A method, system, and intelligent computing center operation and maintenance platform for dynamic scheduling of scheduled tasks
By splitting scheduled tasks into phase diagrams and generating scheduling contracts, combining resource envelope and freshness calculations, selecting low-conflict resource intervals and execution windows, and issuing dual-term leases, the resource conflict problem in task scheduling in intelligent computing centers is solved, achieving dynamic task scheduling and uniqueness of results.
Patent Information
- Application Number
- CN202611313690.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-10
AI Technical Summary
Existing timed task scheduling schemes cannot effectively resolve resource conflicts in training jobs in intelligent computing centers, leading to repeated or missed task executions, and failing to guarantee the freshness of results and the effective use of resources.
By splitting scheduled tasks into phase diagrams, a task scheduling contract is generated. Combining resource envelopes, checkpoints, and side effect submission points, the freshness deficit is calculated, a low-conflict resource interval and execution window are selected, and a dual-period lease is issued. The execution agent only executes or submits the task phase when the lease verification passes, thus achieving dynamic scheduling.
It enables dynamic scheduling of tasks in the intelligent computing center, avoids resource conflicts, ensures the freshness and uniqueness of results, reduces the risk of duplicate execution, and improves resource utilization efficiency.
Smart Images

Figure CN123068035A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of dynamic scheduling and execution consistency control technology for scheduled tasks in intelligent computing center operation and maintenance, specifically to a dynamic scheduling method, system, and intelligent computing center operation and maintenance platform for scheduled tasks. Background Technology
[0002] The intelligent computing center's inspections, backups, model synchronization, log archiving, and firmware checks are typically triggered at fixed times. These fixed times cannot reflect the phased resource usage of training jobs. When multiple tasks start simultaneously at the top of the hour, they will compete for management network, storage bandwidth, and CPU, thus affecting online computing tasks.
[0003] Existing schedulers mostly determine whether to start based on task priority and idle resources, without taking into account factors such as the allowable stale time of results, segmented execution, peak resource envelope, and the backlog caused by the previous delay. Furthermore, duplicate or missed executions may occur when switching control nodes.
[0004] Common scheduled task scheduling schemes mainly select the execution time based on priority, available resources, or fixed delay window, without simultaneously constraining the validity period of operation and maintenance results, the uniqueness of task side effects, the reuse of completed stages, and the validity of leases after controller switching.
[0005] Task definitions, resource telemetry, job phase predictions, and execution agent status originate from different interfaces. The nominal trigger time is not equivalent to the safe execution time; if only the current utilization rate is extended, the result may exceed the validity period, and during controller switching, two task instances may repeatedly submit the same external registration or backup result.
[0006] This application decomposes the task into a phase graph with resource envelopes, checkpoints, and side-effect commit points. It uses the age of valid results to form a freshness deficit, which is then used in conjunction with low-conflict resource intervals to solve for windows and nodes. The execution of a lease carries both the controller period and the task period; when feedback exceeds the limit, only uncommitted phases are migrated, while completed phases with consistent summaries continue to serve as input for the new period.
[0007] The nominal trigger time of a scheduled task is not directly equivalent to its execution time. The task scheduling contract simultaneously limits the validity period of the result, the latest completion boundary, stage side effects, and checkpoints. The scheduler uses freshness debt to constrain the movable range and then solves it together with the low-conflict resource interval; the dual-time lease prevents old controllers or old task instances from submitting results. Summary of the Invention
[0008] Fixed Cron times cannot reflect the differences in the training, inference, data loading, and checkpoint writing phases of intelligent computing jobs. If scheduled tasks are launched during storage checkpoints or management network peaks, they will compete with online jobs for hardware queues; simple delays may cause results to expire or lead to long-term starvation of low-priority tasks. Controller failover, execution lease expiration, and partial task submission can also cause side effects such as duplicate registration and uploads. This invention aims to create an executable dynamic scheduling closed loop that balances result freshness, physical resource capacity, completion deadlines, and the uniqueness of side effects.
[0009] The first aspect of this application provides a dynamic scheduling method for scheduled tasks, characterized by comprising: parsing the triggering period, result validity period, latest completion time, stage resource requirements, read / write objects, interruptible points, checkpoint interfaces, and side-effect submission points of a scheduled task; generating a task scheduling contract, wherein the stage resource requirements are represented by upper and lower bounds of computation, management network, and storage queue occupancy; calculating freshness deficit based on the age of the most recent valid result, target update interval, and number of consecutive delays, and determining a movable execution window by combining the upper bound of the predicted execution duration and the latest completion time; obtaining the upper and lower bounds of the occupancy of each stage of the online computing job in the dimensions of computation, management network, and storage; determining a continuous time interval in which the sum of the upper bound of online occupancy and the upper bound of the resource requirements of the corresponding stage of the scheduled task does not exceed the lower bound of resource capacity as a low-conflict resource interval; and forming candidate pairs from the movable execution window and the low-conflict resource interval. In this case, candidate solutions with an estimated completion time upper bound later than the latest completion time, execution nodes lacking checkpoint recovery conditions, or side effect stages not within the allowed window are eliminated. Then, the execution window, execution node, and segmentable boundary are selected based on the resource peak overlap cost, checkpoint and data migration cost, and freshness deficit at the time of candidate completion. An execution lease containing the controller period, task period, expiration time, stage checkpoint summary, stage idempotent key, and input summary is issued for the selected solution. The execution agent executes or submits the corresponding stage only when the lease verification passes and both periods are not lower than the locally seen valid values. The stage resource envelope and completion time prediction are updated based on the actual stage progress and resource usage. When the completion time prediction or resource usage reaches the rescheduling condition, a new task period is generated. Only stages that have not yet submitted side effects are rescheduled on a rolling basis, and completed stages whose output summary is consistent with the input summary of the new period are reused.
[0010] Furthermore, the task scheduling contract includes at least a task stability identifier, a nominal trigger time, a target update interval, a result validity period, a latest completion time, a phase diagram, resource envelopes for each phase, a read set, a write set, an interruptible point, a checkpoint location, a compensation action, and a side effect submission point; after the task is completed, the age of the most recent valid result is updated with the actual completion time, and the freshness deficit is not cleared to zero with the nominal trigger time.
[0011] Furthermore, let the age of the most recent valid result be A, the target update interval be T, and the number of consecutive delays be n. The current freshness deficit Dcurrent = max(0, AT) + λnT. For candidate solution x, the freshness deficit when the candidate is completed is Dfinish(x) = max(0, A + tfinish(x) - T) + λnT. The objective function is J(x) = αCconflict(x) + βCmigrate(x) + γDfinish(x). Among the candidate solutions that satisfy the hard constraints, the one with the smallest J(x) is selected.
