A computing power allocation method based on dynamic windows and multi-version correction

Abstract

This invention discloses a computing power allocation method based on dynamic windows and multi-version correction, belonging to the field of intelligent resource scheduling technology. This method constructs a scheduling state set by real-time monitoring of the task set to be scheduled and the platform resource status of the cloud platform resource scheduling system; extracts a stable and effective task set based on dynamic judgment of time windows; calculates scheduling priority coefficients by combining task resource density, urgency, and historical completion performance to determine the priority scheduling set and execute the first version of the scheduling scheme; evaluates the gap between the current scheduling version and the historical best scheme through residual vectors and offset coefficients to determine whether version correction is needed; and calculates the feedback change rate based on resource usage feedback during scheduling execution to determine whether the scheduling execution state is stable. If unstable, the scheduling parameters are dynamically adjusted. This method achieves precise allocation of computing power resources, adaptive scheduling optimization, and anomaly risk control.

Description

Technical Field

[0001] This invention relates to the field of intelligent resource scheduling technology, specifically a computing power allocation method based on dynamic windows and multi-version correction. Background Technology

[0002] With the widespread application of cloud computing, edge computing, and high-performance computing platforms, massive heterogeneous computing tasks are submitted concurrently in multi-tenant shared environments. Computing resource scheduling systems face complex scheduling scenarios, such as large differences in task attributes, rapid changes in resource status, and strong fluctuations in execution results. In this context, traditional static scheduling algorithms or single-best strategies are difficult to adapt to the challenges of dynamic loads and uncertain resources, and scheduling efficiency, resource utilization, and task completion stability are all difficult to guarantee.

[0003] Existing scheduling systems mostly rely on the following two types of methods: priority scheduling methods based on fixed weight strategies: these methods set static priorities for tasks or schedule tasks according to their submission time order, ignoring time-varying factors such as task resource density and latency tolerance, lacking adaptability, and resulting in insufficient scheduling accuracy; single-round scheduling methods based on prediction or learning models: although they introduce execution time prediction, resource demand modeling and other means to improve the level of intelligence, they usually only generate a single scheduling scheme, which is difficult to cope with resource state fluctuations and real-time feedback of scheduling results, and has the problem of getting trapped in local optima and lacking correction mechanisms.

[0004] Furthermore, due to the influence of multiple factors such as resource contention and concurrent interference during task execution, there is a significant deviation between the actual completion time and the predicted value. If the system cannot make scheduling adjustments and version updates based on operational feedback, it can easily lead to resource waste, task congestion, or even scheduling failure. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a computing power allocation method based on dynamic windows and multi-version correction to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a computing power allocation method based on dynamic windows and multi-version correction, characterized by comprising the following steps:

[0007] Step 1: Monitor the set of tasks to be scheduled and the status of platform resources in the cloud platform resource scheduling system in real time; collect task request frequency (PL) and resource occupancy growth rate coefficient. The data includes: resource utilization fluctuation value ZB, resource demand vector D for each task, total available resources A, maximum latency tolerance L, task submission time Tj, historical completion time CSJ and predicted completion time CYC for each task instance; after data processing, the data is aggregated into the scheduling state set S.

[0008] Step 2: Based on the scheduling state set S, construct a time window Wt, and extract the task request frequency PL and resource consumption growth rate coefficient within the time window. Calculate and obtain the dynamic window index factor DTY, and compare it with the window stability threshold Dth to determine whether the current window is stable and effective. If it is stable and effective, generate the first set of tasks to be assigned; if it is unstable and ineffective, give a strategy.

[0009] Step 3: Extract the resource requirement vector of each task in the first set of tasks to be assigned through the task attribute mapping table, construct the resource density ZR, task urgency index UI, and historical average completion time CI, calculate the task scheduling priority coefficient PI, and compare it with the scheduling priority threshold Pth. When PI≥Pth, it enters the first priority scheduling set. Then compare the resource density ZR with the resource density threshold Zth. If ZR≥Zth, execute resource allocation immediately and include it in the V1 scheduling scheme; if ZR<Zth, a strategy is given. When PI<Pth, the current scheduling is suspended and it enters the next round of scheduling candidate set.

