SLURM-based quantum classical hybrid computing task dynamic scheduling system and method

Abstract

The invention discloses a quantum classical hybrid computing task dynamic scheduling system and method based on SLURM. The system comprises a quantum task feature extraction module, a dynamic priority evaluation module, a dependency analysis module, a quantum perception backfilling module and a uniform resource abstraction layer module. The method comprises the following steps: extracting quantum features of a to-be-processed task and a computing resource to form a quantum feature set and caching the quantum feature set; the feature set is obtained in real time, task priorities are output through multi-dimensional evaluation, and a real-time priority sequence is generated; based on the task type and the feature set, forming a dependency relationship between the classic task and the quantum task, and converting the dependency relationship into a dependency constraint and / or resource reservation instruction; according to the task priority, the resource reservation instruction and the real-time resource state, a future idle period is predicted, and a short-time quantum task is inserted for backfilling; computing resources are distributed according to the task priority and the dependency constraint, and the resource utilization state is fed back to the backfill and priority evaluation module in real time. According to the method, efficient scheduling of hybrid computing tasks can be realized, and the overall performance of the system is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of computer scheduling technology, specifically relating to a dynamic scheduling system and method for quantum-classical hybrid computing tasks based on SLURM. Background Technology

[0002] With the rapid development of quantum computing technology and the expansion of classical computing clusters, quantum-classical hybrid computing environments face severe scheduling challenges. The main reason is the difference in characteristics between classical and quantum computing. For example, quantum computing tasks have strict time constraints and bursty characteristics, while classical computing tasks exhibit stable and continuous resource requirements. The two have fundamental differences in computing modes. At the same time, the existing SLURM (Simple Linux Resource Management Tool) scheduling system cannot effectively identify the urgency of quantum tasks, resulting in low utilization of quantum devices. Furthermore, traditional scheduling strategies cannot achieve efficient coordination between quantum and classical tasks.

[0003] Although related technologies disclose quantum-classical hybrid computing task scheduling methods and quantum-classical hybrid architectures, which use static machine learning prediction models for task scheduling, they still cannot adapt to the dynamic decay characteristics of coherence time of quantum tasks in real time. Furthermore, the separate task pools lead to low efficiency in the collaboration between quantum and classical tasks. Therefore, there is an urgent need for a SLURM-based quantum-classical hybrid computing task dynamic scheduling system and method to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to propose a dynamic scheduling system and method for quantum-classical hybrid computing tasks based on SLURM, in order to solve the problems in the prior art.

[0005] Therefore, this invention provides a dynamic scheduling system for quantum-classical hybrid computing tasks based on SLURM, comprising:

[0006] The quantum task feature extraction module is used to extract quantum features from the task to be processed containing quantum features to obtain the first quantum feature, and to extract quantum features from the computing resources to obtain the second quantum feature, and to cache the extracted first and second quantum features as features of the quantum feature set.

[0007] The dynamic priority evaluation module is used to obtain the quantum feature set from the cache in real time, and output the corresponding priority for the quantum task to be processed through multi-dimensional evaluation processing. It also adjusts the task processing priority in real time based on the update mechanism and outputs a real-time updated task priority sequence.

[0008] The dependency resolution module, based on the type of the task to be processed and the quantum feature set, forms the dependency relationship between the classical task and the quantum task, and forms the dependency relationship into the dependency constraint of the corresponding task to be processed, and / or forms the resource reservation instruction of the corresponding task to be processed.

[0009] The quantum sensing backfilling module, based on the priority of the task to be processed, the resource reservation instruction, and the real-time feedback of the computing resource status, predicts the idle time periods that will appear in the future, and inserts the eligible short-term quantum tasks into the computing resources of the idle time periods for task backfilling.

[0010] The unified resource abstraction layer module allocates corresponding computing resources based on the priority and dependency constraints of the task to be processed, and provides real-time feedback on the utilization status of computing resources to the quantum sensing backfill module and the dynamic priority evaluation module.

[0011] In one embodiment, the quantum task feature extraction module includes a parameter resolver module. The parameter resolver module creates a quantum parameter context structure for a task to be processed that contains quantum-specific parameters. The quantum parameter upper and lower structures include a base layer and an extension layer. The base layer is used to store standard SLURM parameters, and the extension layer is used to store quantum property parameters.

[0012] In one embodiment, the quantum task feature extraction module includes a legitimacy verifier module, which is used to verify the legitimacy of the extracted first quantum feature in terms of quantum computing type, resource requirements, and time feasibility.

[0013] In one embodiment, the quantum task feature extraction module includes a metadata conversion module, which is used to convert the first quantum feature in text format into a binary structure and write the converted binary structure into an extended job description structure.

[0014] In one embodiment, the dynamic priority evaluation module includes a task coherence time constraint decay module. The task coherence time constraint decay module calculates the remaining value of task coherence time by taking the initial coherence time window of the quantum task, the waiting time after task submission, the effective coherence time constant of the task determined by the quantum device and circuit depth factor, and the additional decay rate caused by environmental noise, so as to quantify the time urgency of the quantum task.

[0015] In one embodiment, the dynamic priority evaluation module includes a task importance evaluation module, which calculates the importance of the quantum task by considering user level, quantum gate operation fidelity requirements, the proportion of bits occupied by the task, and weighting, in order to quantify the importance of the quantum task.

[0016] In one embodiment, the final priority of the dynamic priority evaluation module is calculated by taking the remaining value of the task coherence time, the importance of the quantum task, and the initial coherence time window, and the priority matrix is ​​updated based on the final priority.

[0017] In one embodiment, the dynamic priority evaluation module employs a real-time update mechanism, which includes an event-driven update mode and a periodic scan update mode.

[0018] In one embodiment, the dynamic priority evaluation module includes a priority matrix management module. The priority matrix management module maintains a task queue sorted by final priority using a red-black tree to quickly obtain the highest priority task, uses read-write locks and segmented locks to handle concurrent access, and combines time window merging to partition and process update requests in parallel and merge the red-black tree for rebalancing.

[0019] In one embodiment, the dependency resolution module includes a directed acyclic graph (DAG) modeling module, which abstracts task dependencies into a weighted topological graph structure.

[0020] In one embodiment, the dependency resolution module includes a collaborative reservation module, which employs a multi-dimensional conflict detection and three-level resolution strategy based on a quantum resource reservation table structure, and implements distributed resource collaboration based on a three-phase commit protocol.

[0021] In one embodiment, the quantum sensing backfill module includes a first GRES quantum extension module, which is used to add a quantum processing unit to support the definition of key parameters of the quantum device.

