Task scheduling method and system based on heterogeneous video storage system

Abstract

The invention discloses a task scheduling method and system based on a heterogeneous video storage system. The task scheduling method comprises the following steps: acquiring data in the heterogeneous video storage system and calculating a system congestion degree baseline; generating a video task actual scheduling scheme and a virtual migration-free baseline scheduling scheme, and calculating completion time; calculating a normalized net income index and a migration oscillation index according to the completion time difference in combination with a preset task importance weight, a node congestion degree change and a round-trip migration frequency; generating node-level migration threshold configuration based on the self-adaptive migration threshold updating rule; calculating video task migration net earnings for the video tasks on the high-congestion nodes; and executing video task rescheduling and forming closed-loop control based on an execution result. According to the adaptive scheduling method based on congestion perception and income oscillation constraint, the dynamic migration decision of the cross-node video task is realized, and the adaptive scheduling method has the advantages of accurate scheduling, efficient resources and stable system.

Description

Technical Field

[0001] This invention relates to the field of video task scheduling, and in particular to a task scheduling method and system based on a heterogeneous video storage system. Background Technology

[0002] Heterogeneous video storage systems are widely used in urban surveillance, park security, and intelligent analytics scenarios. Different types of storage nodes form a multi-level collaborative architecture due to differences in processing power, bandwidth, and storage media. Existing technologies typically schedule video tasks through static load balancing, fixed priorities, or simple queue rules, triggering task migration only when node utilization is high to alleviate local congestion. However, with massive continuous video access, intensive alarm analysis, and frequent access to keyframes, storage node load exhibits rapid fluctuations. Traditional scheduling methods struggle to accurately reflect the overall system congestion status, and the errors in task queuing time and execution time prediction accumulate, leading to insufficient stability of the scheduling scheme. Furthermore, the impact of cross-node scheduling behavior on global performance is difficult to quantify and assess.

[0003] Existing technologies generally suffer from drawbacks such as crude migration decision-making, unmeasurable benefits, and a lack of business sensitivity. Most migration strategies rely solely on a single utilization threshold or queue length threshold to determine migration timing, failing to differentiate between video task types, keyframe sensitivity, and service level requirements, and thus unable to handle priority conflicts in mixed business scenarios. Furthermore, existing methods lack quantifiable metrics for queuing latency fluctuations, node congestion changes, and round-trip migration phenomena caused by task migration, leading to volatile migration behavior and even overall performance degradation. In addition, traditional methods lack a dynamic update mechanism based on migration execution results, preventing the system from adjusting subsequent scheduling strategies based on actual operational deviations, resulting in a lack of closed-loop control capabilities in the scheduling process.

[0004] Therefore, how to provide a task scheduling method and system based on a heterogeneous video storage system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a task scheduling method and system based on a heterogeneous video storage system. This invention constructs an adaptive scheduling method with congestion awareness and revenue oscillation constraints to realize dynamic migration decision-making for cross-node video tasks, which has the advantages of accurate scheduling, high resource efficiency and system stability.

[0006] A task scheduling method based on a heterogeneous video storage system according to an embodiment of the present invention includes the following steps: Obtain the utilization rate of each storage node, the waiting time of video tasks, and the queue length in the heterogeneous video storage system; calculate the congestion of storage nodes and summarize them to form a system congestion baseline. Within each scheduling window, an actual scheduling scheme for video tasks is generated based on the system congestion baseline, and a virtual no-migration baseline scheduling scheme that prohibits task migration is constructed to obtain the set of completion times for the same batch of video tasks under the two scheduling schemes. Based on the difference in completion time between the virtual no-migration baseline scheduling scheme and the actual video task scheduling scheme, and combined with the preset task importance weight, node congestion changes and round-trip migration times, the normalized net benefit index and migration oscillation index are calculated. The system congestion baseline, normalized net revenue index and migration oscillation index are input into the adaptive migration threshold update rule to obtain the adaptive migration threshold of the scheduling window and generate the node-level migration threshold configuration for each storage node. For video tasks located on highly congested nodes, estimate the completion time and migration cost under the two scenarios of keeping the migration unchanged and performing the migration, and calculate the net benefit of video task migration; When the net benefit of video task migration exceeds the corresponding node-level migration threshold and meets the migration constraints, video task rescheduling is performed, and the normalized net benefit index and migration oscillation index of subsequent scheduling windows are updated based on the execution results to form a closed-loop control.

[0007] Optionally, the generation of the system congestion baseline specifically includes: In a heterogeneous video storage system, the processor utilization rate, input / output bandwidth utilization rate, and disk read / write utilization rate of each storage node are periodically collected to obtain the original sampling sequence of each storage node utilization rate. Denoising processing, outlier removal processing, and time alignment processing are then performed to generate a cleaned sequence of each storage node utilization rate. The video tasks queued on each storage node are statistically analyzed. The time difference between the time when each video task enters the queue and the current time is calculated to obtain the original sampling sequence of the waiting time of video tasks on each storage node. Interpolation and truncation processing are performed to generate the cleaned sequence of the waiting time of video tasks on each storage node. The number of unfinished video tasks in each storage node queue is counted to obtain the original sampled sequence of the queue length of each storage node. The sequence is then normalized according to the preset maximum queue capacity to generate the normalized sequence of the queue length of each storage node. The utilization cleaning sequence of each storage node and the waiting time cleaning sequence of video tasks are subjected to interval normalization processing, and then concatenated with the queue length normalization sequence on a unified time scale to form a local congestion feature vector sequence. Based on the local congestion feature vector sequence, the normalized values ​​of storage node utilization, video task waiting time, and queue length are weighted and summed according to the preset weighting rules to obtain the storage node congestion sequence of each storage node. The sequence is then averaged over time within the current scheduling window to generate the storage node congestion degree of the scheduling window. Within the current scheduling window, the congestion of each storage node in the scheduling window is fused at the node level. Based on the storage capacity, video task type distribution, and service level weight of each storage node, a system-level weighting coefficient is generated. The system-level weighting coefficient is then summed to obtain the system congestion baseline that represents the overall load status of the current scheduling window.