[0010] Step 4: Extract the historical completion time (CSJ) and predicted completion time (CYC) of each task instance from the historical scheduling records within the n time windows in the task completion performance comparison table, and calculate the residual vector. Further calculate and obtain the residual offset coefficient PYX, and compare and analyze it with the residual tolerance threshold Xth to determine whether the current version scheduling effect is within the acceptable range. If it is, it will directly enter the scheduling sequence to be executed; otherwise, a strategy will be given.

[0011] Step 5: Based on the execution status of each task in the pending scheduling sequence under the real running environment, collect the changes in its core resource usage data over time, construct a task resource usage feedback status set, calculate the feedback change rate Ft, and compare it with the feedback threshold Fth to determine whether the current scheduling execution status is stable. If it is unstable, a strategy is given.

[0012] Preferably, step one includes:

[0013] S11. Real-time monitoring of the set of tasks to be scheduled and the status of platform resources in the cloud platform resource scheduling system;

[0014] S111. Statistically analyze the number of computing task instances in the pending task set that enters the cloud platform resource scheduling system, use the sliding time window statistical method to analyze the number of new requests per unit time, obtain the task request frequency PL, and establish a task arrival time series record.

[0015] S112. Monitor the historical usage of various computing resources on the platform, including: CPU cores, GPU memory capacity, RAM, TPU units, bandwidth channels, and disks; use a multi-period moving average method to analyze the resource usage growth trend per unit time and obtain the resource usage growth rate coefficient. And establish a time series record of resource changes;

[0016] S113. Collect real-time utilization rate change data of each computing node in the resource usage status of the platform, use the standard deviation statistical method to analyze the fluctuation range of resource usage, obtain the resource utilization rate fluctuation value ZB, and establish a resource fluctuation time series record.

[0017] S114. Extract the structural parameters of the computation task instances in the task set to be scheduled, and use the task attribute vector construction method to obtain the resource requirement vector D, total available resources A, maximum latency tolerance L and task submission time Tj for each task, and establish a task attribute mapping table.

[0018] S115. Retrieve historical scheduling records within the last n time windows of the platform resource usage status, use the task trajectory alignment method to obtain the historical completion time CSJ and predicted completion time CYC of each task instance, and establish a task completion performance comparison table.

[0019] S12. Construct a scheduling state set S, and process the collected parameters by noise reduction, anomaly removal, normalization, time series calibration and missing data filling. Summarize the processed data into the scheduling state set S, and simultaneously generate a resource usage statistics table and a task execution record table.

[0020] Preferably, step two includes:

[0021] S21. Based on the scheduling state set S, construct a time window Wt, and extract the task request frequency PL and resource consumption growth rate coefficient within the time window. After dimensionless processing, the dynamic window exponent factor DTY is calculated and obtained.

[0022] Preferably, step two also includes:

[0023] S22. Based on the judgment of window validity, a window stability threshold Dth is set, and the dynamic window exponent factor DTY is compared and analyzed with the window stability threshold Dth to obtain the first evaluation result, including:

[0024] When the dynamic window exponent factor DTY is greater than or equal to the window stability threshold Dth, it indicates that the current window is stable and valid, and the first set of tasks to be assigned is generated.

[0025] When the dynamic window index factor DTY < the window stability threshold Dth, it indicates that the current window is unstable and invalid, the task load and resource status are unstable, and there is a risk that the system's scheduling efficiency will be affected. This triggers the first warning instruction and generates the first strategy: increase the resource monitoring frequency by 20% to detect abnormal fluctuations in system load by monitoring resource changes more frequently; queue or postpone the scheduling of current task requests and reallocate them when resources are sufficient; expand the dynamic window and adjust the time range of the dynamic window to cope with changes in task requests; recalculate until the dynamic window index factor DTY ≥ the window stability threshold Dth.

[0026] Preferably, step three includes:

[0027] S31. Extract the resource requirement vector of each task in the first set of tasks to be assigned from the task attribute mapping table, obtain the proportion of the total available resources, and after dimensionless processing, construct the resource density ZR.

[0028] S32. Extract the maximum latency tolerance L and the time Tn when the scheduling request is received for each task from the task attribute mapping table. Combine this with the task submission time Tj, and after dimensionless processing, calculate and obtain the task urgency index UI.