[0022] In one embodiment, the quantum sensing backfill module includes a window prediction module, which calculates the basic available window by analyzing the historical calibration logs of the quantum device and combining them with the stability curve of the cooling system, and then dynamically corrects the window by combining a dynamic adjustment factor and similar historical tasks.

[0023] In one embodiment, the quantum sensing backfilling module includes a short-term task detection module, which identifies short-term tasks that can be backfilled based on a three-layer filtering mechanism and generates a short-term task list based on priority sorting.

[0024] In one embodiment, the Unified Resource Abstraction Layer module includes a second GRES quantum extension module, which creates quantum resource types to define a set of physical characteristic parameters for quantum devices.

[0025] In one embodiment, the unified resource abstraction layer module includes a quantum device driver interface module, which shields hardware differences through an instruction conversion layer, enabling different quantum architectures to present a unified operating interface.

[0026] In one embodiment, the unified resource abstraction layer module includes a node extension module, which is embedded with quantum dynamic sensing capabilities and extends with a topology descriptor, a fidelity evaluator, and a calibration stater.

[0027] On the other hand, the present invention provides a dynamic scheduling method for quantum-classical hybrid computing tasks, including:

[0028] Quantum features are extracted from the task to be processed containing quantum features to obtain the first quantum feature, and quantum features are extracted from the computing resources to obtain the second quantum feature. The extracted first and second quantum features are cached as features of the quantum feature set.

[0029] The quantum feature set is obtained from the cache in real time, and the corresponding priority is output for the quantum task to be processed through multi-dimensional evaluation. The task processing priority is adjusted in real time based on the update mechanism, and the task priority sequence is updated in real time.

[0030] Based on the type of the task to be processed and the quantum feature set, a dependency relationship between classical tasks and quantum tasks is formed, and the dependency relationship is formed into a dependency constraint for the corresponding task to be processed, and / or a resource reservation instruction for the corresponding task to be processed.

[0031] Based on the priority of the tasks to be processed, the resource reservation instructions, and the real-time feedback of computing resource status, the idle periods that will appear in the future are predicted, and short-term quantum tasks that meet the conditions are inserted into the computing resources of the idle periods to fill the tasks.

[0032] Based on the priority and dependency constraints of the task to be processed, corresponding computing resources are allocated, and the utilization status of computing resources is fed back to the quantum sensing backfill module and the dynamic priority evaluation module in real time.

[0033] Beneficial effects:

[0034] 1. This invention extends the job submission interface of SLURM to accurately capture the characteristics of quantum tasks, establish a dynamic evaluation model, and enhance the dependency management mechanism, thereby supporting intelligent collaboration between quantum and classical tasks. Simultaneously, it optimizes the backfill scheduling algorithm to improve the utilization rate of quantum devices. Furthermore, by leveraging an extended resource abstraction layer, it provides unified heterogeneous resource management, enabling efficient scheduling of hybrid computing tasks and significantly improving the overall system performance.

[0035] 2. This invention responds in real time to the task coherence time constraint decay of quantum tasks through a dynamic priority mechanism, prioritizing the scheduling of quantum tasks that are about to fail, thus significantly reducing the failure rate of quantum tasks caused by queuing timeouts.

[0036] 3. This invention uses an intelligent dependency resolution engine and quantum dependency types to accurately manage the complex dependencies between quantum tasks and classical tasks, ensuring that related tasks are executed on demand and in an orderly manner, reducing waiting between tasks, and improving the overall efficiency of the hybrid computing process.

[0037] 4. This invention, through quantum-sensing intelligent backfill scheduling, can keenly capture the idle gaps of quantum devices and quickly fill them with suitable short quantum tasks, maximizing the running time of quantum computing devices and reducing resource idle waste.

[0038] 5. By combining dynamic prioritization, intelligent dependency management, and efficient quantum backfilling, this invention enables the system to more rationally arrange all tasks, shorten the average waiting time and completion time of jobs, and improve the throughput of the entire HPC cluster (high-performance computer cluster).

[0039] 6. This invention integrates the configuration, management, and scheduling of quantum devices with traditional CPUs, GPUs, and other resources under the SLURM framework through an extended GRES quantum resource abstraction layer, simplifying the management complexity of hybrid computing environments and improving the flexibility and efficiency of resource management. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are 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.

[0041] Figure 1 This is a schematic diagram of the scheduling system provided by the present invention.

[0042] Figure 2 The flowchart shows the dynamic calculation process in the dynamic priority evaluation module provided by this invention.

[0043] Figure 3 This is a flowchart illustrating the real-time updates in the dynamic priority evaluation module provided by the present invention.

[0044] Figure 4 The flowchart illustrates the implementation of the dependency resolution module provided in this invention.

[0045] Figure 5 A flowchart illustrating the construction of the DAG parsing engine provided by this invention.

[0046] Figure 6 The flowchart for conflict detection provided by this invention.

[0047] Figure 7 A flowchart for conflict strategy selection provided by the present invention.

[0048] Figure 8 The flowchart illustrates the implementation of the quantum sensing backfill module provided by this invention.

[0049] Figure 9 A flowchart illustrating the construction process of the window prediction module provided by this invention.

[0050] Figure 10 The flowchart for short-time task detection provided by this invention.

[0051] Figure 11 The flowchart illustrates the implementation of the unified abstraction layer module provided by this invention.

[0052] Figure 12 A flowchart illustrating the implementation of the GRES quantum extension provided by this invention.

[0053] Figure 13 This is a schematic diagram of the three-layer hardware abstraction control layer provided by the present invention.

[0054] Figure 14 A flowchart illustrating the integration process of the unified resource view provided by this invention.

[0055] Figure 15 This is a schematic diagram of explicit dependency annotation based on DAG provided by the present invention.

[0056] Figure 16 A flowchart of the scheduling method provided by the present invention.

[0057] In the diagram: 1. Quantum task feature extraction module; 2. Dynamic priority evaluation module; 3. Dependency resolution module; 4. Quantum perception backfilling module; 5. Unified resource abstraction layer module. Detailed Implementation

[0058] The invention will be more readily understood by referring to the following detailed description of preferred embodiments and included examples. Unless otherwise specified, 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. In case of conflict, the definitions in this specification shall prevail.

[0059] This invention provides a dynamic scheduling system and method for quantum-classical hybrid computing tasks based on SLURM, solving the problems of low quantum computing resource utilization caused by the static resource allocation strategy of existing SLURM scheduling systems, which cannot adapt to the dynamic characteristics of quantum tasks. Existing scheduling mechanisms lack consideration for the time-sensitive nature of quantum tasks and cannot dynamically adjust priorities. Insufficient management of the complex dependencies between quantum and classical tasks results in low task collaboration efficiency. Furthermore, the fragmented system resource view makes it difficult to achieve unified and optimized allocation of quantum and classical computing resources.