[0008] Optionally, the generation of the completion time set specifically includes: At the beginning of each scheduling window, based on the current system congestion baseline, a set of video tasks in the pending scheduling state is obtained, and their video task type label, video access mode label, key frame priority label, cold and hot storage level label and service level label are associated to form a video task description set for the scheduling window. For each video task in the video task description set in the scheduling window, the comprehensive priority of the video task is calculated. The comprehensive priority sequence of video tasks in the scheduling window is obtained by linearly weighting the service level weight, video task type weight, video access mode and key frame priority, and cold and hot storage level weight according to the preset adjustment coefficient. Within the scheduling window, for each video task, candidate storage nodes in the heterogeneous video storage system are traversed. Based on the congestion of storage nodes in the scheduling window, the system congestion baseline, and the estimated queuing waiting time and processing time of video tasks on candidate storage nodes, the scheduling matching score of candidate storage nodes for video tasks is calculated to form a scheduling matching score matrix. Based on the scheduling matching score matrix, target storage nodes are selected for video tasks according to the rule of scheduling matching scores from high to low, so as to obtain the actual scheduling scheme of video tasks facing the scheduling window and calculate the completion time of video tasks. While maintaining the initial binding relationship between each video task and each storage node at the start of the scheduling window, a virtual no-migration baseline scheduling scheme that prohibits task migration is constructed. The storage nodes are sorted according to the arrival order of the video tasks. The expected start time and expected execution duration are estimated based on the execution capabilities of the corresponding storage nodes, and the completion time of the video tasks is calculated. The scheduling scheme and video task completion time are matched to generate a set of video task completion times for the same batch of video tasks in the actual video task scheduling scheme and in the virtual no-migration baseline scheduling scheme.

[0009] Optionally, the generation of the normalized net return index and the migration oscillation index specifically includes: Within each scheduling window, based on the difference between the video task completion time of the actual video task scheduling scheme and the virtual no-migration baseline scheduling scheme, the video task completion time difference value is obtained. The ratio of the difference value to the video task completion time in the virtual no-migration baseline scheduling scheme is used as the video task completion time saving ratio, thus obtaining the video task completion time difference value sequence and the video task completion time saving ratio sequence within the scheduling window. The importance of video task descriptions in the scheduling window is characterized according to preset rules. Video task type, video access mode, key frame priority, cold and hot storage level and service level are mapped to service level score, video task type score, video access mode and key frame priority score and cold and hot storage level score, respectively. The task importance weight calculation function is input and linearly weighted according to preset adjustment coefficient to form the preset task importance weight sequence of the scheduling window. Within the scheduling window, the normalized net benefit of each video task is calculated by multiplying the preset task importance weight corresponding to each video task by the video task completion time saving ratio, and then aggregating them in a weighted average manner to obtain the normalized net benefit index. Within the same scheduling window, video tasks that have undergone migration operations in heterogeneous video storage systems are filtered, and congestion fluctuation components, round-trip migration count components, and keyframe access delay fluctuation components are calculated. The migration oscillation index is then input into the migration oscillation index calculation function and weighted and summed according to a preset weighting rule to obtain the migration oscillation index.

[0010] Optionally, the generation of the node-level migration threshold configuration specifically includes: Within each scheduling window, the system congestion baseline, normalized net revenue index, and migration oscillation index are input into the adaptive migration threshold update rule. The system congestion baseline is compared with the preset congestion reference interval, and the basic migration threshold of the scheduling window is generated based on the overall load level of the current scheduling window. Based on the basic migration threshold of the scheduling window, the scheduling revenue adjustment is calculated using the normalized net revenue index, and the amplitude is truncated to generate the scheduling revenue adjustment result that represents the degree of influence of the overall scheduling revenue of the current scheduling window on the migration threshold. The oscillation suppression adjustment amount is calculated using the migration oscillation index and the amplitude is truncated to generate the oscillation suppression adjustment result that characterizes the degree of impact of migration on the stability of video services within the current scheduling window; The scheduling window's basic migration threshold, scheduling revenue adjustment result, and oscillation suppression adjustment result are superimposed according to a preset combination rule to obtain the instantaneous migration threshold of the scheduling window. This instantaneous migration threshold is then smoothed by a sliding weighted average with the historical migration thresholds of the previous scheduling windows. Preset upper and lower bound constraints on the migration threshold are applied to generate the adaptive migration threshold of the scheduling window. Based on the adaptive migration threshold of the scheduling window, and taking into account the storage medium type, storage capacity, service level weight of each storage node, as well as the migration success rate, migration failure rate and number of round trips in the historical scheduling window, a node-level offset factor is applied to generate a node-level migration threshold for each storage node. The node-level migration threshold is then organized into a node-level migration threshold configuration and distributed to each storage node in the heterogeneous video storage system.

[0011] Optionally, the generation of the net benefit of video task migration specifically includes: From the heterogeneous video storage system, storage nodes with a congestion level higher than a preset high congestion threshold in the scheduling window are selected as a set of high-congestion storage nodes. Video tasks that are currently in a queue or have not yet been completed are extracted to generate a video task migration candidate set. Iterate through all storage nodes except for the high-congestion storage nodes where the video tasks are located in the video task migration candidate set. Add storage nodes whose congestion level in the scheduling window is lower than the preset migration congestion threshold and whose corresponding node-level migration threshold in the node-level migration threshold configuration is not higher than the scheduling window adaptive migration threshold to the migration target storage node candidate set, and record the mapping relationship to form a migration target storage node candidate mapping table. For the candidate mapping table of migration target storage nodes, the estimated remaining queuing wait time of the video task on its current high-congestion storage node is added to the estimated execution time to obtain the completion time of the video task without migration. For the candidate mapping table of the target storage node, the estimated data migration time, queuing wait time and estimated execution time required to migrate the video task from the high-congestion storage node to the target storage node are added together to obtain the completion time of the video task under the migration scenario. Subtract the completion time under the condition of not migrating from the completion time under the condition of performing migration to obtain the video task migration time benefit for the corresponding video task; Combining the network bandwidth distribution of the heterogeneous video storage system, the input / output bandwidth utilization rate and disk read / write utilization rate of each storage node, the video task migration cost is calculated for each pair of highly congested storage nodes and the migration target storage node. The estimated data migration time, the additional input / output load increment introduced on the source and migration target storage nodes during the migration process, and the disturbance penalty factor caused by the key frame access delay are weighted by corresponding weighting coefficients and then summed to obtain the video task migration cost. For each pair of high-congestion storage nodes and migration target storage nodes, the net benefit of video task migration is calculated by subtracting the video task migration cost from the video task migration time benefit of the corresponding video task. Among all the migration target storage node candidates for the same video task, the migration target storage node with the largest net benefit of video task migration is selected, and the corresponding net benefit of video task migration is recorded as the video task migration decision result. The results are then summarized to form the video task migration net benefit result set.