[0012] Furthermore, for each resource dimension, the lower bound of resource capacity is Cavailable_low = Cphysical_low - Rcommitted - Emeasure - Ssafety, where Cphysical_low is the lower bound of measurable physical capacity, Rcommitted is the committed reservation, Emeasure is the upper bound of measurement error, and Ssafety is the safety margin. When the predicted interval width exceeds the prediction deviation boundary during the online operation phase, a short-term low-conflict interval of resources is formed based on the real-time lower bound of resource capacity, and the lease term is shortened.
[0013] Furthermore, the controller period is generated by the atomic increment operation stored in the persistent period; the controller generates an integrity authentication code for the controller identifier, controller period, task period, input digest, phase checkpoint digest, and expiration time; after the authentication code is verified, the execution agent persists the maximum controller period and the maximum task period that has been submitted. Execution leases with lower controller periods may not submit results even if they have not yet expired.
[0014] Furthermore, when a lease expires or a network partition occurs, the execution agent stops after setting checkpoints for phases that have no side effects and are interruptible. For phases that have already entered the side effect submission point, it only completes confirmation or performs predefined compensation without reallocation. After connectivity is restored, valid results are selected according to the controller period and task period, and results from lower periods are only used as audit records.
[0015] A second aspect of this application provides a dynamic scheduling system for timed tasks, characterized by comprising a contract generation module, a freshness calculation module, a resource interval prediction module, a joint selection module, a period lease execution module, and a rolling rescheduling module; the contract generation module is used to generate task scheduling contracts with phased resource envelopes and side-effect submission points; the freshness calculation module is used to calculate freshness deficit and movable execution windows; the resource interval prediction module is used to generate low-conflict resource intervals based on the upper and lower bounds of online computing power job occupancy and the lower bound of resource capacity; the joint selection module is used to perform late completion, resource capacity, and checkpoint recovery constraint filtering, and select an execution scheme based on resource peak overlap, migration cost, and freshness deficit at candidate completion; the period lease execution module is used to issue and verify execution leases containing controller periods and task periods; the rolling rescheduling module is used to only reorder unsubmitted stages and reuse completed stages with consistent digests when the prediction is completed or resource occupancy reaches the rescheduling condition.
[0016] A third aspect of this application provides an electronic device, characterized in that it includes a processor, a memory, a communication interface, and a computer program stored in the memory and capable of running on the processor, wherein the communication interface is used to connect a scheduling controller, a task execution agent, and a resource monitor, and the computer program, when executed by the processor, implements the method according to any one of claims 1 to 6.
[0017] A fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the method described in any one of claims 1 to 6.
[0018] A fifth aspect of this application provides a computer program product comprising a computer program, characterized in that, when executed by a processor, the computer program implements the method described in any one of claims 1 to 6.
[0019] Phase resource envelopes enable scheduling based on the measurable capacity boundaries of computing, management networks, and storage queues, while freshness deficit limits continuous delays. Dual-term leases isolate primary and backup controllers and multiple executions of the same task, and phase idempotent keys prevent duplicate side effects. Rolling rescheduling preserves completed technical outputs and only moves uncommitted phases, thereby reducing resource peak overlap without compromising result uniqueness. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the closed-loop dynamic scheduling of timed tasks. It shows that the task contract separately constrains the freshness deficit and the low resource conflict interval. Under hard constraints, the two types of results jointly select the execution plan, and the actual completion prediction and resource deviation feedback are used to recalculate the unsubmitted stage.
[0021] Figure 2 This is a schematic diagram of the structure of the timed task dynamic scheduling system and the intelligent computing center operation and maintenance platform. It shows the joint selection module for contract generation, freshness calculation and resource range prediction as common inputs. The completion prediction and resource deviation generated by the time lease execution module are then fed back to the corresponding calculation module.
[0022] Figure labeling: Contract generation module 11, Freshness calculation module 12, Resource range prediction module 13, Joint selection module 14, Term lease execution module 15, Rolling rescheduling module 16. Detailed Implementation
[0023] The following implementation method uses the task phase and side effect submission point as the execution boundary to explain the specific handling of freshness deficit, resource capacity range, time lease and checkpoint reuse.
[0024] The “Phase Resource Envelope” records the upper and lower bounds of the task phase’s requirements in terms of CPU or accelerator computation, management network bandwidth, storage queue usage, and duration. For each resource dimension, the “Resource Capacity Lower Bound” Cavailable_low = Cphysical_low - Rcommitted - Emeasure - Ssafety is a conservative value of the measurable physical capacity lower bound minus the committed reservation, measurement error upper bound, and safety margin. The “Resource Low Conflict Interval” is a continuous time interval in which the sum of the online job usage upper bound and the scheduled task requirement upper bound does not exceed Cavailable_low in all critical resource dimensions. The safety margin is not deducted repeatedly during comparison.
[0025] A task scheduling contract should include at least a task stability identifier, nominal trigger time, target update interval, result validity period, latest completion time, phase graph, resource envelope, read / write set, breakable point, checkpoint interface, side effect commit point, idempotent key, and compensation action. An execution lease should include at least a controller phase, task phase, issuance time, expiration time, phase checkpoint summary, phase idempotent key, and input summary; both structures use append-only versions.
[0026] Task scheduling contracts can be in the following states: pending trigger, movable, leased, executing, phase committed, pending rescheduling, expired, and compensating. Candidate solutions must first pass four types of hard constraints: latest completion, resource capacity, checkpoint recovery, and side effect window, before participating in cost ranking. If no feasible solution exists, conflict constraints are output and the scheduling state is retained, without forcibly executing by reducing safety margins or ignoring the validity period of the result.
[0027] The persistent period storage uses atomic comparisons to exchange incrementing controller periods; when multiple controllers are deployed, the period value is only committed and returned by the storage that receives majority confirmation. The controller serializes the controller identifier, controller period, task period, input digest, stage checkpoint digest, and expiration time in a fixed field order, and generates an integrity authentication code using a pre-configured controller identity key. The execution agent first verifies the authentication code and controller identity; only after successful verification does it persist the maximum known controller period, then compares the task period and checks the stage checkpoint digest with the current time; execution and commit are rejected if authentication fails, any period is lower than the locally known value, the lease has expired, or the digest is inconsistent. Renewal only takes effect before the original lease expires, provided the input digest and checkpoint digest remain unchanged, and the controller period is the same; new task periods can only reuse completed stages whose output digest matches the new period's input.