[0060] The technical concept of this invention lies in its ability to accurately capture the characteristics of quantum tasks by extending the job submission interface of SLURM, establishing a dynamic evaluation model, and enhancing the dependency management mechanism, thereby supporting intelligent collaboration between quantum and classical tasks. Simultaneously, it optimizes the backfill scheduling algorithm to improve the utilization rate of quantum devices. Furthermore, by leveraging an extended resource abstraction layer, it provides unified heterogeneous resource management, enabling efficient scheduling of hybrid computing tasks and significantly improving the overall system performance.

[0061] like Figure 1-15 As shown, a dynamic scheduling system for quantum-classical hybrid computing tasks based on SLURM includes:

[0062] like Figure 1 As shown, the quantum task feature extraction module 1 is used to extract quantum features from the task to be processed containing quantum features to obtain a first quantum feature, and to extract quantum features from the computing resources to obtain a second quantum feature. The extracted first and second quantum features are cached as a feature set. The first quantum feature includes parameters such as the number of qubits in each quantum node required by the quantum task, the task coherence time constraint range, and the supported set of gate operations. The second quantum feature is the configuration file of the quantum computing device, including parameters such as the number of qubits in each quantum node, the task coherence time constraint range, and the supported set of gate operations.

[0063] Specifically, the quantum task feature extraction module 1 is a core extension component of the SLURM system. While maintaining the original SLURM job submission system, it adds a quantum task processing channel through a plug-in approach. It consists of three parts: a parameter parser module, a validity verifier module, and a metadata conversion module. The three parts interact with each other through a standard interface, thereby ensuring seamless integration with the native SLURM components.

[0064] The quantum task feature extraction module 1 achieves deep integration with the SLURM system through the SLURM Quantum Plugin. This plugin is dynamically loaded when the control daemon starts and registers itself through the SLURM standard interface. During the initialization phase, the quantum task feature extraction module 1 reads the second quantum feature, i.e., the configuration file, of the quantum computing device to establish a device capability database, including key parameters such as the number of qubits of each quantum node, the task coherence time constraint range, and the supported set of gate operations. Simultaneously, the module initializes a quantum task description structure cache pool in memory for efficient processing of job submission requests.

[0065] In one embodiment, when submitting a job (such as a quantum task or a classical task), the job undergoes parameter preprocessing. For tasks with quantum-specific parameters (such as `--quantum-type`), the command parser module creates a dedicated quantum parameter context structure. Exemplarily, the quantum parameter context structure employs a layered design, including a base layer and an extension layer. The base layer contains standard SLURM parameters such as job ID, submitting user, required number of CPU cores, memory requirements, task execution time limits, and output log paths. The extension layer stores quantum characteristic parameters such as `--quantum-type` (quantum task type), `--coherence-time` (task coherence time constraint), `--qubits-required` (number of qubits required), and `qpu_type` (target quantum processor type). During parameter parsing, the system performs preliminary format checks to ensure that time parameters in the extension layer use microseconds and that the number of qubits is a positive integer, among other basic constraints. After parsing, the parsed parameter set is encapsulated into a binary message and sent to the control daemon via the SLURM scheduling system's internal communication protocol.

[0066] In one embodiment, when the control daemon (slurmctld) receives a job request, the SLURMQuantum Plugin (quantum plugin) first extracts the first quantum feature, i.e., the quantum feature parameter, and inputs it into the legality verifier module for deep verification. Specifically, the verification process consists of three stages. The first stage checks the legality of the quantum computing type to ensure it matches the type of quantum device registered in the cluster. The second stage verifies the resource requirements, checking whether the requested number of qubits is within the node's capabilities. The third stage evaluates time feasibility, calculating whether the task execution time is within the specified task coherence time constraint window. It should be noted that a detailed error code system is set up for each verification stage. For example, when the requested number of qubits exceeds the limit, detailed error information containing the maximum number of qubits available to the node will be returned.

[0067] In one embodiment, when the verified first quantum feature, i.e., the quantum feature parameter, enters the metadata conversion module, the system converts the text-formatted quantum feature parameter into an optimized binary structure. This structure converts task coherence time constraints into relative timestamps, and maps qubit requirements to device topology representations. The converted data is written to an extended job description structure, which adds a quantum-specific field area while maintaining compatibility with traditional job descriptions. Simultaneously, the module persists the quantum metadata to the SLURM database via an asynchronous write thread. For example, fields such as qpu_type and coherence_window are added to the existing job_table, and an associated index is established with the quantum device table.

[0068] In one embodiment, the final step in feature extraction is the real-time updating of quantum resource states. The module maintains a distributed quantum resource state table. Once a new quantum task is validated, the system atomically updates the "reserved bits" and "estimated availability time" fields of the relevant quantum nodes. This state information is synchronized to the entire cluster via SLURM's node feature broadcasting mechanism, providing real-time data support for scheduling decisions. State updates employ an optimistic locking mechanism to handle concurrent conflicts, ensuring data consistency under high load scenarios.

[0069] In one embodiment, a complete error handling chain is established during the extraction of quantum feature parameters. If an anomaly occurs at any stage, the system performs a transaction rollback, releases temporarily allocated resource identifiers, clears pre-stored metadata, and restores the quantum device state flag. In some embodiments, for retryable errors, the quantum feature extraction module implements an exponential backoff retry algorithm. In some embodiments, for unrecoverable errors, detailed diagnostic information can be returned via the SLURM standard error channel. All error handling processes are recorded in the quantum operation audit log, supporting post-event analysis.

[0070] In summary, the quantum feature extraction module is deeply integrated with SLURM through a plug-in approach. It parses the quantum task-specific parameters submitted by users, verifies the matching of these parameters with the capabilities of quantum devices, converts the parameters into standardized binary structures for storage and updates the quantum resource status, and handles parameter errors and rollbacks. This provides accurate quantum task feature data for subsequent scheduling processes such as dynamic priority evaluation and dependency resolution, achieving seamless compatibility with the native SLURM architecture.

[0071] like Figure 2-3As shown, the dynamic priority evaluation module 2 is used to obtain the quantum feature set from the cache of the quantum task feature extraction module in real time, and output the corresponding priority for the quantum task to be processed through multi-dimensional evaluation processing, and adjust the task processing priority in real time based on the update mechanism, and output the real-time updated task priority sequence.