[0012] Optionally, the generation of the closed-loop control specifically includes: Within each scheduling window, obtain the net benefit set of video task migration and the node-level migration threshold configuration, set migration constraints, compare the net benefit of video task migration with the corresponding node-level migration threshold and migration constraints, and obtain a set of candidate migration actions. The candidate migration action set is sorted from high to low according to the net benefit of video task migration. The number of concurrent migrations and the expected network bandwidth usage of each storage node are dynamically counted. If a new candidate migration action still meets the migration constraints after being added, it is added to the video task rescheduling plan. The video task rescheduling plan is sent to the corresponding source storage node and the migration target storage node. The task removal operation in the task queue of the source storage node, the task insertion operation in the task queue of the migration target storage node, and the video data migration trigger operation are executed in sequence, and the video task rescheduling execution record set is collected. Based on the set of video task rescheduling execution records, the actual video task completion time is calculated, the expected start time and expected execution duration of the corresponding video task in the actual video task scheduling scheme are updated, the updated actual video task scheduling scheme is obtained, and the video task completion time difference, key frame access delay fluctuation and storage node congestion change before and after migration are re-statistically analyzed within the scheduling window to obtain the execution deviation statistics. Based on the execution deviation statistics, the historical statistics of the normalized net return index and migration oscillation index are updated across scheduling windows, and the adjustment coefficients are updated in the task importance weight calculation function and migration oscillation index calculation function to complete the closed-loop control link.

[0013] A task scheduling system based on a heterogeneous video storage system according to an embodiment of the present invention includes: The congestion baseline module is used to obtain the utilization rate of each storage node, the waiting time of video tasks, and the queue length, calculate the congestion degree of storage nodes, and form the system congestion degree baseline. The scheduling generation module is used to generate actual scheduling schemes for video tasks based on the system congestion baseline within each scheduling window, and to construct a virtual non-migration baseline scheduling scheme that prohibits task migration, thereby obtaining the set of completion times for video tasks under the two scheduling schemes. The revenue oscillation module is used to calculate the normalized net revenue index and migration oscillation index based on the difference in completion time between the two scheduling schemes, combined with the task importance weight, node congestion changes and round-trip migration times. The threshold update module is used to input the system congestion baseline, normalized net revenue index and migration oscillation index into the adaptive migration threshold update rule to generate the scheduling window adaptive migration threshold and node-level migration threshold configuration. The migration benefit module is used to estimate the completion time and migration cost for video tasks located on highly congested nodes under two scenarios: keeping the migration in place and performing the migration, and to calculate the net migration benefit of the video task. The rescheduling module is used to perform video task rescheduling and update the normalized net return index and migration oscillation index based on the rescheduling execution results, forming a closed-loop control.

[0014] The beneficial effects of this invention are: This invention constructs a dynamic scheduling mechanism oriented towards congestion evolution characteristics within a heterogeneous video storage system. This achieves a unified quantitative description of system-level load status, task-level revenue changes, and migration-side oscillation risks, enabling scheduling behavior to maintain stability and foresight under the joint constraints of multiple dimensions. Based on the collaborative calculation of the system congestion baseline, normalized net revenue index, and migration oscillation index, this invention can accurately identify the true value of task migration at different time periods, transforming migration behavior within the scheduling window from the previous coarse-grained triggering to a refined, calculable, and verifiable decision-making process. Simultaneously, by introducing adaptive migration thresholds and node-level migration threshold configurations, the system can execute differentiated scheduling strategies based on the capacity differences, service density, and historical migration performance of different storage nodes, effectively suppressing queuing expansion on high-load nodes and improving overall resource utilization efficiency.

[0015] With the support of a closed-loop control link, this invention further achieves continuous self-correction capabilities across scheduling windows. The actual execution results of video task rescheduling are quantified and fed back to update the normalized net benefit index and migration oscillation index, enabling the system to continuously correct latency estimation errors and congestion prediction biases. This avoids scheduling degradation caused by the cumulative distortion of migration decisions over time in traditional methods. Through the combined effect of dynamic benefit evaluation and oscillation suppression strategies, the completion time of video tasks in high-congestion scenarios is significantly improved, keyframe access latency fluctuations are significantly reduced, and the effectiveness and stability of migration are continuously enhanced. Therefore, this invention achieves higher scheduling accuracy, better global performance, and stronger autonomous adaptability in complex, multi-service concurrent heterogeneous video storage environments. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0017] Figure 1 This is a flowchart of a task scheduling method based on a heterogeneous video storage system proposed in this invention; Figure 2 This is a schematic diagram of the video task migration net benefit calculation process of a task scheduling method based on a heterogeneous video storage system proposed in this invention; Figure 3 This is a schematic diagram of the adaptive migration threshold update process of a task scheduling method based on a heterogeneous video storage system proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A task scheduling method based on a heterogeneous video storage system includes the following steps: Obtain the utilization rate of each storage node, the waiting time of video tasks, and the queue length in the heterogeneous video storage system; calculate the congestion of storage nodes and summarize them to form a system congestion baseline. Within each scheduling window, an actual scheduling scheme for video tasks is generated based on the system congestion baseline, and a virtual no-migration baseline scheduling scheme that prohibits task migration is constructed to obtain the set of completion times for the same batch of video tasks under the two scheduling schemes. Based on the difference in completion time between the virtual no-migration baseline scheduling scheme and the actual video task scheduling scheme, and combined with the preset task importance weight, node congestion changes and round-trip migration times, the normalized net benefit index and migration oscillation index are calculated. The system congestion baseline, normalized net revenue index and migration oscillation index are input into the adaptive migration threshold update rule to obtain the adaptive migration threshold of the scheduling window and generate the node-level migration threshold configuration for each storage node. For video tasks located on highly congested nodes, estimate the completion time and migration cost under the two scenarios of keeping the migration unchanged and performing the migration, and calculate the net benefit of video task migration; When the net benefit of video task migration exceeds the corresponding node-level migration threshold and meets the migration constraints, video task rescheduling is performed, and the normalized net benefit index and migration oscillation index of subsequent scheduling windows are updated based on the execution results to form a closed-loop control.