[0028] Let A be the age of the most recent valid result, T be the target update interval, and n be the number of consecutive delays. The current freshness deficit Dcurrent = max(0, AT) + λnT. λ is pre-configured according to the task category and written into the contract parameter snapshot; A is recalculated from the actual completion time after the task is completed. For candidate solution x, let tfinish(x) represent the upper bound of the duration from the candidate evaluation time to the expected completion time. The freshness deficit Dfinish(x) when the candidate is completed is Dfinish(x) = max(0, A + tfinish(x) - T) + λnT; the evaluation function is J(x) = αCconflict(x) + βCmigrate(x) + γDfinish(x), where Cconflict is the peak overlap between the resource envelope and the online job occupancy interval, Cmigrate is the cost of checkpoint recovery and cross-node data movement, and α, β, and γ are all non-negative weights.
[0029] Each cost component is normalized within the effective upper and lower bounds of the candidate set in this round or within a pre-fixed physical upper bound; when the upper and lower bounds of a component are equal, the normalized value of that component is zero. Let Oobserved_high be the sum of the observed online occupancy upper bound and the actual task occupancy upper bound. When Oobserved_high > Cavailable_low for Nstop sampling periods in the parameter snapshot consecutively, or when there is a hard error such as resource interface report queue overflow, device unschedulability, or storage write failure, the resource conflict stopping condition is met; when a single measurement spike occurs and the next sampling resumes, only the envelope is updated, and stopping is not triggered. α, β, γ, Ssafety, prediction bias boundary, Nstop, sampling period, and lease term are fixed with the contract version before candidate evaluation and cannot be adjusted after seeing the candidate results.
[0030] When a prediction interval is missing or its width exceeds the prediction deviation boundary, a short-term execution interval is formed only based on the lower bound of the real-time resource capacity, and the lease is shortened. If the latest completion time cannot be met, the system remains pending scheduling or manual intervention is requested. During network partitioning, each partition can only renew its existing leases that have not expired, and no new cross-partition side effects are issued. After the lease expires, phases without side effects stop at the checkpoint, and phases that have entered the side effect submission point only complete confirmation or execute predefined compensation.
[0031] The configuration backup task includes read, compression, upload, and registration phases. After read and compression are completed, the online training job enters the checkpoint write phase, causing the sum of the upper bound of the storage queue occupancy and the upper bound of the upload phase demand to exceed the lower bound of the resource capacity. The system retains the output digests of read and compression, only moving uncommitted upload and registration phases to the next low-conflict resource interval; if the registration action has already been committed, it is not reassigned, but confirmed along the original node or compensated according to the contract. This example does not depend on specific bandwidth values.
[0032] The resource monitor aggregates the upper bound of online job occupancy and the upper bound of scheduled task occupancy by the same resource dimension and sampling period. When N consecutive sampling periods exceed Cavailable_low or the resource interface reports a hard error, the execution agent forms a checkpoint at the most recent breakable point and triggers a new task period. The new controller must first obtain a higher period from the persistent majority period store and generate a verifiable integrity authentication code. Leases with old periods or failed authentication cannot be continued to be submitted because they have not yet expired.
[0033] The following combination Figure 1 and Figure 2 This explains the processing boundaries for freshness deficit, resource range, dual-term leases, and rolling rescheduling.
[0034] Figure 1 This represents the computation and feedback boundaries within a single rolling scheduling period. Task contracts are only updated when the definition or side effect boundaries change, and execution deviations typically only update the selection results for the backlog, resource range, and uncommitted phases.
[0035] Step S100 breaks down the scheduled task into a phase diagram, registering the read / write objects, computational load, network bandwidth management, upper and lower bounds of storage queue occupancy, interruptible points, and compensation actions for each phase. The nominal trigger time, result validity period, and latest completion time are written into the same contract version; the side effect phase is also bound to a unique idempotent key and an allowed execution window.
[0036] Step S200 calculates the result age and freshness deficit from the most recently verified actual completion time, and uses the upper bound of the predicted execution duration to infer the latest start boundary. The number of consecutive delays only increases the deficit and does not change the latest completion time; stages with submitted side effects are not included in the movable set, only stages that have not yet been submitted can be rearranged from the checkpoint.
[0037] Step S300 predicts the upper and lower bounds of computation, management network, and storage occupancy according to the online operation stage, and deducts submitted reservations, measurement errors, and safety margins from the lower bound of physical capacity. Continuous intervals where all critical resources meet capacity constraints constitute low-conflict intervals; if the prediction is too wide, only short intervals based on real-time measurements are generated and the leases are shortened.
[0038] Step S400 combines movable windows and low-conflict intervals into candidate solutions. It first checks the latest completion time, checkpoint recoverability, and side-effect windows, then calculates resource peak overlap, data migration, and freshness deficit at candidate completion. Solutions that fail the hard constraints are not ranked; in cases of complete tie, the task stability identifier determines the final result.
[0039] Step S500 retrieves the new controller period from the majority period store and generates the task period for this task. The execution agent verifies the lease authentication code, expiration time, two periods, and phase checkpoint digests; leases with low controller periods or whose input digests do not match the checkpoints cannot be executed or committed, even if their nominal terms have not yet expired.
[0040] Step S600 continuously compares the actual stage progress, resource usage, and predicted completion time. When the resource conflict stopping condition is met or the expected completion upper bound is exceeded, the agent sets a checkpoint at the nearest breakable point, the controller generates a new task period and only reorders uncommitted stages; completed stages whose output summary matches the new input are directly reused.
[0041] Task contract phases, resource forecasts, leases, and execution feedback are associated with the input summary via the parent version. When job phases or resource measurements change, only the affected phases are re-entered into the candidate set; completed phases with submitted side effects and no dependencies remain confirmed, and the results log saves the reason for removal and the next feasible window.
[0042] Furthermore, the task scheduling contract also records the read / write objects, interruptible points, and compensation actions for each execution phase.
[0043] Furthermore, the freshness deficit is determined based on the age of the valid result, the target update interval, and the number of consecutive delays, and is recalculated according to the actual completion time after a new result is completed.
[0044] Furthermore, the low-conflict resource interval is a continuous time interval in which the sum of the upper bound of online job occupancy and the upper bound of timed task demand does not exceed the lower bound of resource capacity after deducting reserves, measurement errors, and safety margins.
[0045] Furthermore, when jointly selecting the execution period and location, candidate solutions that cannot meet the latest completion time are first eliminated, and then the resource peak overlap and checkpoint migration cost are minimized.
[0046] Furthermore, the execution lease includes both controller period and task period, and the execution agent refuses to submit results below the locally seen period.