[0072] Specifically, the Dynamic Priority Evaluation Module 2 implements dynamic priority evaluation through a deep modification of the native SLURM priority / basic plugin. It extends the priority_type enumeration type, adding dedicated priority categories QUANTUM_URGENT and QUBIT_CRITICAL. The Dynamic Priority Evaluation Module 2 adopts a microservice design, encapsulating the core algorithm into independent computational units that interact with the SLURM master scheduler via shared memory. It maintains a real-time priority matrix in memory, containing the original priority, remaining value, task importance score, and final effective priority for each job.

[0073] In one embodiment, the existing priority type system is extended for quantum computing by adding two priority categories specifically for quantum computing tasks: quantum urgent tasks and bit critical tasks. This can be achieved by modifying the implementation of the `priority_type_t` enumeration structure in the enumeration definition file of the SLURM source code.

[0074] In one embodiment, a novel priority matrix data structure is designed to manage the various parameters required for dynamic priority calculation. This data structure achieves data sharing among multiple processes through a shared memory mechanism. The data structure includes job ID, base priority, remaining value, task importance score, final dynamic priority, priority type, last update time, and red-black tree nodes. It should be noted that this priority matrix data structure is maintained in two independent but synchronized data structures: an indexable hash table for fast lookup and a red-black tree for the scheduler to quickly retrieve the highest priority task.

[0075] like Figure 2As shown, in one embodiment, the dynamic priority evaluation module 2 includes a task coherence time constraint decay module. This module calculates the remaining task coherence time value by considering the initial coherence time window of the quantum task, the waiting time after task submission, the effective coherence time constant determined by the quantum device and circuit depth factors, and the additional decay rate caused by environmental noise. This quantifies the time urgency of the quantum task. Specifically, the task coherence time constraint decay module is a task coherence time constraint decay model. The task coherence time constraint (TCW) refers to the maximum allowed time window from task readiness to completion; exceeding this window results in unreliable calculations due to quantum state decoherence. In the task flow of this invention, each quantum task independently possesses a task coherence time constraint parameter, which can be declared by the user when submitting the task or automatically derived by the system. The task coherence time constraint decay indicates that after submission, the task coherence time constraint decreases linearly with the waiting time, and the remaining value determines the task urgency. Therefore, a decay model is established for this parameter to quantify the time urgency of the quantum task. Its calculation model is based on the following formula:

[0076]

[0077] Among them, TCW initial T represents the initial coherence time window of the task (declared by the user when submitting the task or automatically derived by the system). wait Γ represents the waiting time after task submission, τ is the effective coherence time constant of the task determined by both the quantum device and the circuit depth factor, and Γ represents the waiting time after task submission. noise This represents the additional attenuation rate caused by environmental noise (the quantum device monitors this in real time and transmits the data to the system). RemainingTCW(t) represents the remaining task coherence time; the smaller the remaining task coherence time, the higher the task urgency.

[0078] In one embodiment, a dynamic priority evaluation module 2 is further included, which includes a task importance evaluation module. The task importance evaluation module calculates the importance of the quantum task by considering user level, quantum gate operation requirement fidelity, the proportion of bits occupied by the task, and weight, so as to quantify the importance of the quantum task.

[0079] Specifically, the task importance module assesses task importance using a comprehensive formaldehyde evaluation system, calculated using the following formula:

[0080]

[0081] Among them, U lever This indicates the user level (0.1-1.0), derived from the SLURM account configuration. Regular users have a level of 0.1-0.5, while VIP users have a level of 0.6-1.0; F reqThis indicates the fidelity requirement for quantum gate operations; This represents the number of bits occupied by the task / the total number of bits in the cluster. When a task occupies a high proportion of the cluster's quantum bit resources, its execution process will significantly reduce the overall parallel capability of the system. If the task fails or is delayed, it will result in a greater opportunity cost loss. At the same time, high-bit-occupancy tasks usually correspond to more critical quantum computing stages. Therefore, the system will automatically increase the overall priority of the task to ensure that scarce quantum resources are tilted towards the task that has the greatest impact on the overall efficiency of the system. This represents the weight amount, and the constraint conditions are:

[0082] After calculating the remaining task coherence time and task importance, the final priority is calculated using the following formula:

[0083]

[0084] In summary, calculating the final priority provides the scheduler with a clear basis for task sorting. By comprehensively considering the time sensitivity (e.g., coherence time window), resource requirements (e.g., number of qubits), and priority type (e.g., quantum-specific priority) of quantum tasks with the basic parameters of classical tasks (e.g., submission time, resource scale), a unified priority value is generated. This ensures that high-priority, time-sensitive tasks (especially quantum tasks) can obtain resources first, avoiding quantum state decoherence or resource conflicts caused by delays. At the same time, it balances the reasonable scheduling of classical tasks, ultimately achieving efficient resource allocation and orderly task execution in the quantum-classical hybrid computing system, improving overall scheduling efficiency and task success rate.

[0085] like Figure 3As shown, in one embodiment, the dynamic priority evaluation module 2 adopts a real-time update mechanism, which includes an event-driven update mode and a periodic scan update mode to balance the timeliness and comprehensiveness of updates, ensuring that the scheduler always makes decisions based on accurate and up-to-date system states. Specifically, in the event-driven mode, tasks enqueuing / dequeuing, user parameter modifications, and other operations immediately trigger the priority calculation unit to perform real-time updates. This allows for rapid response to sudden state changes, preventing time-sensitive quantum tasks from failing due to delays by triggering updates instantly. In the periodic scan mode, the background daemon automatically traverses the task queue, recalculates the final priority of all quantum tasks, and updates it. The update process follows a standardized procedure: first, task characteristic parameters are obtained; then, the remaining values ​​are calculated; then, the task importance is evaluated; and finally, dynamic priorities are synthesized and the matrix data is updated. The entire process is designed with pipelines to ensure efficient execution, and key state changes are reflected in the priority matrix within a certain time. The periodic scan mode comprehensively inspects the system state at fixed intervals, capturing gradual changes without event triggers, such as the natural decay of quantum bit coherence over time and the slow increase in classical node load, thus compensating for the blind spots of event-driven updates in covering latent changes.

[0086] In one embodiment, the dynamic priority evaluation module 2 includes a priority matrix management module. This module maintains a task queue sorted by final priority using a red-black tree (self-balancing binary search tree) to quickly obtain the highest priority task. It employs read-write locks and segmented locks to handle concurrent access, and combines time window merging to partition and process update requests in parallel, then merges the red-black tree for rebalancing. Specifically, the dynamic priority evaluation module 2 adopts a hierarchical optimization strategy. First, the data structure design uses a red-black tree to maintain an ordered queue of final priorities, ensuring that the highest priority task is always at the root of the tree. Concurrency control uses a read-write lock mechanism, where the scheduler read operation uses a shared read lock to allow concurrent access by multiple threads, single-job updates use an exclusive write lock but strictly limit the holding time, and full scans use segmented locks to partition by priority range. The batch update strategy uses time window merging technology to collect update requests, divide them into blocks according to priority range, process them in parallel, and finally merge them for red-black tree rebalancing.