[0020] In this embodiment, the generation of the system congestion baseline specifically includes: In a heterogeneous video storage system, the processor utilization rate, input / output bandwidth utilization rate, and disk read / write utilization rate of each storage node are periodically collected to obtain the original sampling sequence of each storage node utilization rate. Denoising processing, outlier removal processing, and time alignment processing are then performed to generate a cleaned sequence of each storage node utilization rate. The video tasks queued on each storage node are statistically analyzed. The time difference between the time when each video task enters the queue and the current time is calculated to obtain the original sampling sequence of the waiting time of video tasks on each storage node. Interpolation and truncation processing are performed to generate the cleaned sequence of the waiting time of video tasks on each storage node. The number of unfinished video tasks in each storage node queue is counted to obtain the original sampled sequence of the queue length of each storage node. The sequence is then normalized according to the preset maximum queue capacity to generate the normalized sequence of the queue length of each storage node. The utilization cleaning sequence of each storage node and the waiting time cleaning sequence of video tasks are subjected to interval normalization processing, and then concatenated with the queue length normalization sequence on a unified time scale to form a local congestion feature vector sequence. Based on the local congestion feature vector sequence, the normalized values ​​of storage node utilization, video task waiting time, and queue length are weighted and summed according to the preset weighting rules to obtain the storage node congestion sequence of each storage node. The sequence is then averaged over time within the current scheduling window to generate the storage node congestion degree of the scheduling window. Within the current scheduling window, the congestion of each storage node in the scheduling window is fused at the node level. Based on the storage capacity, video task type distribution, and service level weight of each storage node, a system-level weighting coefficient is generated. The system-level weighting coefficient is then summed to obtain the system congestion baseline that represents the overall load status of the current scheduling window.

[0021] In this embodiment, the generation of the completion time set specifically includes: At the beginning of each scheduling window, based on the current system congestion baseline, a set of video tasks in the pending scheduling state is obtained, and their video task type label, video access mode label, key frame priority label, cold and hot storage level label and service level label are associated to form a video task description set for the scheduling window. For each video task in the video task description set in the scheduling window, the comprehensive priority of the video task is calculated. The comprehensive priority sequence of video tasks in the scheduling window is obtained by linearly weighting the service level weight, video task type weight, video access mode and key frame priority, and cold and hot storage level weight according to the preset adjustment coefficient. Within the scheduling window, for each video task, candidate storage nodes in the heterogeneous video storage system are traversed. Based on the congestion of storage nodes in the scheduling window, the system congestion baseline, and the estimated queuing waiting time and processing time of video tasks on candidate storage nodes, the scheduling matching score of candidate storage nodes for video tasks is calculated to form a scheduling matching score matrix. The scheduling matching score is obtained by subtracting the weighted value of storage node congestion and the weighted value of queuing time from the comprehensive priority of the video task. The weighted value of storage node congestion and the weighted value of queuing time are controlled by the congestion influence coefficient and the queuing time influence coefficient, respectively. Based on the scheduling matching score matrix, target storage nodes are selected for video tasks according to the rule of scheduling matching scores from high to low, so as to obtain the actual scheduling scheme of video tasks facing the scheduling window and calculate the completion time of video tasks. The actual video task scheduling scheme sorts video tasks on the same target storage node from high to low according to the comprehensive priority of the video tasks. Combined with the estimated execution time of the video tasks already existing in the current queue of the target storage node, it determines the expected start time and expected execution duration of each video task on the target storage node. While maintaining the initial binding relationship between each video task and each storage node at the start of the scheduling window, a virtual no-migration baseline scheduling scheme that prohibits task migration is constructed. The storage nodes are sorted according to the arrival order of the video tasks. The expected start time and expected execution duration are estimated based on the execution capabilities of the corresponding storage nodes, and the completion time of the video tasks is calculated. The scheduling scheme and video task completion time are matched to generate a set of video task completion times for the same batch of video tasks in the actual video task scheduling scheme and in the virtual no-migration baseline scheduling scheme.

[0022] In this embodiment, the generation of the normalized net return index and the migration oscillation index specifically includes: Within each scheduling window, based on the difference between the video task completion time of the actual video task scheduling scheme and the virtual no-migration baseline scheduling scheme, the video task completion time difference value is obtained. The ratio of the difference value to the video task completion time in the virtual no-migration baseline scheduling scheme is used as the video task completion time saving ratio, thus obtaining the video task completion time difference value sequence and the video task completion time saving ratio sequence within the scheduling window. The importance of video task descriptions in the scheduling window is characterized according to preset rules. Video task type, video access mode, key frame priority, cold and hot storage level and service level are mapped to service level score, video task type score, video access mode and key frame priority score and cold and hot storage level score, respectively. The task importance weight calculation function is input and linearly weighted according to preset adjustment coefficient to form the preset task importance weight sequence of the scheduling window. The importance characterization is based on the video task type label, video access mode label, keyframe priority label, cold and hot storage level label, and service level label pre-associated with each video task. The characterization is based on the following order: video alarm analysis tasks take priority over ordinary video analysis tasks, video analysis tasks take priority over video playback tasks, video access modes with dense keyframes take priority over ordinary sequential access modes, video tasks located in high-speed cold and hot storage levels take priority over video tasks located in low-speed cold and hot storage levels, and video tasks with higher service level labels take priority over video tasks with lower service level labels. Within the scheduling window, the normalized net benefit of each video task is calculated by multiplying the preset task importance weight corresponding to each video task by the video task completion time saving ratio, and then aggregated in a weighted average manner to obtain the normalized net benefit index that represents the overall scheduling benefit level of the current scheduling window in the heterogeneous video storage system. Within the same scheduling window, video tasks that have undergone migration operations in heterogeneous video storage systems are screened, and congestion fluctuation components, round-trip migration count components, and keyframe access delay fluctuation components are calculated. The migration oscillation index is input into the migration oscillation index calculation function and weighted and summed according to the preset weighting rules to obtain the migration oscillation index used to characterize the degree of impact of migration on the stability of video services within the current scheduling window. The congestion fluctuation component is obtained by calculating the difference in congestion of each storage node before and after migration and the absolute value of the change in congestion of each storage node, accumulating them, and averaging them over the number of migrations. The round-trip migration count component is obtained by counting the number of round-trip migrations of the same video task between multiple storage nodes in both the actual video task scheduling scheme and the virtual no-migration baseline scheduling scheme. The keyframe access latency fluctuation component is obtained by calculating the fluctuation amplitude of access latency for keyframes within the scheduling window for video alarm analysis tasks and keyframe intensive access requests.