[0047] Furthermore, rolling rescheduling only migrates stages that have not yet committed side effects, and reuses completed stages via idempotent keys and output summaries.
[0048] The method organizes step inputs and state changes around a task scheduling contract, and avoids mixing old and new records through processing periods. Specifically, the processing includes: organizing the trigger period, result validity period, resource envelope, and side effect boundaries into a scheduling contract; jointly selecting the execution period, node, and segment boundary using freshness deficit and low-conflict intervals with uncertainty; and maintaining side effect uniqueness through dual-period execution leases, checkpoints, and rolling rescheduling. The state "Confirmed" corresponds to the existence of direct technical observations or mutually independent indirect observations with consistent input summaries; the state "Unknown" indicates that the evidence gap has not been eliminated, but is not equivalent to the condition not being met.
[0049] The entry point obtains the trigger cycle and result validity period, read / write objects and resource envelopes, and predicts and manages the network and storage queues for online job phases. Interface requests carry object stability identifiers, expected versions, calling permissions, and idempotent keys, while responses carry actual versions, collection time intervals, quality flags, and content summaries. Data from caches, mirrors, or derivations is stored from its direct source; unobtainable fields use explicit unknown values, and missing items are stored with explicit unknown markers, without using default zero values or old values to fill in the gaps.
[0050] The calculation parameters, thresholds, and weights are fixed simultaneously according to the task scheduling contract version. The thresholds can be determined by the allowable range of equipment specifications, the quantile of the field-calibrated sample, the upper limit of historical normal window fluctuations, and the task delay budget or risk budget; the parameters are not changed within the same processing period. When the sample is insufficient or the denominator is zero, the calculation result is marked as uncalculated, and this item is not replaced by a fixed value lacking a source.
[0051] After obtaining the task scheduling contract, the control side first performs checks on field type, object existence, version continuity, permission scope, and integrity summary. Then, it calls the generation of the task scheduling contract to drive the rolling rescheduling of the corresponding processor. Each module appends the results according to field permissions without directly rewriting the source data. When rules or topology need to be corrected, the next version is generated, and the set of affected objects is retrieved from the dependency index.
[0052] The preceding and following steps complete the data handover using structured objects. The preceding step outputs the freshness deficit or movable execution window and writes it to the parent version field. The following step re-verifies that the parent version is still the current valid version before generating the rolling rescheduling result. If a task execution timeout occurs during this process, only the branches that depend on the failed interface are frozen, while other unrelated branches remain in the confirmed state.
[0053] After generating candidates, results that violate hard constraints are first eliminated, and then the remaining results are sorted: The first level is based on eliminating candidates that cannot meet the latest completion time, prohibiting the migration and redoing of stages with submitted side effects, and establishing a feasible set of hard constraints by simultaneously verifying the controller period and task period of the lease; the second level sorts only within the feasible set based on risk, latency, resource consumption, or evidence completeness. If there are no executable candidates, conflicting constraints are listed in the results, and the safety boundary does not change due to sorting requirements.
[0054] Before committing, the system rereads the completed checkpoint and side effect summaries and the current task scheduling contract header version, comparing them with the input summaries saved during the preparation phase. Committing is allowed if the summaries match; if the summaries change but do not affect the current dependency branches, confirmed outputs are reused; if a summary change hits a dependency index, only the affected branches are recalculated, and unaffected branches are not rolled back entirely.
[0055] The scheduling result interface returns the contract version, execution window, node, phase boundary, dual-period lease, and unmet constraints. The execution agent only accepts results that are not yet due in the current controller period; if there is no feasible window, it keeps the resource pending scheduling and reports conflicting resources, without forcibly starting by reducing the safety margin or ignoring the validity period of the result.
[0056] The task scheduling contract is a structured record that runs through the processing chain of this invention, and includes at least a task stability identifier, trigger period, earliest start time, latest completion time, result validity period, resource envelope, segmentable flag, idempotent key, checkpoint, scheduling period, and lease status. Specifically, the result validity period limits the maximum time the task result can be used for subsequent operational decisions; the resource envelope records the upper and lower bounds of CPU, storage, and network requirements for each execution stage; and the scheduling period distinguishes between valid decisions before and after rescheduling.
[0057] Freshness deficit represents the cumulative deviation of the current result's age from the target update interval, and it increases monotonically when the task is continuously delayed; after the task is completed, it is recalculated based on the actual completion time rather than simply being reset to zero.
[0058] Load phase is the joint state of resources across stages such as training, inference, data loading, or checkpoint writing, and is represented using phase with prediction intervals rather than a single utilization rate.
[0059] The scheduling system is deployed in the intelligent computing center's operation and maintenance control domain, receiving scheduled task definitions, job scheduler load, storage and network telemetry, and maintenance windows. Task execution agents reside on management nodes or computing power nodes, and these agents report checkpoints using scheduling periods and idempotent keys.
[0060] In step S100, a task scheduling contract is generated. The controller breaks down the script or workflow into retryable units, recording the read / write objects, maximum execution duration, interruptible points, and compensation actions for each unit; and binds unique idempotent keys to non-repeatable external writes.
[0061] In step S200, the freshness deficit is calculated. The system accumulates the result age starting from the completion time of the valid result, and subtracts the predicted execution time from the result's validity period to obtain the latest start boundary; continuous delays increase the deficit weight, so that low-priority tasks will not be permanently starved.
[0062] In step S300, the prediction of low-conflict resource intervals is performed. Upper and lower bounds of occupancy are generated for each resource dimension, and then the intervals are intersected with the task resource envelope; candidate intervals are formed only when all key dimensions have available capacity.
[0063] When the prediction interval is missing, the scheduler forms a short-time interval based only on the lower bound of the real-time resource capacity and shortens the execution lease accordingly; this short-time interval still needs to meet the latest completion time and side effect window constraints.
[0064] In step S400, a joint selection of execution period and location is performed. Candidate schemes are first filtered according to the latest completion constraint, and then the resource peak overlap between each task stage and the online job stage is minimized; segmentable tasks are allowed to be executed across two candidate intervals, but checkpoint writes are included in the storage resource envelope.
[0065] In step S500, the issuance and execution of the time-limited lease are performed. The agent only accepts instructions with a scheduling period not less than the locally known time and whose leases have not expired. Before executing side effects, the idempotent key is registered. After the control node switches, the registration status is read to determine whether to continue, compensate, or skip.
[0066] In step S600, feedback bias is performed and rolling rescheduling is initiated. Submitted side effects are not repeated in the new period; unsubmitted phases are migrated from the most recent checkpoint. The residual between actual occupancy and the predicted interval is used to update the profile of similar tasks, while retaining historical versions for review.