[0087] A red-black tree structure ensures fast querying, insertion, and deletion of task priorities, while time window merging reduces performance overhead from frequent updates, adapting to the surge in task numbers required in hybrid computing scenarios. Read-write locks and segmented locks ensure efficient sharing of read operations and atomicity of update operations when multiple tasks simultaneously request priority information, preventing data inconsistency. In summary, by maintaining a real-time, ordered priority matrix, a clear task sorting basis is provided to the scheduler, ensuring that high-priority, time-sensitive quantum tasks are scheduled first, ultimately improving resource utilization and task execution reliability in hybrid computing systems.

[0088] like Figure 4-7 As shown, dependency resolution module 3, based on the type of the task to be processed (e.g., quantum task and / or classical task) and the quantum feature set, forms the dependency relationship between classical and quantum tasks, and transforms the dependency relationship into dependency constraints for the corresponding task to be processed, and / or resource reservation instructions for the corresponding task to be processed. Dependency resolution module 3 achieves intelligent dependency management for quantum and classical tasks through a deep extension of the native SLURM dependency mechanism. First, at the system architecture level, the dependency type system is extended, adding two special dependency relationships: quantum-classical hybrid (quantum_classic) and quantum sequence (qubit_sequence). Furthermore, a directed acyclic graph modeling module is used in dependency resolution module 3 to abstract task dependencies into a weighted topological graph structure, ensuring the reliability of dependency execution through a real-time resource reservation table and conflict detection algorithm. A multi-dimensional dependency state matrix is ​​maintained in memory, containing data flow relationships between tasks, resource occupation mapping, and timing constraint information.

[0089] like Figure 4 As shown, the DAG parsing engine, which is the module for modeling directed acyclic graphs, is constructed through the following steps:

[0090] First, node creation is performed. During the node creation phase, a node structure with type labels is generated for each job. Quantum nodes record bit requirements, gate sets, and task coherence time constraints, while classical nodes label CPU / GPU resource requirements.

[0091] Then, edge relationships are constructed. During the edge relationship construction stage, dependency declarations are resolved to generate five types of dependency edges: classical to quantum edges to indicate data format conversion requirements, quantum to classical edges to define result processing specifications, and quantum sequence edges to record bit continuity requirements.

[0092] Then, graph verification is performed, which involves a triple check: using an improved Tarjan algorithm to prevent quantum sequence loops; ensuring physical connectivity between quantum tasks through qubit overlap analysis; and establishing a set of time-constrained equations for verification.

[0093] Finally, resource estimation is performed, taking into account both quantum and classical requirements, the peak memory for classical tasks is calculated, 20% of error correction spare bits are reserved for quantum tasks, and calibration time overhead is considered.

[0094] In summary, the DAG parsing engine parses user-submitted hybrid tasks into a Directed Acyclic Graph (DAG) structure, where nodes represent tasks and edges represent dependencies between tasks. It also verifies the validity of dependencies and transforms these dependencies into metadata recognizable by the scheduling system. The DAG parsing engine provides a dependency chain for subsequent collaborative reservation mechanisms, ensuring that related tasks reserve resources sequentially. It also supports dynamic scheduling adjustments; when the state of a task changes, it can quickly locate affected downstream tasks and trigger rescheduling. Furthermore, it avoids resource waste caused by dependency chaos, ultimately ensuring that hybrid tasks are executed in an orderly and efficient manner according to dependency logic.

[0095] like Figure 5-6 As shown, in one embodiment, the dependency resolution module 3 includes a collaborative reservation module. The collaborative reservation module adopts a multi-dimensional conflict detection and three-level resolution strategy based on the quantum resource reservation table structure, and realizes distributed resource collaboration based on a three-phase commit protocol.

[0096] Specifically, a quantum resource reservation table structure is constructed, including a QPU identifier, a time window, a qubit map, and a list of associated classical nodes. Upon receiving a new reservation request, the system first checks for conflicts. It performs a time window overlap check, determining if the new request's time period overlaps with any existing reservation time period. If overlap exists, quantum resource conflict detection is initiated by comparing the qubit map—the bit mask of the new request—with the existing reservation mask using bitwise operations. If the same physical bits are occupied simultaneously, a quantum resource conflict is identified. Simultaneously, classical resource conflict detection scans the list of classical computing nodes bound to the new request. If any node is already occupied by an existing reservation, a classical resource conflict is marked. The system ultimately determines a conflict as valid and triggers a resolution process only if both time windows overlap and at least one of a quantum or classical resource conflict occurs. For advanced scenarios supporting quantum entanglement, the system further verifies the physical connection status of the qubit group in the new request, ensuring its topological connectivity meets the task requirement threshold.

[0097] like Figure 7As shown, the system employs a three-tiered strategy framework to dynamically resolve conflicts of different types. Time window conflicts are handled using an elastic offset strategy. Quantum resource conflicts trigger topology-aware remapping; classical resource conflicts execute isomorphic node replacement. When encountering mixed conflicts, a quantum-classical decoupling mechanism is activated to split the task into asynchronous execution units. If the above strategies fail, priority arbitration is triggered, allowing high-priority tasks to preempt low-priority resource reservations.

[0098] For the reservation system, the system achieves distributed resource coordination based on a three-phase commit protocol: in the reservation phase, the scheduler generates a global transaction ID, locks quantum processor and classical node resources in parallel, and creates a joint reservation certificate with a version number; in the dependency verification phase, the system detects the status of the preceding tasks and the consistency of resource versions in the DAG graph in real time; in the commit phase, the system atomically updates the resource status to "committed" and broadcasts an execution ready signal.

[0099] like Figure 15 As shown, in one embodiment, based on explicit dependency annotations of a directed acyclic graph (DAG), after a user submits a DAG definition containing quantum task nodes and classical task nodes via a YAML / JSON file, the DAG parsing engine deconstructs the file content and constructs a dependency graph. This graph is transformed into an executable plan during the scheduler integration phase, where the quantum property extension module injects quantum computing-specific constraints: quantum task nodes must satisfy qubit coherence time limits, and dependency edges between tasks apply a maximum wait threshold to prevent quantum state decoherence. Before task execution, the runtime dependency checking module dynamically verifies the validity of dependencies between tasks, especially ensuring that the timing constraints in the quantum-classical hybrid task chain are satisfied, ultimately triggering task execution under the premise of satisfying all quantum property and classical dependency conditions.

