7. The scheduling and execution unit receives instructions, starts inference tasks, monitors task execution status in real time, and collects data showing that the actual inference latency of each task is stable within 30ms, fully meeting the 50ms deadline requirement, with a QoS compliance rate of 100%. 8. The closed-loop self-iterative optimization unit collects all execution data from this scheduling, calculates that the QoS compliance rate of the hard real-time task is 100%, and the GPU node resource utilization rate is increased from 40% in the traditional scheduling scheme to 75%. The scheduling data is used to construct incremental training samples to complete the parameter optimization of the DRL model. 9. During off-peak hours in the evening, if the resource elastic scaling unit detects that the overall resource utilization of the cluster has been below 30% for more than 10 minutes, it will automatically release 10 idle GPU inference nodes, reducing the overall energy consumption of the cluster by 20%.

[0077] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A computer resource intelligent scheduling and optimization device for multi-task concurrency, comprising a task access unit, a resource pool management unit, and a scheduling execution unit, characterized in that, It also includes a task multi-dimensional profiling unit, a heterogeneous resource full-state perception unit, a two-layer intelligent scheduling and decision-making unit, a resource optimization and conflict resolution unit, and a closed-loop self-iterative optimization unit; The input end of the task multi-dimensional profiling unit is connected to the output end of the task access unit, and is used to extract full life cycle features of concurrently accessed scheduled tasks to generate a multi-dimensional task profile that includes task static attributes, dynamic resource requirements, execution timing constraints, QoS level requirements and computing power dependence features. The input end of the heterogeneous resource full-state perception unit is connected to the output end of the resource pool management unit, and is used to collect the real-time full-dimensional state of the computer heterogeneous resource pool, and generate a resource state matrix that includes the real-time load, available margin, performance bottleneck and energy consumption characteristics of CPU, memory, GPU, storage IO, network bandwidth resources. The first input of the dual-layer intelligent scheduling decision unit is connected to the output of the task multi-dimensional profiling unit, and the second input is connected to the output of the heterogeneous resource full-state perception unit. It is used to perform dual-layer scheduling decisions of global pre-scheduling and local real-time scheduling based on multi-dimensional task profiling and resource state matrix, and generate initial resource allocation scheme and real-time scheduling strategy. The input end of the resource optimization and conflict resolution unit is connected to the output end of the dual-layer intelligent scheduling decision unit, and the output end is connected to the input end of the scheduling execution unit. It is used to perform fragmentation management and optimization on the initial resource allocation scheme, and at the same time to detect and resolve resource competition conflicts in the process of concurrent execution of multiple tasks in real time, and generate the final executable resource scheduling instructions. The input of the closed-loop self-iterative optimization unit is connected to the output of the scheduling execution unit, and the output is connected to the optimization input of the dual-layer intelligent scheduling decision unit. It is used to collect actual task execution data and resource operation data, generate scheduling effect evaluation indicators, and iteratively optimize the scheduling model parameters of the dual-layer intelligent scheduling decision unit based on the evaluation indicators.

2. The intelligent scheduling and optimization device for computer resources oriented towards multi-task concurrency as described in claim 1, characterized in that, The multi-dimensional task profiling unit includes a task feature extraction subunit, a resource demand prediction subunit, and a profiling classification subunit. The task feature extraction subunit is used to extract the static attribute features of the task to be scheduled, including task type, tenant, priority, execution deadline, parallelizability and dependency. The resource demand prediction subunit is used to predict the dynamic CPU, memory, GPU, IO and network bandwidth resource demand curves of a task throughout its entire execution cycle based on the task's historical execution data and the computational logic characteristics of the current task, using a lightweight time-series prediction model. The profile classification subunit is used to classify tasks into three categories—hard real-time guarantee tasks, soft real-time guarantee tasks, and non-real-time throughput tasks—based on the task's QoS level requirements, timing constraint strictness, and resource demand characteristics, and to generate corresponding profile tags for each category of tasks.

3. The intelligent scheduling and optimization device for computer resources oriented towards multi-task concurrency as described in claim 1, characterized in that, The heterogeneous resource full-state perception unit includes a resource state acquisition subunit, a performance bottleneck identification subunit, and a resource availability assessment subunit. The resource status acquisition subunit adopts a distributed probe architecture and collects data on CPU core utilization, memory usage, GPU computing power utilization, video memory usage, storage IOPS, network throughput and node energy consumption of each node in the computer heterogeneous resource pool in real time with a preset microsecond-level acquisition cycle. The performance bottleneck identification subunit is used to identify performance bottleneck nodes, overloaded resource dimensions, and idle resource fragments within the resource pool based on the collected resource status data and the isolated forest algorithm. The resource availability assessment subunit is used to combine real-time load, historical stability and performance bottleneck data of resources to generate a multi-dimensional availability score for each resource node and construct a real-time updated resource status matrix.