[0067] J(x)=α·Cconflict(x)+β·Cmigrate(x)+γ·Dfinish(x)
[0068] In the formula, x represents the candidate execution scheme, Cconflict(x) represents the peak overlap between the upper bound of the stage resource demand and the online job occupancy interval, Cmigrate(x) represents the migration cost formed by the amount of data recovered at the checkpoint and the amount of movement across nodes, and Dfinish(x) represents the freshness deficit when the candidate is completed. Each component is normalized according to the boundary fixed before scheduling and multiplied by non-negative weights α, β and γ. The scheme with the smallest J(x) is selected from the candidates that satisfy the hard constraints.
[0069] When multiple periodic tasks share resources, first fix the stages with external time limits according to the immovable boundary, and then use the resource peak phase misalignment rule to arrange the remaining stages; the phase misalignment amount should at least cover the rising edge of resource occupation and measurement uncertainty.
[0070] The lease includes both the controller phase and the task phase. The agent stores a summary of the output of completed phases and only releases uncommitted resources when it receives instructions for a newer phase, preventing two controllers from advancing the same side effect simultaneously.
[0071] Example 2 illustrates a scenario where three types of periodic tasks—asset inventory, configuration backup, and log archiving—are deployed alongside online training jobs. The validity period of the results, network traffic, storage writes, and whether segmentation is allowed differ among these three task types, meaning the actual start time cannot be determined solely by the Cron expression.
[0072] The task scheduling contract records the task identifier, nominal trigger time, result validity period, latest completion time, phase diagram, resource envelopes for each phase, read / write objects, checkpoint interfaces, and side effect commit points. Configuration backup tasks can be paused before the upload is complete, while external registration actions for asset inventory, once submitted, can only be revoked through compensation records.
[0073] Let A be the age of the most recent valid result, T be the target update interval, and n be the number of consecutive delays. Then, the freshness deficit can be represented as D = max(0,AT) + λnT. λ is used to gradually transform multiple delays into hard constraints and is pre-configured according to the task category. After a task generates a new result, A is recalculated based on the actual completion time, and the deficit cannot be cleared by the nominal trigger time.
[0074] Resource prediction uses interval envelopes. For a candidate period w, the upper and lower bounds of computation, management network, and storage usage for online jobs are summarized, and then intersected with the resource envelopes of each stage of the scheduled task. If the prediction interval width is too large, it is not misclassified as low-conflict; instead, the conflict cost is increased or the executable segment is shortened.
[0075] Candidate solutions first undergo feasibility screening: the expected completion time must not be later than the latest completion time, the execution node must have checkpoint recovery conditions, and the side effect phase must fall within the allowed window. Feasible solutions then calculate the comprehensive cost based on freshness deficit, peak overlap, data migration volume, and checkpoint recovery volume.
[0076] Controller periods are used to isolate the switching of the primary and backup controllers, while task periods are used to isolate multiple executions of the same task. The execution agent stores the maximum controller period and the submitted task periods; the old controller cannot submit results of lower periods even if the lease has not yet reached its nominal term, thus avoiding concurrent writes between the two primary controllers.
[0077] For example, the upload phase of a configuration backup task requires 40%–55% of the storage queue, with a lower bound of 80% resource capacity. The upper bound of storage usage during the online training checkpoint phase is 45%, and the sum of these upper bounds is 100%, which does not constitute a low-conflict interval. After training enters the computation phase, the upper bound of storage usage drops to 20%, and the sum of these upper bounds is 75%. The scheduler retains the completed read and compressed digests and only moves the upload and registration phases to this interval. The percentages mentioned above are used to illustrate the interval comparison process.
[0078] If the actual progress of the task is slower than the predicted completion time, but the remaining stages can still be completed before the latest completion time, only the resource envelope is updated; rolling rescheduling only occurs when the expected completion upper bound is exceeded or resource conflicts reach the stopping condition. Stages that have already been registered externally are not migrated, but are confirmed or compensated along the original node.
[0079] The platform saves a summary of each selection, including the candidate window, reason for rejection, resource prediction version, lease, and stage output. When reproducing the problem, it should be solved again using the same resource period and task contract, yielding the same feasible set. When comparing with a fixed trigger scheme, the platform verifies the delay limit, peak resource overlap, and number of duplicate submissions.
[0080] In terms of data structure, each task scheduling contract includes at least a stage resource envelope, read / write objects, checkpoints, completion prediction, freshness deficit, and lease summary. The business identity and the database auto-incrementing primary key are stored separately; the business identity is determined by the task identifier, controller period, task period, and stage idempotent key, while the auto-incrementing primary key is only used for physical location. Updates use append-only versions, with the previous version recording its termination reason and remaining readable, thus enabling differentiation between object changes, repeated data collection, and reprocessing of the same object.
[0081] After a task contract is established, it is in a pending trigger state; after the movable window is calculated, it becomes movable; after a dual-period lease is issued, it becomes leased; and at the start of the proxy phase, it is in execution. A phase is marked as committed after being written to an external object and confirmed by a readback. Phases that predict out-of-bounds events but have not yet been committed enter a phase awaiting reordering; phases that exceed the latest completion time or enter compensation phases are marked as expired and in compensation, respectively. Each transition carries a controller period, task period, and checkpoint summary; the recovery end only accepts records with periods not lower than the maximum seen value.
[0082] External registration, backup object writes, or completion markers use a prepare-execute-acknowledge log. The prepare log stores the expected write set, the stage idempotent key, and the compensation entry; the execution acknowledgment must be consistent with the object readback result. If the acknowledgment times out and the object status is unclear, the stage remains in the "commit unclear" state and is not reallocated to other nodes.
[0083] The idempotent index for a stage consists of a task identifier, controller period, task period, stage identifier, and input digest. When the same index is arrived repeatedly, confirmed outputs are reused; if the input digests are different, overwriting is rejected. When multiple agents compete for the same stage, only the agent whose conditional write is successful is eligible to commit.
[0084] The nominal trigger time, movable window, lease term, and latest completion time are all represented using a unified clock reference and error limits. Before the agent executes, it confirms that the current time is still within the lease interval. When the controller re-plans, it closes the old window with the new task period, and late results can only be recorded in the audit log.
[0085] The validity period and target update interval are derived from the task contract, and the freshness coefficient is used to convert consecutive delays into outstanding debts. The upper limit of resource conflicts is the remaining amount after deducting reservations, measurement errors, and safety margins from the lower limit of physical capacity. The prediction deviation boundary is formed based on historical errors of similar stages, and the lease term is not shorter than the upper limit of the stage confirmation delay. A single solution uses the same parameter snapshot, and parameter adjustments only affect subsequent task periods and cannot change the feasibility conclusions of already issued leases.