[0100] The DAG definition specification requires users to explicitly describe the nodes and dependencies of a quantum-classical hybrid task using structured files (YAML / JSON). Quantum task nodes must include a device ID, a quantum circuit file path, and a virtual scheduling window (TCW). v and Physical Scheduling Window TCW p Specific attributes, while classic task nodes require defining the execution command, resource requirements, and task time (TCW). classic Dependency edges declare the data flow direction through the inputs / outputs fields and set the maximum wait time constraint for quantum-sensitive links to ensure the validity of quantum states.

[0101] After the file is parsed, the system will automatically construct a task graph DAG file with temporal constraints.

[0102] Subsequently, by resolving the epidemic through dependencies, loading the DAG file and performing topological sorting, the order of task execution is established;

[0103] Subsequently, critical path analysis was used to identify the task paths that had the greatest impact on the overall execution time, with a particular focus on quantum computing tasks included within them. Total path duration calculation:

[0104] T path =∑T classic +max(TCW v TCW p ).

[0105] During the constraint injection phase, the system performs specific processing for quantum characteristics: setting a maximum delay constraint for the path of the quantum task, ensuring the task is completed before quantum state decoherence; and automatically binding the optimal calibration strategy for the quantum device. The resulting dependency graph satisfies both classical computational dependencies and strictly adheres to the physical constraints of quantum computing, achieving efficient and reliable execution of the hybrid workflow.

[0106] Finally, the coordinated scheduling of quantum-classical tasks is achieved through dynamic priority evaluation and real-time dependency verification: after the task is submitted, the dependency checking module will continuously monitor the state changes of the data entity. If the output of the predecessor task is ready within the maximum waiting time, the successor task will be added to the ready queue.

[0107] When the system resource arbiter allocates devices for a quantum task, it will forcibly verify whether the current device T1 (T1 (relaxation time): the average time for a qubit to spontaneously decay from a high-energy state to the ground state, reflecting the energy dissipation rate.) / T2 ((decoherence time): the time for the qubit's phase information to be retained, reflecting the phase decoherence rate.) value of the target quantum processor meets the task's TCW. p Constraints are imposed, and subsequent classical tasks are triggered to execute via hardware interrupt signals; the exception handler monitors error signals from quantum devices in real time, automatically activates backup strategies, and dynamically adjusts the maximum waiting time and priority of constraints for affected tasks, ultimately achieving nanosecond-level response hybrid task scheduling in a unified resource view.

[0108] like Figure 8-10 As shown, it also includes a quantum-sensing backfill module 4. The quantum-sensing backfill module 4 predicts idle periods that will occur in the future based on the priority of the task to be processed, the resource reservation instruction, and the real-time feedback of the computing resource status. It then inserts eligible short-term quantum tasks into the computing resources of the idle periods for task backfilling. Specifically, the quantum-sensing backfill module 4 deeply modifies the original backfill scheduling plugin of SLURM to increase support for the special characteristics of quantum computing tasks.

[0109] Specifically, the quantum sensing backfill module 4 includes a first GRES quantum extension module, which is used to add quantum processing units to support the definition of key parameters of the quantum device. The key parameters include the number of qubits, gate operation fidelity, topological connectivity, etc.

[0110] like Figure 9 As shown, in one embodiment, the quantum sensing backfill module 4 includes a window prediction module, which calculates the basic available window by analyzing the historical calibration log of the quantum device and combining it with the stability curve of the cooling system, and dynamically corrects the window by combining a dynamic adjustment factor and similar historical tasks.

[0111] The construction of the window prediction module, i.e., the window prediction model, specifically includes the following:

[0112] First, the baseline window is calculated by extracting the standard period from the calibration log of the quantum controller and combining it with the stability curve of the cooling system to calculate the basic available window.

[0113] W base =T coherence -T calibration -T cooling

[0114] Among them, T coherence It is the decoherence time of a quantum device, T calibration It is the calibration time, T cooling It is the stabilization time of the refrigeration system.

[0115] Secondly, the dynamic adjustment factor is calculated by continuously collecting dynamic parameters such as equipment temperature fluctuations and gate error rate changes. The window prediction value is corrected using a linear regression model, and the gate error rate matrix is ​​obtained in real time through a quantum controller. Environmental monitoring collects electromagnetic interference intensity ε EMI The final result is obtained through the following formula:

[0116]

[0117] Where α, β, and γ are empirical weights. This is the average gate error rate. A sliding window linear regression method is used to dynamically update the weights.

[0118]

[0119] Where η represents the learning rate, which controls the step size of parameter updates and prevents oscillations; The gradient of the loss function represents the degree to which changes in weights affect the error.

[0120] Finally, the window was corrected as follows:

[0121] W dynamic =W base -(1-min(α t ,|ΔW|))

[0122] Finally, the output is weighted and fused to generate the final predicted value. The model weight parameters are automatically updated within a certain period of time to ensure that the prediction results always reflect the latest status of the equipment.

[0123] Weighted fusion mechanism of output:

[0124] W predict =ω b ·W base +ω d ·W dynamic +ω h ·W final

[0125] Where, ω b ω d ω h As a weighted amount.

[0126] In summary, the construction of a window prediction model endows the system with the ability to accurately capture the temporal characteristics of quantum devices. By integrating historical calibration logs, real-time state parameters, and task execution data of quantum devices, a prediction algorithm adapted to quantum characteristics is designed, enabling it to dynamically generate future available time windows for the devices and match the coherent time constraints of quantum tasks. This construction process ensures that the model can respond to changes in device state in real time and accurately assess the feasibility of tasks within the window, providing the scheduler with a time-dimensional decision-making basis. Ultimately, this achieves efficient utilization of quantum resources and reliable execution of time-sensitive tasks, serving as the core support for balancing timeliness and utilization in quantum-classical hybrid scheduling.

[0127] like Figure 10 As shown, in one embodiment, the quantum sensing backfill module 4 includes a short-term task detection module, which identifies short-term tasks that can be backfilled based on a three-layer filtering mechanism and generates a short-term task list based on priority sorting.

[0128] Specifically, the primary filter scans the submitted parameters to quickly identify candidate tasks that require fewer than 5 qubits and require fewer than 100 request gate operations.

[0129] The secondary predictor uses the K-nearest neighbor algorithm to compare the feature vector of the new task with the historical database, find the five historical tasks with the highest similarity, and take the median of their execution time as the predicted value.

[0130] A dynamic regulator monitors the real-time queue status. When the number of tasks to be scheduled increases, the duration threshold is automatically lowered to ensure that the selection criteria always adapt to the current resource pressure. The system performs a full scan of the task queue within a certain timeframe, labeling each candidate task with its predicted execution time and confidence score, ultimately generating a list of short tasks sorted by urgency.