[0023] In this embodiment, the generation of the node-level migration threshold configuration specifically includes: Within each scheduling window, the system congestion baseline, normalized net revenue index, and migration oscillation index are input into the adaptive migration threshold update rule. The system congestion baseline is compared with the preset congestion reference interval, and the basic migration threshold of the scheduling window is generated based on the overall load level of the current scheduling window. Based on the basic migration threshold of the scheduling window, the scheduling revenue adjustment is calculated using the normalized net revenue index, and the amplitude is truncated to generate the scheduling revenue adjustment result that represents the degree of influence of the overall scheduling revenue of the current scheduling window on the migration threshold. Specifically, when the normalized net return index is higher than the preset return target, the scheduling return adjustment amount is set to a negative offset that lowers the migration threshold; when the normalized net return index is lower than the preset return target, the scheduling return adjustment amount is set to a positive offset that raises the migration threshold. The oscillation suppression adjustment amount is calculated using the migration oscillation index and the amplitude is truncated to generate the oscillation suppression adjustment result that characterizes the degree of impact of migration on the stability of video services within the current scheduling window; Specifically, when the migration oscillation index is higher than the preset oscillation tolerance threshold, the oscillation suppression adjustment amount is set to a positive offset that increases the migration threshold; when the migration oscillation index is lower than the preset oscillation tolerance threshold, the oscillation suppression adjustment amount is set to a negative offset that decreases the migration threshold. The scheduling window's basic migration threshold, scheduling revenue adjustment result, and oscillation suppression adjustment result are superimposed according to a preset combination rule to obtain the instantaneous migration threshold of the scheduling window. This instantaneous migration threshold is then smoothed by a sliding weighted average with the historical migration thresholds of the previous scheduling windows. Preset upper and lower bound constraints on the migration threshold are applied to generate the adaptive migration threshold of the scheduling window. Based on the adaptive migration threshold of the scheduling window, and taking into account the storage medium type, storage capacity, service level weight of each storage node, as well as the migration success rate, migration failure rate and number of round trips in the historical scheduling window, a node-level offset factor is applied to generate a node-level migration threshold for each storage node. The node-level migration threshold is then organized into a node-level migration threshold configuration and distributed to each storage node in the heterogeneous video storage system.

[0024] In this embodiment, the generation of the net benefit of video task migration specifically includes: From the heterogeneous video storage system, storage nodes with a congestion level higher than a preset high congestion threshold in the scheduling window are selected as a set of high-congestion storage nodes. Video tasks that are currently in a queue or have not yet been completed are extracted to generate a video task migration candidate set. Iterate through all storage nodes except for the high-congestion storage nodes where the video tasks are located in the video task migration candidate set. Add storage nodes whose congestion level in the scheduling window is lower than the preset migration congestion threshold and whose corresponding node-level migration threshold in the node-level migration threshold configuration is not higher than the scheduling window adaptive migration threshold to the migration target storage node candidate set, and record the mapping relationship to form a migration target storage node candidate mapping table. For the candidate mapping table of migration target storage nodes, the estimated remaining queuing wait time of the video task on its current high-congestion storage node is added to the estimated execution time to obtain the completion time of the video task without migration. For the candidate mapping table of the target storage node, the estimated data migration time, queuing wait time and estimated execution time required to migrate the video task from the high-congestion storage node to the target storage node are added together to obtain the completion time of the video task under the migration scenario. Subtract the completion time under the condition of not migrating from the completion time under the condition of performing migration to obtain the video task migration time benefit for the corresponding video task; Combining the network bandwidth distribution of the heterogeneous video storage system, the input / output bandwidth utilization rate and disk read / write utilization rate of each storage node, the video task migration cost is calculated for each pair of highly congested storage nodes and the migration target storage node. The estimated data migration time, the additional input / output load increment introduced on the source and migration target storage nodes during the migration process, and the disturbance penalty factor caused by the key frame access delay are weighted by corresponding weighting coefficients and then summed to obtain the video task migration cost. For each pair of high-congestion storage nodes and migration target storage nodes, the net benefit of video task migration is calculated by subtracting the video task migration cost from the video task migration time benefit of the corresponding video task. Among all the migration target storage node candidates for the same video task, the migration target storage node with the largest net benefit of video task migration is selected, and the corresponding net benefit of video task migration is recorded as the video task migration decision result. The results are then summarized to form the video task migration net benefit result set.

[0025] In this embodiment, the generation of the closed-loop control specifically includes: Within each scheduling window, obtain the net benefit set of video task migration and the node-level migration threshold configuration, set migration constraints, compare the net benefit of video task migration with the corresponding node-level migration threshold and migration constraints, and obtain a set of candidate migration actions. The migration constraints include the maximum number of concurrent migrations, the maximum number of migrations, the minimum available network bandwidth, and the maximum migration latency corresponding to the service level. The candidate migration action set is sorted from high to low according to the net benefit of video task migration. The number of concurrent migrations and the expected network bandwidth usage of each storage node are dynamically counted. If a new candidate migration action still meets the migration constraints after being added, it is added to the video task rescheduling plan. The video task rescheduling plan is sent to the corresponding source storage node and the migration target storage node. The task removal operation in the task queue of the source storage node, the task insertion operation in the task queue of the migration target storage node, and the video data migration trigger operation are executed in sequence, and the video task rescheduling execution record set is collected. The video task rescheduling execution record set includes the actual start time of migration, the actual end time of migration, the actual amount of data migrated, changes in storage node congestion, and changes in keyframe access latency. Based on the set of video task rescheduling execution records, the actual video task completion time is calculated, the expected start time and expected execution duration of the corresponding video task in the actual video task scheduling scheme are updated, the updated actual video task scheduling scheme is obtained, and the video task completion time difference, key frame access delay fluctuation and storage node congestion change before and after migration are re-statistically analyzed within the scheduling window to obtain the execution deviation statistics. Based on the execution deviation statistics, the historical statistics of the normalized net return index and migration oscillation index are updated across scheduling windows, and the adjustment coefficients are updated in the task importance weight calculation function and migration oscillation index calculation function to complete the closed-loop control link.