[0086] The recovery process follows these technical boundaries: reuse the output summary of completed stages, only undo unclosed side effect stages, and continue from the most recent checkpoint. Each rollback request has an independent action number; upon successful execution, a recovery observation and a post-recovery summary are appended. If a step cannot be recovered, subsequent compensation dependent on that step is discontinued, on-site evidence is preserved, and the object is placed in isolation to prevent null values from masquerading as recovery completion.
[0087] A smart computing center implementation includes a scheduling controller, a task execution agent, a resource monitor, and an operation and maintenance platform. These components exchange resource envelopes, leases, and object readback results through structured contracts with length, mode version, and checksum values. When optional fields are unknown, the original message is retained for further processing; when required fields are unknown, the period is rolled back, or the digest check fails, the request is rejected and the supported contract version is returned.
[0088] On the same task set with known task cycles, result validity periods, and resource peaks and troughs, a comparison is made between fixed Cron triggering, delay based solely on resource load, and contract scheduling of this invention. The number of repeated executions when valid results have not expired, the number of deadline defaults, the number of resource conflicts, the number of lease re-executions, and the next window error after deviation feedback are recorded.
[0089] After the scheduler controller recovers, it first reads the most recent complete contract checkpoint, then scans for unconfirmed leases and phase commit records. Confirmed output summaries are restored as reusable inputs; only side effect phases that need to be recorded first query external objects, and based on the query results, confirm, compensate, or maintain isolation, without directly re-executing.
[0090] The computational usage, management network bandwidth, and storage queue depth returned by the resource monitor are first preserved in their original dimensions, and then converted into capacity ratios according to the contract. Checkpoints, completion predictions, and lease summaries save the generated components and versions. Resource values that cannot be converted remain unknown and their corresponding usage ranges are expanded. When multiple sources are inconsistent, the envelope union is used instead of averaging. If the upper bound of the union exceeds the lower bound of the capacity, the relevant candidate windows are directly judged as infeasible.
[0091] When two candidate windows have the same objective function value, the one with the larger latest completion margin is selected first, followed by a comparison of the amount of data to be moved and the number of reusable checkpoints; if they are still the same, a fixed order is obtained based on the node stability identifier. The scheduling record retains the resource upper bound, freshness deficit, migration cost, and reason for removal for each candidate, so that the same contract, resource period, and parameter snapshot can reproduce the same selection.
[0092] During task contract persistence, the task stability identifier, controller period, and task period are written into the header. Then, the resource envelope, read / write objects, checkpoints, and completion predictions are encoded in stage topology order. Units, mode versions, and lease summaries are all included in the content digest. The display name is not used as the identity of the task or resource. Therefore, the execution agent can independently recalculate the digest after receiving a lease and reject leases formed from old checkpoints or foreign objects with the same name.
[0093] The contract version is calculated by encoding the task stability identifier, controller period, task period, phase input summary, and parameter snapshot summary by field length. The execution agent also compares the phase checkpoint summary with the pattern version to prevent old encodings or old checkpoints from generating contracts that appear identical.
[0094] Task input quality is recorded for source authenticity, time integrity, field integrity, and cross-consistency. Resource prediction only checks the telemetry component, while side effect submission also checks the object readback and permission components; when any component changes, only the stages and candidate windows referencing that component are recalculated.
[0095] When a field is missing, the impact range is determined based on the stage read set. Fields unrelated to the current stage only form quality markers; if the upper bound of resources, checkpoint summary, or readback results of side effect objects are missing, the corresponding candidate window or submission action is paused. The supplementary acquisition message carries the task stability identifier, controller period, task period, and stage idempotency key. Upon return, only the relevant stages are re-predicted; confirmed stages are not rerun, and unknown resource values are not treated as zero occupancy.
[0096] Each side effect phase stores preparation, invocation, object readback, and compensation records in the transaction catalog. If the acknowledgment is lost but the object readback shows that the write has been completed, the phase is acknowledged and the output summary is reused; if the object status is still unclear, retransmission is prohibited and the process is handled by a higher-level controller after the lease expires.
[0097] The task execution queue is partitioned by task stability identifier and task duration. Within the same phase, only agents holding valid high-term leases can write to the idempotent key. When the queue is congested, lease revocation, object readback, checkpoint completion, and compensation confirmation are preserved. Duplicate resource sampling in phases not yet started can be compressed within the same resource duration, but must not cross the nominal trigger time, lease expiration point, or latest completion time.
[0098] The recovery test sequentially simulates lease duplication, delayed messages from the old controller, agent restart after a checkpoint, object write receipt timeout, and resource mode version upgrade. It verifies whether commits are consistently disabled during the low-controller period, whether confirmed phases are reused, and whether retransmission of side-effect phases with unclear states is stopped. Furthermore, if compensation is not yet confirmed, the controller needs to be switched again. The new controller should continue processing unclosed phases from the most recent checkpoint, without revoking committed phases with consistent object readbacks.
[0099] Boundary tests are set up with candidates that exactly meet the latest completion time, resource upper bound equals capacity lower bound, lease expires, and a missing checkpoint summary. Verification includes not only whether the task started, but also the reason for removal, task duration, reusable phase, and side effect submission status.
[0100] Interfaces involving external action permissions use the least privilege credential. The credential itself is not written into the task scheduling contract; only an irreversible credential reference summary, permission scope, and audit number are recorded. If the executing entity is not authorized to complete external registration, backup upload completion, or other task stages with side effects, the process terminates during the preparation phase, preventing permission errors from being mistakenly identified as target object failures. Audit records and business records are associated using the same object key and processing period, but access is controlled separately. When exporting implementation data, a stable anonymous identifier replaces the real asset identifier.
[0101] Completed continuous contract versions that have exceeded the audit period can be compressed into phase checkpoints, but the controller phase, task phase, state transition summary chain, and external object location must be preserved. Phase records that are unclear in commit, in compensation, or still referenced by new task phases must not be compressed.
[0102] When replicating the example, first fix the task contract, the most recent valid result, and the resource monitoring period, and calculate the occupancy range and freshness deficit for each stage; the scheduler generates candidate windows and completes hard constraint filtering, then issues leases containing controller periods and task periods. The agent submits checkpoints and object readbacks stage by stage, and only updates stages that have not yet started when the actual occupancy deviates from the prediction. If a higher controller period is generated, the results of the old period are retained for auditing purposes, and the confirmed outputs are used as reusable inputs for the new period.