[0029] Preferably, step three also includes:

[0030] S33. By constructing the resource intensity ZR and task urgency index UI, and combining them with the corresponding historical average completion time CI, after dimensionless processing, the task scheduling priority coefficient PI is calculated and obtained.

[0031] S34. Prioritize and judge all tasks in the first set of tasks to be assigned, set a scheduling priority threshold Pth and a resource density threshold Zth, and compare and analyze the task scheduling priority coefficient PI with the scheduling priority threshold Pth to obtain the second evaluation results, including:

[0032] When the task scheduling priority coefficient PI is greater than or equal to the scheduling priority threshold Pth, it enters the first priority scheduling set. Then, the resource density ZR is compared and analyzed with the resource density threshold Zth. If the resource density ZR is greater than or equal to the resource density threshold Zth, resource allocation is immediately performed and it is included in the V1 scheduling scheme. If the resource density ZR is less than the resource density threshold Zth, a second warning instruction is triggered, and a second strategy is generated: dynamic resource downsizing technology is used to reduce the number of CPU cores, GPU memory, memory and bandwidth in the task request by 20% and reduce the precision by 15%. In the case of insufficient resources, scheduling is completed and it is included in the V1 scheduling scheme.

[0033] When the task scheduling priority coefficient PI is less than the scheduling priority threshold Pth, the current scheduling is suspended, and the task enters the next round of scheduling candidate set for recalculation until the task scheduling priority coefficient PI is greater than or equal to the scheduling priority threshold Pth.

[0034] Preferably, step four includes:

[0035] S41. Extract the historical completion time CSJ and predicted completion time CYC of each task instance corresponding to the historical scheduling records within n time windows from the task completion performance comparison table, construct a version comparison dataset, and perform time alignment and record filtering based on task number;

[0036] S42. Extract the historical completion time CSJ and the predicted completion time CYC. After dimensionless processing, calculate the residual vector for each task in the current version scheduling scheme V1. .

[0037] Preferably, step four also includes:

[0038] S43. By retrieving the optimal version with the lowest mean residual vector from historical versions, set the mean residual value to be... And combined with the mean of the residual vector in the current version, after dimensionless processing, the residual offset coefficient PYX is calculated and obtained;

[0039] S44. Set the residual tolerance threshold Xth, and compare and analyze the residual offset coefficient PYX with the residual tolerance threshold Xth to obtain the third evaluation results, including:

[0040] When the residual offset coefficient PYX ≤ the residual tolerance threshold Xth, it means that the current version of the scheduling effect is within the acceptable range, the system determines it as a valid scheduling version, and directly enters the scheduling sequence to be executed;

[0041] When the residual offset coefficient PYX > the residual tolerance threshold Xth, it indicates that the current version scheduling effect is not within the acceptable range. The system judges it as an invalid scheduling version, which poses a risk of system performance regression. This triggers the third warning instruction and generates the third strategy: start the reconstruction of the structure mapping based on the historical best version, adjust the current scheduling order and resource allocation structure, build a new scheduling scheme Vx, replace the current version cache and mark the correction status.

[0042] Preferably, step five includes:

[0043] S51. Based on the execution status of each task in the pending scheduling sequence under the real operating environment, collect the changes in their core resource usage data over time, and construct a task resource usage feedback status set, including: the rate of change of resource utilization per unit time. and current standard deviation of resource usage .

[0044] Preferably, step five also includes:

[0045] S52, using the rate of change in resource utilization per unit time and current standard deviation of resource usage After dimensionless processing, the feedback rate of change Ft is calculated and obtained;

[0046] S53. Set a feedback threshold Fth, and compare the feedback change rate Ft with the feedback threshold Fth to obtain the fourth evaluation results, including:

[0047] When the feedback change rate Ft ≤ the feedback threshold Fth, it indicates that the current scheduling execution status is stable and no adjustment is needed; continuous monitoring is required.