[0026] A task scheduling system based on a heterogeneous video storage system includes: The congestion baseline module is used to obtain the utilization rate of each storage node, the waiting time of video tasks, and the queue length, calculate the congestion degree of storage nodes, and form the system congestion degree baseline. The scheduling generation module is used to generate actual scheduling schemes for video tasks based on the system congestion baseline within each scheduling window, and to construct a virtual non-migration baseline scheduling scheme that prohibits task migration, thereby obtaining the set of completion times for video tasks under the two scheduling schemes. The revenue oscillation module is used to calculate the normalized net revenue index and migration oscillation index based on the difference in completion time between the two scheduling schemes, combined with the task importance weight, node congestion changes and round-trip migration times. The threshold update module is used to input the system congestion baseline, normalized net revenue index and migration oscillation index into the adaptive migration threshold update rule to generate the scheduling window adaptive migration threshold and node-level migration threshold configuration. The migration benefit module is used to estimate the completion time and migration cost for video tasks located on highly congested nodes under two scenarios: keeping the migration in place and performing the migration, and to calculate the net migration benefit of the video task. The rescheduling module is used to perform video task rescheduling and update the normalized net return index and migration oscillation index based on the rescheduling execution results, forming a closed-loop control.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a video surveillance center in a coastal city. This center employs a heterogeneous video storage system consisting of high-speed solid-state nodes, ordinary mechanical nodes, and historical archive nodes. It has long faced problems such as low task scheduling efficiency, persistently high congestion on some nodes, and significant latency fluctuations in video alarm tasks, leading to accumulated access delays and delayed alarm processing during peak daytime hours. After the method of this invention was fully deployed in the center, its verification was conducted on various video tasks, including video analysis, alarm identification, historical review, and structured extraction.

[0028] In actual operation, the system first continuously collects the processor utilization, bandwidth utilization, disk read / write load, queue length, and video task waiting time of each storage node to form a unified congestion status input. The original scheduling mechanism of this center only made coarse-grained allocations based on the remaining space of a single node and the task type, which made it difficult to accurately identify the node congestion evolution trend. After the deployment of this invention, a system congestion baseline is generated within the scheduling window, which can clearly reflect the decline in node processing capacity during high-load phases and serve as an important basis for subsequent scheduling. For each batch of video tasks to be scheduled, the system generates both the actual scheduling scheme and a virtual non-migration baseline path that prohibits migration, so that the completion time of the task under the two schemes forms a comparable time set, providing a basis for subsequent revenue calculation.

[0029] During the continuous operation of the scheduling window, this invention constructs a normalized net benefit index and a migration oscillation index based on differences in task completion time, service level, and the number of round trips, enabling the system to dynamically evaluate scheduling benefits and business fluctuations in multiple rounds of scheduling. The monitoring center experiences intensive structured tasks and sudden alarm events at night; the benefit oscillation joint quantification mechanism of this invention can effectively identify access fluctuations caused by migration behavior, thereby forming more robust node-level scheduling characteristics.

[0030] When the system enters a high-congestion state, such as when some nodes experience continuous access backlog, this invention forms an adaptive migration threshold for the scheduling window through adaptive migration threshold update rules. Combined with node-level migration threshold configuration, this ensures that task migration actions follow congestion, benefit, and oscillation constraints in parallel. During verification, for tasks located within highly congested nodes, this invention estimates the completion time for both "staying on without migration" and "performing migration," while simultaneously quantifying the additional load introduced by data migration, thus achieving an accurate estimate of the net benefit of video task migration.

[0031] Subsequently, the system reschedules tasks whose net migration benefit exceeds the corresponding node-level migration threshold. Information such as migration time, node congestion changes, and keyframe access latency changes during the rescheduling process is fed back to the update link of the benefit index and oscillation index, enabling cross-window self-correction of the scheduling strategy. The scheduling operation records of the monitoring center show that the closed-loop control chain of this invention can maintain a stable scheduling rhythm through multiple load fluctuations, achieving a better balance between resource utilization and task latency.

[0032] The overall operational results show that the application of this invention in the monitoring center can solve the problems that the original scheduling mechanism cannot accurately identify congestion, the migration strategy lacks benefit constraints, and the business fluctuations cannot be quantified. This enables the video storage system to maintain a more stable and efficient task scheduling capability in long-term operation, and significantly improves the overall business smoothness.

[0033] Table 1. Performance comparison between a task scheduling method based on a heterogeneous video storage system and traditional methods.

[0034] As shown in Table 1, in terms of scheduling efficiency, this invention reduces the average task completion time from 12.86 seconds to 8.21 seconds, a decrease of 36.18%. This result demonstrates that, through the combined effects of system congestion baseline, comprehensive video task priority, dual-path time estimation, and scheduling scoring matrix, this invention significantly improves the allocation accuracy of tasks across different storage nodes, making the scheduling process more closely reflect the actual load capacity of the nodes. Simultaneously, the peak-period task queuing time is reduced from 19.44 seconds in the traditional mechanism to 11.03 seconds, a reduction of 43.28%, proving that the adaptive migration threshold mechanism of this invention can guide task migration in a timely manner when the load increases rapidly, thereby avoiding long queue backlogs.

[0035] In terms of congestion control, this invention also demonstrates outstanding performance. The peak occupancy rate of highly congested nodes decreased from 94.7% to 78.2%, a reduction of 17.36%, effectively preventing nodes from entering the performance degradation range. The system-level average congestion rate decreased from 73.3% to 55.7%, a reduction of 24.03%, indicating that the node-level migration threshold configuration successfully achieved dynamic load balancing among nodes in multiple rounds of scheduling, reducing the risk of long-term congestion accumulation.

[0036] Regarding migration quality, this invention significantly improves the link stability of keyframe access. Keyframe access latency fluctuation decreased from 41.2 milliseconds to 22.5 milliseconds, a reduction of 45.39%, indicating that this invention significantly reduces the impact of migration on keyframe-intensive tasks by constraining the frequency and scale of migration actions through the migration oscillation index. Simultaneously, the migration failure rate decreased from 6.8% to 2.1%, a reduction of 69.12%, fully demonstrating the invention's ability to provide refined estimations of migration costs, bandwidth usage, and node I / O changes, making migration decisions safer and more reliable.

[0037] Regarding system stability, this invention reduces the scheduling window performance fluctuation index from 1.00 to 0.58, a decrease of 42%. This demonstrates that within a multi-round scheduling loop, the continuous update mechanism of the normalized net benefit index and migration oscillation index can continuously correct scheduling behavior, enabling the system to maintain a more consistent service level across different business periods. Simultaneously, the number of round-trip migrations decreases from 13 to 5, a reduction of 61.54%, proving that the migration threshold update rule avoids the repeated migration phenomenon common in traditional scheduling, making the task execution path more stable.