[0131] In summary, the short-time task detection module, or short-time task detection system, can accurately identify and efficiently adapt to quantum or classical tasks with extremely short execution times (such as microseconds or milliseconds). By working in conjunction with the window prediction model and resource scheduler, it maximizes the utilization of system resource fragments while ensuring the timeliness of short-time tasks.

[0132] In one embodiment, the quantum-sensing backfill module 4, i.e., the backfill decision engine, is modified. Specifically, quantum-sensing decision logic is embedded in the backfill plugin to establish a four-stage workflow:

[0133] First, identification is performed during idle periods. The idle detector continuously scans the timeline of the quantum device to identify gap windows between adjacent tasks that exceed a safety threshold.

[0134] Secondly, a backfillability assessment is performed. The candidate evaluator performs a four-dimensional check on the short task list: verifying that the request bit is available in the target window, confirming that the predicted duration is less than 80% of the window, checking that all task dependencies are met, and evaluating the device temperature stability.

[0135] Secondly, for eligible tasks, the fragment optimizer initiates a dual spatial-temporal sorting process: in the temporal dimension, adjacent micro-windows are merged to form continuous time periods; in the spatial dimension, discrete qubits are integrated into continuous blocks through bit remapping.

[0136] Finally, the task is bound to the target window through an atomic commit protocol, while reserving time margin to deal with the risk of quantum state decay.

[0137] In summary, the quantum sensing backfill module 4 periodically scans the status of quantum devices. When an upcoming idle period is detected, it triggers the window prediction model to update the prediction results. Simultaneously, it traverses the scheduling queue and uses a short-time task detection algorithm to filter out candidate tasks. It performs a backfillability assessment on each candidate task, and tasks that pass the assessment enter the resource reservation process. If resource fragmentation is found, it initiates a fragmentation defragmentation strategy to attempt to expand the window. Finally, it inserts eligible tasks into the idle window for execution.

[0138] like Figure 11-14As shown, it also includes a unified resource abstraction layer module 5. The unified resource abstraction layer module 5 allocates corresponding computing resources based on the priority of the task to be processed and the dependency constraints of the task to be processed, and provides real-time feedback on the utilization status of computing resources to the quantum sensing backfill module 4 and the dynamic priority evaluation module 2.

[0139] like Figure 12 As shown, in one embodiment, the Unified Resource Abstraction Layer module 5 includes a second GRES quantum extension module. This module creates quantum resource types to define the set of physical characteristic parameters for quantum devices. Specifically, quantum resource types are created within the GRES core architecture to define the set of physical characteristic parameters for quantum devices. By declaring quantum-specific attribute fields in the `gres.conf` configuration file, the system establishes a standardized description system for qubit resources: qubit size describes the computational capability boundary, the topology connection matrix encodes the physical correlation between bits, the coherence time window constrains the upper limit of task execution, and the gate operation precision sets the service quality baseline. A topology-aware scheduling algorithm is developed to automatically convert discrete qubit requests into optimal physical mapping schemes, ensuring that the distance constraints of quantum gate operations are met while avoiding high-error-rate qubit regions.

[0140] like Figure 13 As shown, in one embodiment, the unified resource abstraction layer module 5 includes a quantum device driver interface module. This module shields hardware differences through an instruction conversion layer, enabling different quantum architectures to present a unified operating interface. Specifically, the quantum device driver interface module constructs a three-layer hardware abstraction control layer, including a bottom-level monitoring module, a middle-level control engine, and an upper-level fault-tolerant hub. The bottom-level monitoring module establishes a comprehensive picture of the quantum device's state, continuously collecting bit-level operating parameters; the middle-level control engine encapsulates the entire lifecycle of tasks, achieving standardized management of quantum operations; and the upper-level fault-tolerant hub establishes a quantum-specific error classification system, designing dedicated recovery strategies for quantum-specific faults such as decoherence. This interface shields hardware differences through an instruction conversion layer, enabling different quantum architectures to present a unified operating interface, significantly reducing platform adaptation complexity.

[0141] In one embodiment, the unified resource abstraction layer module 5 includes a node extension module, which incorporates quantum dynamic sensing capabilities. This node extension module extends to include a topology descriptor, a fidelity evaluator, and a calibration state machine. Quantum dynamic sensing capabilities are embedded into the node characteristic system. The topology descriptor transforms the physical connections of the quantum chip into programmable structured data; the fidelity evaluator integrates multiple physical parameters to construct a real-time quality scoring model; and the calibration state machine implements a seamless switching strategy between three modes. Through real-time visualization of the modified system commands, a dynamically updated digital profile of the quantum device is formed, providing an accurate hardware state reference for scheduling decisions.

[0142] like Figure 14 As shown, a panoramic hybrid resource management system can be created through the cooperation of the aforementioned device driver interface module and node extension module. The cross-resource query interface breaks down the display barriers between quantum and classical computing, outputting a complete resource topology with a single command; the physical mapping engine establishes a hard association between the quantum controller and classical units, optimizing data transmission paths; and the logical resource pool technology allows for the definition of quantum-classical combined units, enabling users to obtain matching computing resource packages through simple declarations. This integration allows quantum devices to naturally integrate into the existing resource scheduling system, forming a unified hybrid computing resource landscape.

[0143] like Figure 16 As shown, the present invention also provides a dynamic scheduling method for quantum-classical hybrid computing tasks, comprising:

[0144] The first quantum feature is obtained by extracting quantum features from the task to be processed containing quantum features, and the second quantum feature is obtained by extracting quantum features from the computing resources. The extracted first and second quantum features are cached as features of the quantum feature set.

[0145] The quantum feature set is obtained from the cache in real time, and the corresponding priority is output for the quantum task to be processed through multi-dimensional evaluation processing. The task processing priority is adjusted in real time based on the update mechanism, and the task priority sequence is updated in real time.

[0146] Based on the type of the task to be processed and the quantum feature set, a dependency relationship between classical tasks and quantum tasks is formed, and the dependency relationship is formed into a dependency constraint for the corresponding task to be processed, and / or a resource reservation instruction for the corresponding task to be processed.

[0147] Based on the priority of the tasks to be processed, the resource reservation instructions, and the real-time feedback of computing resource status, the idle periods that will appear in the future are predicted, and short-term quantum tasks that meet the conditions are inserted into the computing resources of the idle periods to fill the tasks.

[0148] Based on the priority and dependency constraints of the task to be processed, corresponding computing resources are allocated, and the utilization status of computing resources is fed back to the quantum sensing backfill module and the dynamic priority evaluation module in real time.