[0048] When the feedback change rate Ft > the feedback threshold Fth, it indicates that the current scheduling execution state is unstable, triggering the fourth warning instruction and generating the fourth strategy: adopting adaptive scheduling parameter update technology to dynamically adjust the resource allocation ratio, priority evaluation coefficient and version replacement threshold parameters in the next round of task scheduling, and increasing the sampling frequency by 10% and the window sliding amplitude by 8% to realize the scheduling system's timely response to changes in task load and self-learning optimization.

[0049] This invention provides a computing power allocation method based on dynamic windows and multi-version correction. It has the following beneficial effects:

[0050] (1) The computing power allocation method based on dynamic window and multi-version correction constructs a dynamic time window and calculates the window index factor DTY. Combined with the stability threshold, it judges the trend of task requests and resource changes, thereby ensuring that the scheduling operation is only executed when the window is stable, avoiding scheduling errors caused by sudden increase in tasks or sudden change in resources.

[0051] (2) This computing power allocation method based on dynamic window and multi-version correction introduces resource density ZR, task urgency UI and task scheduling priority coefficient PI. The system can sort tasks based on task structure and time requirements, realize priority scheduling of high-priority and resource-efficient tasks, and improve the overall computing power utilization efficiency.

[0052] (3) This computing power allocation method based on dynamic window and multi-version correction analyzes the residual offset between the current version scheduling effect and the historical best solution, introduces correction strategy Vx, and ensures that the resource allocation structure has traceability and version self-repair capability, thereby improving the global optimal performance and stability of the scheduling system.

[0053] (4) This computing power allocation method based on dynamic window and multi-version correction collects the rate of change and standard deviation of resource usage in real time during the task execution phase, and judges the stability of system operation by feedback rate of change Ft. If it becomes unstable, it triggers strategy self-adjustment and dynamically updates scheduling parameters to achieve self-learning and continuous optimization in response to load changes. Attached Figure Description

[0054] Figure 1 This is a schematic diagram illustrating the steps of a computing power allocation method based on dynamic windows and multi-version correction according to the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1

[0057] Please see Figure 1 This invention provides a computing power allocation method based on dynamic windows and multi-version correction, comprising the following steps:

[0058] Step 1: Monitor the set of tasks to be scheduled and the status of platform resources in the cloud platform resource scheduling system in real time; collect task request frequency (PL) and resource occupancy growth rate coefficient. The data includes: resource utilization fluctuation value ZB, resource demand vector D for each task, total available resources A, maximum latency tolerance L, task submission time Tj, historical completion time CSJ and predicted completion time CYC for each task instance; after data processing, the data is aggregated into the scheduling state set S.

[0059] Step 2: Based on the scheduling state set S, construct a time window Wt, and extract the task request frequency PL and resource consumption growth rate coefficient within the time window. Calculate and obtain the dynamic window index factor DTY, and compare it with the window stability threshold Dth to determine whether the current window is stable and effective. If it is stable and effective, generate the first set of tasks to be assigned; if it is unstable and ineffective, give a strategy.

[0060] Step 3: Extract the resource requirement vector of each task in the first set of tasks to be assigned through the task attribute mapping table, construct the resource density ZR, task urgency index UI, and historical average completion time CI, calculate the task scheduling priority coefficient PI, and compare it with the scheduling priority threshold Pth. When PI≥Pth, it enters the first priority scheduling set. Then compare the resource density ZR with the resource density threshold Zth. If ZR≥Zth, execute resource allocation immediately and include it in the V1 scheduling scheme; if ZR<Zth, a strategy is given. When PI<Pth, the current scheduling is suspended and it enters the next round of scheduling candidate set.

[0061] Step 4: Extract the historical completion time (CSJ) and predicted completion time (CYC) of each task instance from the historical scheduling records within the n time windows in the task completion performance comparison table, and calculate the residual vector. Further calculate and obtain the residual offset coefficient PYX, and compare and analyze it with the residual tolerance threshold Xth to determine whether the current version scheduling effect is within the acceptable range. If it is, it will directly enter the scheduling sequence to be executed; otherwise, a strategy will be given.

[0062] Step 5: Based on the execution status of each task in the pending scheduling sequence under the real running environment, collect the changes in its core resource usage data over time, construct a task resource usage feedback status set, calculate the feedback change rate Ft, and compare it with the feedback threshold Fth to determine whether the current scheduling execution status is stable. If it is unstable, a strategy is given.