[0038] In summary, the data in the table fully demonstrates that the present invention, through five core mechanisms—congestion baseline construction, comparison of completion time of two schemes, quantification of revenue oscillation, adaptive threshold control, and cross-window closed-loop update—significantly improves the scheduling efficiency, congestion control, migration quality, and system stability of heterogeneous video storage systems, achieving fine scheduling capabilities and long-term operational stability that traditional mechanisms cannot achieve.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A task scheduling method based on a heterogeneous video storage system, characterized in that, Includes the following steps: Obtain the utilization rate of each storage node, the waiting time of video tasks, and the queue length in the heterogeneous video storage system; calculate the congestion of storage nodes and summarize them to form a system congestion baseline. Within each scheduling window, an actual scheduling scheme for video tasks is generated based on the system congestion baseline, and a virtual no-migration baseline scheduling scheme that prohibits task migration is constructed to obtain the set of completion times for the same batch of video tasks under the two scheduling schemes. Based on the difference in completion time between the virtual no-migration baseline scheduling scheme and the actual video task scheduling scheme, and combined with the preset task importance weight, node congestion changes and round-trip migration times, the normalized net benefit index and migration oscillation index are calculated. The system congestion baseline, normalized net revenue index and migration oscillation index are input into the adaptive migration threshold update rule to obtain the adaptive migration threshold of the scheduling window and generate the node-level migration threshold configuration for each storage node. For video tasks located on highly congested nodes, estimate the completion time and migration cost under the two scenarios of keeping the migration unchanged and performing the migration, and calculate the net benefit of video task migration; When the net benefit of video task migration exceeds the corresponding node-level migration threshold and meets the migration constraints, video task rescheduling is performed, and the normalized net benefit index and migration oscillation index of subsequent scheduling windows are updated based on the execution results to form a closed-loop control.

2. The task scheduling method based on a heterogeneous video storage system according to claim 1, characterized in that, The generation of the system congestion baseline specifically includes: In a heterogeneous video storage system, the processor utilization rate, input / output bandwidth utilization rate, and disk read / write utilization rate of each storage node are periodically collected to obtain the original sampling sequence of each storage node utilization rate. Denoising processing, outlier removal processing, and time alignment processing are then performed to generate a cleaned sequence of each storage node utilization rate. The video tasks queued on each storage node are statistically analyzed. The time difference between the time when each video task enters the queue and the current time is calculated to obtain the original sampling sequence of the waiting time of video tasks on each storage node. Interpolation and truncation processing are performed to generate the cleaned sequence of the waiting time of video tasks on each storage node. The number of unfinished video tasks in each storage node queue is counted to obtain the original sampled sequence of the queue length of each storage node. The sequence is then normalized according to the preset maximum queue capacity to generate the normalized sequence of the queue length of each storage node. Perform interval normalization processing on the storage node utilization cleaning sequence and the video task waiting time cleaning sequence, and concatenate them with the queue length normalization sequence on a unified time scale to form a local congestion feature vector sequence. Based on the local congestion feature vector sequence, the normalized values ​​of storage node utilization, video task waiting time, and queue length are weighted and summed according to the preset weighting rules to obtain the storage node congestion sequence of each storage node. The sequence is then averaged over time within the current scheduling window to generate the storage node congestion degree of the scheduling window. Within the current scheduling window, the congestion of each storage node in the scheduling window is fused at the node level. Based on the storage capacity, video task type distribution, and service level weight of each storage node, a system-level weighting coefficient is generated. The system-level weighting coefficient is then summed to obtain the system congestion baseline that represents the overall load status of the current scheduling window.

3. The task scheduling method based on a heterogeneous video storage system according to claim 1, characterized in that, The generation of the completion time set specifically includes: At the beginning of each scheduling window, based on the current system congestion baseline, a set of video tasks in the pending scheduling state is obtained, and their video task type label, video access mode label, key frame priority label, cold and hot storage level label and service level label are associated to form a video task description set for the scheduling window. For each video task in the video task description set in the scheduling window, the comprehensive priority of the video task is calculated. The comprehensive priority sequence of video tasks in the scheduling window is obtained by linearly weighting the service level weight, video task type weight, video access mode and key frame priority, and cold and hot storage level weight according to the preset adjustment coefficient. Within the scheduling window, for each video task, candidate storage nodes in the heterogeneous video storage system are traversed. Based on the congestion of storage nodes in the scheduling window, the system congestion baseline, and the estimated queuing waiting time and processing time of video tasks on candidate storage nodes, the scheduling matching score of candidate storage nodes for video tasks is calculated to form a scheduling matching score matrix. Based on the scheduling matching score matrix, target storage nodes are selected for video tasks according to the rule of scheduling matching scores from high to low, so as to obtain the actual scheduling scheme of video tasks facing the scheduling window and calculate the completion time of video tasks. While maintaining the initial binding relationship between each video task and each storage node at the start of the scheduling window, a virtual no-migration baseline scheduling scheme that prohibits task migration is constructed. The storage nodes are sorted according to the arrival order of the video tasks. The expected start time and expected execution duration are estimated based on the execution capabilities of the corresponding storage nodes, and the completion time of the video tasks is calculated. The scheduling scheme and video task completion time are matched to generate a set of video task completion times for the same batch of video tasks in the actual video task scheduling scheme and in the virtual no-migration baseline scheduling scheme.

4. The task scheduling method based on a heterogeneous video storage system according to claim 1, characterized in that, The generation of the normalized net return index and the migration oscillation index specifically includes: Within each scheduling window, based on the difference between the video task completion time of the actual video task scheduling scheme and the virtual no-migration baseline scheduling scheme, the video task completion time difference value is obtained. The ratio of the difference value to the video task completion time in the virtual no-migration baseline scheduling scheme is used as the video task completion time saving ratio, thus obtaining the video task completion time difference value sequence and the video task completion time saving ratio sequence within the scheduling window. The importance of video task descriptions in the scheduling window is characterized according to preset rules. Video task type, video access mode, key frame priority, cold and hot storage level and service level are mapped to service level score, video task type score, video access mode and key frame priority score and cold and hot storage level score, respectively. The task importance weight calculation function is input and linearly weighted according to preset adjustment coefficient to form the preset task importance weight sequence of the scheduling window. Within the scheduling window, the normalized net benefit of each video task is calculated by multiplying the preset task importance weight corresponding to each video task by the video task completion time saving ratio, and then aggregated in a weighted average manner to obtain the normalized net benefit index. Within the same scheduling window, video tasks that have undergone migration operations in heterogeneous video storage systems are filtered, and congestion fluctuation components, round-trip migration count components, and keyframe access delay fluctuation components are calculated. The migration oscillation index is then input into the migration oscillation index calculation function and weighted and summed according to a preset weighting rule to obtain the migration oscillation index.