[0149] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic scheduling system for quantum-classical hybrid computing tasks based on SLURM, characterized in that, include: The quantum task feature extraction module is used to extract quantum features from the task to be processed containing quantum features to obtain the first quantum feature, and to extract quantum features from the computing resources to obtain the second quantum feature, and to cache the extracted first and second quantum features as features of the quantum feature set. The dynamic priority evaluation module is used to obtain the quantum feature set from the cache in real time, and output the corresponding priority for the quantum task to be processed through multi-dimensional evaluation processing. It also adjusts the task processing priority in real time based on the update mechanism and outputs a real-time updated task priority sequence. The dependency resolution module, based on the type of the task to be processed and the quantum feature set, forms the dependency relationship between the classical task and the quantum task, and forms the dependency relationship into the dependency constraint of the corresponding task to be processed, and / or forms the resource reservation instruction of the corresponding task to be processed. The quantum sensing backfilling module, based on the priority of the task to be processed, the resource reservation instruction, and the real-time feedback of the computing resource status, predicts the idle time periods that will appear in the future, and inserts the eligible short-term quantum tasks into the computing resources of the idle time periods for task backfilling. The unified resource abstraction layer module allocates corresponding computing resources based on the priority and dependency constraints of the task to be processed, and provides real-time feedback on the utilization status of computing resources to the quantum sensing backfill module and the dynamic priority evaluation module.

2. The scheduling system according to claim 1, characterized in that, The quantum task feature extraction module includes: The parameter resolver module creates a quantum parameter context structure for a task containing quantum-specific parameters. The quantum parameter upper and lower structures include a base layer and an extension layer. The base layer is used to store standard SLURM parameters, and the extension layer is used to store quantum property parameters.

3. The scheduling system according to claim 1, characterized in that, The quantum task feature extraction module includes a legitimacy verifier module, which is used to verify the legitimacy of the extracted first quantum feature in terms of quantum computing type, resource requirements, and time feasibility.

4. The scheduling system according to claim 1, characterized in that, The quantum task feature extraction module includes a metadata conversion module, which is used to convert the first quantum feature in text format into a binary structure and write the converted binary structure into an extended job description structure.

5. The scheduling system according to claim 1, characterized in that, The dynamic priority evaluation module includes a task coherence time constraint decay module, which calculates the remaining value of task coherence time by using the initial coherence time window of the quantum task, the waiting time after task submission, the effective coherence time constant of the task determined by the quantum device and circuit depth factor, and the additional decay rate caused by environmental noise.

6. The scheduling system according to claim 1, characterized in that, The dynamic priority evaluation module includes a task importance evaluation module, which calculates the importance of quantum tasks based on user level, quantum gate operation fidelity requirements, the proportion of bits occupied by the task, and weighting.

7. The scheduling system according to claim 1, characterized in that, The final priority of the dynamic priority evaluation module is calculated by taking the remaining value of the task coherence time, the importance of the quantum task, and the initial coherence time window, and then updating the priority matrix based on the final priority.

8. The scheduling system according to claim 1, characterized in that, The dynamic priority evaluation module adopts a real-time update mechanism, which includes an event-driven update mode and a periodic scan update mode.

9. The scheduling system according to claim 1, characterized in that, The dynamic priority evaluation module includes a priority matrix management module. The priority matrix management module maintains a task queue sorted by final priority using a red-black tree to quickly obtain the highest priority task. It uses read-write locks and segmented locks to handle concurrent access, and combines time window merging to partition and process update requests in parallel and merge the red-black tree for rebalancing.

10. The scheduling system according to claim 1, characterized in that, The dependency resolution module includes a directed acyclic graph modeling module, which abstracts task dependencies into a weighted topological graph structure.

11. The scheduling system according to claim 1, characterized in that, The dependency resolution module includes a collaborative reservation module, which adopts a multi-dimensional conflict detection and three-level resolution strategy based on the quantum resource reservation table structure, and realizes distributed resource collaboration based on a three-phase commit protocol.

12. The scheduling system according to claim 1, characterized in that, The quantum sensing backfill module includes a first GRES quantum extension module, which is used to add quantum processing units to support the definition of key parameters of the quantum device.

13. The scheduling system according to claim 1, characterized in that, The quantum sensing backfill module includes a window prediction module. The window prediction module calculates the basic available window by analyzing the historical calibration logs of the quantum device and combining them with the stability curve of the cooling system. It then dynamically corrects the window by combining a dynamic adjustment factor and similar historical tasks.

14. The scheduling system according to claim 1, characterized in that, The quantum sensing backfilling module includes a short-time task detection module, which identifies short-time tasks that can be backfilled based on a three-layer filtering mechanism and generates a short-time task list based on priority sorting.

15. The scheduling system according to claim 1, characterized in that, The Unified Resource Abstraction Layer module includes a second GRES quantum extension module, which is used to create quantum resource types to define the set of physical characteristic parameters of quantum devices.

16. The scheduling system according to claim 1, characterized in that, The unified resource abstraction layer module includes a quantum device driver interface module. The quantum device driver interface module shields hardware differences through an instruction conversion layer, enabling different quantum architectures to present a unified operating interface.

17. The scheduling system according to claim 1, characterized in that, The unified resource abstraction layer module includes a node extension module, which is embedded with quantum dynamic sensing capabilities and extends with a topology descriptor, a fidelity evaluator, and a calibration stater.

18. A dynamic scheduling method for quantum-classical hybrid computing tasks, characterized in that, include: The first quantum feature is obtained by extracting quantum features from the task to be processed containing quantum features, and the second quantum feature is obtained by extracting quantum features from the computing resources. The extracted first and second quantum features are cached as features of the quantum feature set. The quantum feature set is obtained from the cache in real time, and the corresponding priority is output for the quantum task to be processed through multi-dimensional evaluation processing. The task processing priority is adjusted in real time based on the update mechanism, and the task priority sequence is updated in real time. Based on the type of the task to be processed and the quantum feature set, a dependency relationship between classical tasks and quantum tasks is formed, and the dependency relationship is formed into a dependency constraint for the corresponding task to be processed, and / or a resource reservation instruction for the corresponding task to be processed. Based on the priority of the tasks to be processed, the resource reservation instructions, and the real-time feedback of computing resource status, the idle periods that will appear in the future are predicted, and short-term quantum tasks that meet the conditions are inserted into the computing resources of the idle periods to fill the tasks. Based on the priority and dependency constraints of the task to be processed, corresponding computing resources are allocated, and the utilization status of computing resources is fed back to the quantum sensing backfill module and the dynamic priority evaluation module in real time.

Images