[0063] In this embodiment, a multi-level dynamic threshold comparison mechanism is introduced, which sets a task scheduling priority coefficient threshold Pth, a resource density threshold Zth, a residual tolerance threshold Xth, and a feedback change rate threshold Fth in stages such as task selection, resource allocation, version correction, and execution feedback. This ensures that each round of scheduling decision has clear discrimination boundaries and dynamic adaptability, effectively improving the robustness, adaptability, and overall resource scheduling efficiency in the task scheduling process.

[0064] Example 2

[0065] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, step one includes:

[0066] S11. Real-time monitoring of the set of tasks to be scheduled and the status of platform resources in the cloud platform resource scheduling system;

[0067] S111. Statistically analyze the number of computing task instances in the pending task set that enters the cloud platform resource scheduling system, use the sliding time window statistical method to analyze the number of new requests per unit time, obtain the task request frequency PL, and establish a task arrival time series record.

[0068] S112. Monitor the historical usage of various computing resources on the platform, including: CPU cores, GPU memory capacity, RAM, TPU units, bandwidth channels, and disks; use a multi-period moving average method to analyze the resource usage growth trend per unit time and obtain the resource usage growth rate coefficient. And establish a time series record of resource changes;

[0069] S113. Collect real-time utilization rate change data of each computing node in the resource usage status of the platform, use the standard deviation statistical method to analyze the fluctuation range of resource usage, obtain the resource utilization rate fluctuation value ZB, and establish a resource fluctuation time series record.

[0070] S114. Extract the structural parameters of the computation task instances in the task set to be scheduled, and use the task attribute vector construction method to obtain the resource requirement vector D, total available resources A, maximum latency tolerance L and task submission time Tj for each task, and establish a task attribute mapping table.

[0071] S115. Retrieve historical scheduling records within the last n time windows of the platform resource usage status, use the task trajectory alignment method to obtain the historical completion time CSJ and predicted completion time CYC of each task instance, and establish a task completion performance comparison table.

[0072] S12. Construct a scheduling state set S, and process the collected parameters by noise reduction, anomaly removal, normalization, time series calibration and missing data filling. Summarize the processed data into the scheduling state set S, and simultaneously generate a resource usage statistics table and a task execution record table.

[0073] In this embodiment, by constructing a scheduling state set S and systematically collecting and standardizing preprocessing data on task requests, resource usage, task attributes, and execution performance, including noise reduction, anomaly removal, normalization, time series calibration, and missing data filling, this invention significantly improves the scheduling system's comprehensive perception capability of complex heterogeneous resources and multi-source dynamic tasks. It provides a high-quality, stable, and reliable data foundation for subsequent dynamic scheduling strategies, enhancing the accuracy and real-time performance of overall scheduling decisions.

[0074] Example 3

[0075] This embodiment is an explanation based on Embodiment 2. Please refer to it. Figure 1 Specifically, step two includes:

[0076] S21. Based on the scheduling state set S, construct a time window Wt, and extract the task request frequency PL and resource consumption growth rate coefficient within the time window. After dimensionless processing, the dynamic window exponent factor DTY is calculated using the following formula:

[0077] Construct a time window Wt that covers the frequency of task request arrivals and changes in resource utilization. The length of the time window is typically adjusted dynamically based on the interval between task requests and real-time changes in resources.

[0078]

[0079] In the formula, Let t represent the resource consumption growth rate coefficient at time t, and e represent the base of the natural constant.

[0080] In this embodiment, by constructing a dynamically adjusted time window Wt and introducing a dynamic window index factor DTY, the present invention comprehensively measures the growth trend of task request frequency and resource consumption. This allows the invention to judge the stability and fluctuation of the scheduling environment in real time, thereby achieving intelligent filtering and dynamic adaptation of the task scheduling window. This improves the sensitivity and response efficiency of the scheduling strategy to changes in resource status and enhances the robustness and adaptability of the system in high-concurrency scenarios.