5. The task scheduling method based on a heterogeneous video storage system according to claim 1, characterized in that, The generation of the node-level migration threshold configuration specifically includes: Within each scheduling window, the system congestion baseline, normalized net revenue index, and migration oscillation index are input into the adaptive migration threshold update rule. The system congestion baseline is compared with the preset congestion reference interval, and the basic migration threshold of the scheduling window is generated based on the overall load level of the current scheduling window. Based on the basic migration threshold of the scheduling window, the scheduling revenue adjustment is calculated using the normalized net revenue index, and the amplitude is truncated to generate the scheduling revenue adjustment result that represents the degree of influence of the overall scheduling revenue of the current scheduling window on the migration threshold. The oscillation suppression adjustment amount is calculated using the migration oscillation index and the amplitude is truncated to generate the oscillation suppression adjustment result that characterizes the degree of impact of migration on the stability of video services within the current scheduling window; The scheduling window's basic migration threshold, scheduling revenue adjustment result, and oscillation suppression adjustment result are superimposed according to a preset combination rule to obtain the instantaneous migration threshold of the scheduling window. This instantaneous migration threshold is then smoothed by a sliding weighted average with the historical migration thresholds of the previous scheduling windows. Preset upper and lower bound constraints on the migration threshold are applied to generate the adaptive migration threshold of the scheduling window. Based on the adaptive migration threshold of the scheduling window, and taking into account the storage medium type, storage capacity, service level weight of each storage node, as well as the migration success rate, migration failure rate and number of round trips in the historical scheduling window, a node-level offset factor is applied to generate a node-level migration threshold for each storage node. The node-level migration threshold is then organized into a node-level migration threshold configuration and distributed to each storage node in the heterogeneous video storage system.

6. The task scheduling method based on a heterogeneous video storage system according to claim 1, characterized in that, The generation of the net benefit from video task migration specifically includes: From the heterogeneous video storage system, storage nodes with a congestion level higher than a preset high congestion threshold in the scheduling window are selected as a set of high-congestion storage nodes. Video tasks that are currently in a queue or have not yet been completed are extracted to generate a video task migration candidate set. Iterate through all storage nodes except for the high-congestion storage nodes where the video tasks are located in the video task migration candidate set. Add storage nodes whose congestion level in the scheduling window is lower than the preset migration congestion threshold and whose corresponding node-level migration threshold in the node-level migration threshold configuration is not higher than the scheduling window adaptive migration threshold to the migration target storage node candidate set, and record the mapping relationship to form a migration target storage node candidate mapping table. For the candidate mapping table of migration target storage nodes, the estimated remaining queuing time of the video task on its current high-congestion storage node is added to the estimated execution time to obtain the completion time of the video task without migration. For the candidate mapping table of the target storage node, the estimated data migration time, the estimated queuing time, and the estimated execution time required to migrate the video task from the high-congestion storage node to the target storage node are added together to obtain the completion time of the video task under the migration scenario. Subtract the completion time under the condition of not migrating from the completion time under the condition of performing migration to obtain the video task migration time benefit for the corresponding video task; Combining the network bandwidth distribution of the heterogeneous video storage system, the input / output bandwidth utilization rate and disk read / write utilization rate of each storage node, the video task migration cost is calculated for each pair of highly congested storage nodes and the migration target storage node. The estimated data migration time, the additional input / output load increment introduced on the source and migration target storage nodes during the migration process, and the disturbance penalty factor caused by the key frame access delay are weighted by corresponding weighting coefficients and then summed to obtain the video task migration cost. For each pair of high-congestion storage nodes and migration target storage nodes, the net benefit of video task migration is calculated by subtracting the video task migration cost from the video task migration time benefit of the corresponding video task. Among all the migration target storage node candidates for the same video task, the migration target storage node with the largest net benefit of video task migration is selected, and the corresponding net benefit of video task migration is recorded as the video task migration decision result. The results are then summarized to form the video task migration net benefit result set.

7. The task scheduling method based on a heterogeneous video storage system according to claim 1, characterized in that, The generation of the closed-loop control specifically includes: Within each scheduling window, obtain the net benefit set of video task migration and the node-level migration threshold configuration, set migration constraints, compare the net benefit of video task migration with the corresponding node-level migration threshold and migration constraints, and obtain a set of candidate migration actions. The candidate migration action set is sorted from high to low according to the net benefit of video task migration. The number of concurrent migrations and the expected network bandwidth usage of each storage node are dynamically counted. If a new candidate migration action still meets the migration constraints after being added, it is added to the video task rescheduling plan. The video task rescheduling plan is sent to the corresponding source storage node and the migration target storage node. The task removal operation in the task queue of the source storage node, the task insertion operation in the task queue of the migration target storage node, and the video data migration trigger operation are executed in sequence, and the video task rescheduling execution record set is collected. Based on the set of video task rescheduling execution records, the actual video task completion time is calculated, the expected start time and expected execution duration of the corresponding video task in the actual video task scheduling scheme are updated, the updated actual video task scheduling scheme is obtained, and the video task completion time difference, key frame access delay fluctuation and storage node congestion change before and after migration are re-statistically analyzed within the scheduling window to obtain the execution deviation statistics. Based on the execution deviation statistics, the historical statistics of the normalized net return index and migration oscillation index are updated across scheduling windows, and the adjustment coefficients are updated in the task importance weight calculation function and migration oscillation index calculation function to complete the closed-loop control link.

8. A task scheduling system based on a heterogeneous video storage system, executing the task scheduling method based on a heterogeneous video storage system as described in any one of claims 1 to 7, characterized in that, include: The congestion baseline module is used to obtain the utilization rate of each storage node, the waiting time of video tasks, and the queue length, calculate the congestion degree of storage nodes, and form the system congestion degree baseline. The scheduling generation module is used to generate actual scheduling schemes for video tasks based on the system congestion baseline within each scheduling window, and to construct a virtual non-migration baseline scheduling scheme that prohibits task migration, thereby obtaining the set of completion times for video tasks under the two scheduling schemes. The revenue oscillation module is used to calculate the normalized net revenue index and migration oscillation index based on the difference in completion time between the two scheduling schemes, combined with the task importance weight, node congestion changes and round-trip migration times. The threshold update module is used to input the system congestion baseline, normalized net revenue index and migration oscillation index into the adaptive migration threshold update rule to generate the scheduling window adaptive migration threshold and node-level migration threshold configuration. The migration benefit module is used to estimate the completion time and migration cost for video tasks located on highly congested nodes under two scenarios: keeping the migration in place and performing the migration, and to calculate the net migration benefit of the video task. The rescheduling module is used to perform video task rescheduling and update the normalized net return index and migration oscillation index based on the rescheduling execution results, forming a closed-loop control.

Images