4. The intelligent scheduling and optimization device for computer resources oriented towards multi-task concurrency as described in claim 1, characterized in that, The dual-layer intelligent scheduling decision unit includes a global intelligent pre-scheduling subunit and a local real-time scheduling subunit. The global intelligent pre-scheduling subunit has a built-in multi-objective deep reinforcement learning scheduling model. It takes multi-dimensional task profiles and resource status matrices as inputs, and takes task completion rate, QoS compliance rate, global resource utilization rate, overall system energy consumption and scheduling fairness among tenants as optimization objectives. It outputs a global task queuing sequence, node-level resource pre-allocation scheme and task scheduling time window planning. The local real-time scheduling subunit is deployed locally on each resource node. It is used to dynamically adjust the execution order of tasks, core-level resource binding, and video memory allocation in real time within the globally pre-allocated resource quota based on the real-time resource status of the node, changes in task execution progress, and sudden task access events, thereby generating a real-time scheduling strategy.

5. The intelligent scheduling and optimization device for computer resources oriented towards multi-task concurrency according to claim 4, characterized in that, The multi-objective deep reinforcement learning scheduling model of the global intelligent pre-scheduling subunit adopts a hierarchical action space design and a multi-scale reward function. The hierarchical action space includes a task allocation action layer and a resource quota action layer. The task allocation action layer is used to determine the target resource node corresponding to the task, and the resource quota action layer is used to allocate the upper and lower limits of the corresponding resource dimension quota to the task. The multi-scale reward function includes immediate rewards and long-term rewards. The immediate rewards are generated based on the QoS compliance and resource utilization efficiency of a single task, while the long-term rewards are generated based on the average resource utilization, average energy consumption, and tenant fairness index within the global scheduling cycle. The model iteratively optimizes the strategy network parameters through a temporal difference learning algorithm.

6. The intelligent scheduling and optimization device for computer resources oriented towards multi-task concurrency as described in claim 1, characterized in that, The resource optimization and conflict resolution unit includes a resource fragmentation governance subunit and a concurrent conflict resolution subunit; The resource fragmentation management subunit is used to optimize the global resource pre-allocation scheme. By resource aggregation scheduling based on task profiles, resource affinity binding of similar tasks, merging and reserving idle resource fragments, it reduces the fragmentation rate of memory, CPU cores and video memory in the resource pool and increases the proportion of continuously available resource blocks. The concurrent conflict resolution subunit is used to detect resource contention conflicts during the concurrent execution of multiple tasks in real time, including shared cache contention, IO bandwidth contention, and network port conflicts. Based on the priority level and profile tags of the tasks, it adopts time-sharing scheduling of conflicting tasks, dynamic adjustment of resource quotas, and isolation allocation of conflicting resources to complete the real-time resolution of conflicts.

7. The intelligent scheduling and optimization device for computer resources oriented towards multi-task concurrency as described in claim 6, characterized in that, The concurrent conflict resolution subunit has a built-in conflict classification processing mechanism. For detected resource contention conflicts, the conflicts are classified into three levels: minor conflict, general conflict and severe conflict according to the scope of conflict impact, task QoS level and conflict duration. For minor conflicts, a dynamic fine-tuning of resource quotas is used to resolve them; For general conflicts, the conflict tasks are resolved by staggering their execution times; for severe conflicts, the high-priority tasks are reserved for resources and the low-priority tasks are suspended and migrated.

8. The intelligent scheduling and optimization device for computer resources oriented towards multi-task concurrency as described in claim 1, characterized in that, The closed-loop self-iterative optimization unit includes an execution data acquisition subunit, a scheduling effect evaluation subunit, and a model iterative optimization subunit; The execution data acquisition subunit is used to collect the task lifecycle execution data output by the scheduling execution unit, including the actual execution time of the task, the actual resource usage curve, QoS compliance status, abnormal events in task execution, and the actual operation data of the resource pool. The scheduling effect evaluation subunit is used to calculate scheduling effect evaluation indicators based on the collected execution data, including task completion rate, hard real-time task QoS compliance rate, average scheduling latency, global resource average utilization rate, resource fragmentation rate, system unit computing power energy consumption, and tenant scheduling fairness index. The model iterative optimization subunit is used to compare the scheduling effect evaluation index with the preset optimization target threshold. When the index deviation exceeds the preset range, an incremental training sample set is constructed based on the collected execution data to perform incremental training and parameter iterative update of the scheduling model of the two-layer intelligent scheduling decision unit.

9. The intelligent scheduling and optimization device for computer resources oriented towards multi-task concurrency according to claim 1, characterized in that, The resource optimization and conflict resolution unit is also connected to a resource elastic scaling unit. The resource elastic scaling unit is used to perform elastic scaling operations on container instances and virtual machine nodes in the resource pool according to the resource status matrix and the overall resource demand trend of multi-task concurrency. At the same time, it dynamically and elastically adjusts the resource quota of tasks, so as to achieve on-demand allocation of resources and energy consumption optimization while ensuring the QoS of tasks.

10. The intelligent scheduling and optimization device for computer resources oriented towards multi-task concurrency according to claim 1, characterized in that, The task access unit supports concurrent access of multi-tenant and multi-type tasks, including real-time computing tasks, batch offline computing tasks, AI inference tasks, AI training tasks, and database transaction processing tasks. The task multi-dimensional profiling unit configures independent tenant weights and QoS guarantee policies for tasks of different tenants. The dual-layer intelligent scheduling decision unit ensures scheduling fairness among multiple tenants based on tenant weights during the scheduling decision process.

Images