[0103] The scheduling controller can use a centralized database, consensus log, or other storage that can atomically increment the controller's epochs, and the execution queue can also be replaced with a different implementation. Regardless of the carrier used, task stability identifiers, dual epochs, phase idempotent keys, and latest completion boundaries must be maintained, and side-effect commits must also undergo object readback; basic components that do not have these semantics should add conditional writes, lease verification, and phase checkpoint records, and cannot be replaced by ordinary Cron locks.
[0104] Example 3 illustrates the input, state changes, and exception handling by taking the model checkpoint backup task as an example where the storage queue is congested due to increased training job writes, and it is necessary to migrate the uncommitted stage and reuse the completed data scanning stage.
[0105] Take, for example, three tasks: daily model repository verification, log archiving every four hours, and hourly liquid cooling inspection, all running in parallel with a nighttime training job.
[0106] Step 1: Establish the validity period of the results, the resource envelope of each stage, and the idempotent key for the three tasks.
[0107] Step 2: Calculate the freshness deficit of liquid cooling inspection based on the most recent valid results, followed by log archiving.
[0108] Step 3: Predict the storage bandwidth usage of training jobs during the checkpoint phase, while the network load will be lower during the subsequent computation phase.
[0109] Step 4: Schedule the liquid cooling inspection in the current short window, move the log archiving to the calculation stage, and split the warehouse verification into two stages: scanning and summary submission.
[0110] Step 5: When the control node is switched, the new controller issues a longer-term lease; the agent identifies the submitted scan results and only continues to run the summary submission.
[0111] Step 6: After the actual training phase ends ahead of schedule, the system updates the low-conflict interval but does not repeat the completed tasks.
[0112] The following describes the boundary states and the corresponding recovery process.
[0113] (1) When the agent loses contact, wait for the lease to expire and then reallocate. After the old agent is restored, the results cannot be submitted due to the low time period.
[0114] (2) When the prediction error leads to insufficient resource reserves, a checkpoint is generated at the interruptible point; in the non-interruptible stage, execution continues according to the latest completion constraint and new tasks are restricted from entering.
[0115] (3) When the task result is corrupted, the process is not considered successful as completion, but is confirmed by both the output summary and independent reading verification.
[0116] (4) When the task definition changes, a new contract version is generated. The old version has already executed side effects, which are audited and processed according to the compensation relationship.
[0117] (1) The low conflict interval of resources can be obtained by quantile prediction, state space model or regularization stage, as long as the upper and lower bounds are output.
[0118] (2) Joint selection can be achieved by constraint programming, integer programming or task-by-task heuristic search, all of which must first satisfy the latest completion hard constraint.
[0119] (3) The execution agent can be implemented through containers, system services or programmable controllers, and the idempotent key and epoch comparison rules remain unchanged.
[0120] Figure 2 The long-term unchanging task constraints, the freshness status that changes with the age of the results, and the resource profile that changes with online tasks are processed separately; the three types of data are only combined in the joint selection module 14 to avoid the resource predictor directly rewriting the task time limit.
[0121] Contract generation module 11 solidifies the period, result validity period, stage resource envelope, read / write objects, breakable points, and side effect commit points into a contract version Cn, which is used by modules 12 to 15.
[0122] The freshness calculation module 12 calculates the outstanding amount and movable window based on the most recently independently verified valid result, the latest completion time, and the number of consecutive delays. If the process returns a success but the output is not verified, the result age is not refreshed.
[0123] The resource interval prediction module 13 provides the upper and lower bounds of computing, management network and storage queue occupancy respectively, and forms a low-conflict interval by combining the maintenance window; when predicting crossing the boundary of the online operation stage, it is divided into different resource periods.
[0124] The joint selection module 14 first eliminates schemes that violate the latest completion, capacity lower bound, or checkpoint recovery constraints, and then compares resource overlap, migration cost, and freshness deficit at candidate completion among the remaining schemes.
[0125] The lease execution module 15 issues leases for selected schemes, including controller period, task period, phase idempotent key, lease term, and checkpoint summary; the execution agent can only submit side effects if neither period is behind.
[0126] The rolling rescheduling module 16 reads the actual stage progress, resource usage, and output readback results, and updates the completion time and resource envelope; when the deviation exceeds the limit, the submitted stage is retained, and new task periods are generated only for stages that have not started or have been checked.
[0127] Figure 2 The two feedback paths in the process have different meanings: the prediction is returned to module 12 to recalculate the result age, and the resource usage deviation is returned to module 13 to update the interval; the two are processed by module 14 to form the next task period, without triggering module 11 to unconditionally rebuild the contract.
[0128] During deployment, the scheduler saves the contract and the two periods, the resource monitor only submits the upper and lower bounds of occupancy with the observation interval, and the execution agent uses the phase idempotency key and object readback to prove whether the side effects have been committed.
[0129] The operations and maintenance platform stores contracts and scheduling periods in persistent storage with comparison and exchange semantics, while the agent only registers idempotent keys and phase summaries through a restricted interface.
[0130] When managing network partitions, each partition can only renew existing leases that have not expired, and no new cross-partition side effects are created; after connectivity is restored, the results are merged according to the period.
[0131] Construct scenarios with fixed cycles, sudden load increases, controller switching, agent disconnection, and task phase time deviations, and compare fixed-time scheduling with the rolling scheduling of this invention.
[0132] The scheduling effect is primarily measured by whether the task produces a verified result before the latest completion time, rather than by whether the process starts.
[0133] Resource interference records the envelope overlap integral of operation and maintenance tasks and online jobs across the computing, management network, and storage dimensions;
[0134] In controller switching and agent recovery scenarios, the number of times the idempotent key was repeatedly submitted as a side effect during the same phase is counted separately.
[0135] Timeliness of results is represented by the cumulative backlog of task results exceeding the target interval and their distribution.
[0136] Resilience is represented by the number of stages from the most recent valid checkpoint to the resubmission of the validated result after an interruption.
[0137] The intelligent computing center's scheduling controller, task execution agents, and resource monitor collaborate through a structured interface on the operations and maintenance platform. Records are stored using a combination of task stability identifier, controller period, task period, and phase idempotent key as identities, respectively saving the nominal trigger time, movable window, lease term, checkpoint summary, and object readback results; no agent can bypass the dual-period verification to directly submit results.
[0138] Figure 2 The module shown can be deployed as a service within the intelligent computing center's operation and maintenance platform, and module 15 can also be deployed to the execution node agent; regardless of the deployment location, the contract version, dual-episode and phase idempotent keys remain unchanged when transmitted across nodes.
[0139] The contract generation module 11 is used to parse the triggering period, result validity period, resource requirements and side effect boundaries of the scheduled task, and generate a task scheduling contract.
[0140] The freshness calculation module 12 is used to calculate the freshness deficit and movable execution window based on the most recent valid result, the latest completion time, and the number of consecutive delays.
[0141] The resource interval prediction module 13 is used to predict low-conflict resource intervals with uncertainty based on the computing power operation stage, management network, storage queue and maintenance window.
[0142] The joint selection module 14 is used to select the task execution period, execution node and segmentable boundary under the constraints of latest completion, resource capacity and checkpoint recovery, with the objective function consisting of resource conflict cost, migration cost and freshness deficit.
[0143] The period lease execution module 15 is used to issue an execution lease for the selected scheme, which includes the controller period, task period, lease term, phase checkpoint summary and idempotent key, and drive task execution based on the checkpoint.
[0144] The rolling rescheduling module 16 is used to collect actual stage progress and resource usage, update resource envelope and completion time prediction, and reschedule unsubmitted stages when deviations exceed limits.
[0145] Electronic devices that carry scheduling controllers or execution agents include processors, memory, and communication interfaces. The memory stores task contracts, resource periods, dual-period leases, and phase checkpoints; the processor compares the controller period with the task period and then executes or rejects the phase; the communication interface returns the current phase state without generating side effects upon receiving an old lease or a repeated idempotent key.
[0146] The embodiments listed in this specification are not exhaustive. Any equivalent implementations that can be made by those skilled in the art, as long as they retain the essential technical features described in the claims and their interrelationships, fall within the scope of protection of this invention.
[0147] The above implementation methods are used to illustrate the implementation relationship between task contracts, freshness deficit, resource intervals, and dual-period leases; when changing prediction algorithms, storage, or execution agents, the boundaries of hard constraints first, low-period rejection, and non-repeated execution of committed phases must still be maintained.
Claims
1. A method for dynamic scheduling of timed tasks, characterized in that, include: Analyze the time limit, phased resource requirements, interruptible points, checkpoint interfaces, and side effect submission points of the scheduled task to generate a task scheduling contract; The freshness deficit is calculated based on the age of the most recent valid results, the target update interval, and the number of consecutive delays. A movable execution window is determined by combining the predicted execution duration and the latest completion time. The upper and lower bounds of online computing power job occupancy in the computing, management network, and storage dimensions are obtained. A continuous interval where the sum of the upper bound of online occupancy and the upper bound of the timed task stage demand does not exceed the lower bound of resource capacity is defined as a low-conflict resource interval. Candidate schemes are formed from the movable execution window and the low-conflict resource interval. Schemes that do not meet the constraints of latest completion, checkpoint recovery, or side-effect window are eliminated. The execution window, execution node, and segmentable boundary are selected based on the resource conflict cost, checkpoint and data migration cost, and the freshness deficit at completion. An execution lease containing the controller period, task period, and stage idempotent key is issued for the selected scheme. The execution agent executes or submits the corresponding stage after the lease and both periods are verified. Based on the actual progress and resource occupancy update completion time prediction, a new task period is generated when the rescheduling conditions are met. Only stages without submitted side effects are rescheduled, and completed stages whose output digests match the input digests of the new period are reused.
2. The method as described in claim 1, characterized in that, The task scheduling contract includes at least the task stability identifier, nominal trigger time, target update interval, result validity period, latest completion time, phase diagram, resource envelope of each phase, read set, write set, interruptible point, checkpoint location, compensation action, and side effect commit point; After the task is completed, the age of the most recent valid result is updated based on the actual completion time, and the freshness deficit is not cleared to zero based on the nominal trigger time.
3. The method as described in claim 1, characterized in that, Let the age of the most recent valid result be A, the target update interval be T, and the number of consecutive delays be n. The current freshness deficit Dcurrent = max(0, AT) + λnT. For a candidate solution x, the freshness deficit when the candidate is completed is Dfinish(x) = max(0, A + tfinish(x) - T) + λnT. The objective function is J(x) = αCconflict(x) + βCmigrate(x) + γDfinish(x). Among the candidate solutions that satisfy the hard constraints, the one with the smallest J(x) is selected.
4. The method as described in claim 1, characterized in that, For each resource dimension, the lower bound of resource capacity is Cavailable_low = Cphysical_low - Rcommitted - Emeasure - Ssafety, where Cphysical_low is the lower bound of measurable physical capacity, Rcommitted is the committed reservation, Emeasure is the upper bound of measurement error, and Ssafety is the safety margin. When the predicted interval width exceeds the prediction deviation boundary during the online operation phase, a short-term low-conflict interval of resources is formed based on the real-time lower bound of resource capacity, and the lease term is shortened.
5. The method as described in claim 1, characterized in that, The controller period is generated by atomic increment operations stored in the persistent period; the controller generates an integrity authentication code for the controller identifier, controller period, task period, input digest, phase checkpoint digest, and expiration time; After the authentication code is verified, the execution agent persists the maximum controller period and the maximum task period. Execution leases in low controller periods cannot submit results even if they have not yet expired.
6. The method as described in claim 1, characterized in that, When a lease expires or a network partition occurs, the execution agent stops after setting checkpoints for phases without side effects and which are interruptible. For phases that have already reached the side effect submission point, it only completes confirmation or performs predefined compensation without reallocation. After connectivity is restored, valid results are selected according to the controller period and task period, and results from lower periods are only used as audit records.
7. A dynamic scheduling system for timed tasks, characterized in that, The system includes a contract generation module, a freshness calculation module, a resource interval prediction module, a joint selection module, a period lease execution module, and a rolling rescheduling module. The contract generation module generates task scheduling contracts with phased resource envelopes and side-effect commit points. The freshness calculation module calculates freshness deficit and movable execution windows. The resource interval prediction module generates low-conflict resource intervals based on the upper and lower bounds of online computing power job occupancy and the lower bound of resource capacity. The joint selection module performs filtering based on latest completion, resource capacity, and checkpoint recovery constraints, and selects an execution scheme based on resource peak overlap, migration cost, and freshness deficit at candidate completion times. The period lease execution module issues and verifies execution leases that include both controller and task periods. The rolling rescheduling module only reorders uncommitted stages and reuses completed stages with consistent digests when predictions are completed or resource occupancy reaches rescheduling conditions.
8. An electronic device, characterized in that, The device includes a processor, a memory, a communication interface, and a computer program stored in the memory and capable of running on the processor. The communication interface is used to connect a scheduling controller, a task execution agent, and a resource monitor. When the computer program is executed by the processor, it implements the method according to any one of claims 1 to 6.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 6.