[0081] Example 4

[0082] This embodiment is an explanation based on Embodiment 3. Please refer to it. Figure 1 Specifically, step two also includes:

[0083] S22. Based on the judgment of window validity, a window stability threshold Dth is set, and the dynamic window exponent factor DTY is compared and analyzed with the window stability threshold Dth to obtain the first evaluation result, including:

[0084] When the dynamic window exponent factor DTY is greater than or equal to the window stability threshold Dth, it indicates that the current window is stable and valid, and the first set of tasks to be assigned is generated.

[0085] When the dynamic window index factor DTY < the window stability threshold Dth, it indicates that the current window is unstable and invalid, the task load and resource status are unstable, and there is a risk that the system's scheduling efficiency will be affected. This triggers the first warning instruction and generates the first strategy: increase the resource monitoring frequency by 20% to detect abnormal fluctuations in system load by monitoring resource changes more frequently; queue or postpone the scheduling of current task requests and reallocate them when resources are sufficient; expand the dynamic window and adjust the time range of the dynamic window to cope with changes in task requests; recalculate until the dynamic window index factor DTY ≥ the window stability threshold Dth.

[0086] The method for setting the window stability threshold Dth is as follows: By statistically analyzing the operation logs of the cloud platform under various task scheduling conditions such as high concurrency and low load, the fluctuation range of the dynamic window index factor composed of task request frequency and resource consumption growth rate is extracted. Combined with the response stability and latency characteristics of the scheduling system under different load pressures, the critical criteria for window stability are determined. Referring to the sliding window scheduling stability evaluation specifications of typical cloud scheduling systems at home and abroad, the processing cycle response threshold of mainstream scheduling frameworks, and the experience feedback of scheduling engineering experts, this threshold is set to accurately reflect the coupling balance between task load and resource status, promptly identify the instability risk of the scheduling window, and ensure the effectiveness of the dynamic window mechanism and the reliability of the pre-analysis.

[0087] In this embodiment, by setting a window stability threshold Dth and introducing a comparison mechanism of dynamic window index factor DTY, the present invention can promptly identify whether the resource and task status within the scheduling window is within a stable and controllable range. When the window is unstable, it can automatically trigger multi-strategy linkage adjustment, including increasing the frequency of resource monitoring, delaying the scheduling of non-critical tasks, and dynamically expanding the time window, thereby effectively avoiding the risk of scheduling efficiency decline caused by sudden changes in resource load and significantly improving the scheduling robustness and task execution reliability of the cloud platform in complex dynamic environments.

[0088] Example 5

[0089] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, step three includes:

[0090] S31. Extract the resource requirement vector of each task in the first set of tasks to be assigned through the task attribute mapping table, obtain the proportion of the total available resources, and after dimensionless processing, construct the resource density ZR, as follows:

[0091]

[0092] In the formula, This represents the amount of resource type k required by the i-th task. This represents the total available amount of the k-th type of resource;

[0093] S32. Extract the maximum latency tolerance L and the time Tn when the scheduling request is received for each task from the task attribute mapping table. Combine this with the task submission time Tj, and after dimensionless processing, calculate the task urgency index UI. The formula is as follows:

[0094]

[0095] When UI is close to 1, it means that the task is close to its maximum acceptable delay and needs to be scheduled as soon as possible; when UI is greater than 1, it means that the task is an overdue task and its priority should be automatically increased or marked as an abnormal task.

[0096] In this embodiment, by constructing the resource intensity ZR and the task urgency index UI, the present invention can accurately quantify the resource occupation intensity and time sensitivity of the task before scheduling, realize the dual evaluation of the task's "resource consumption characteristics" and "timeliness requirements", and thus provide a quantitative basis for subsequent scheduling priority determination and resource allocation strategy, significantly improving the pertinence of scheduling decisions and the timeliness of task response.

[0097] Example 6

[0098] This embodiment is an explanation based on Embodiment 5. Please refer to it. Figure 1 Specifically, step three also includes:

[0099] S33. By constructing the resource intensity ZR and task urgency index UI, and combining them with the corresponding historical average completion time CI, after dimensionless processing, the task scheduling priority coefficient PI is calculated and obtained, as shown in the following formula:

[0100]

[0101] In the formula, a1, a2, and a3 represent weighting coefficients;

[0102] The setting method for a1, a2, and a3: Through systematic analysis of historical task scheduling data, statistics are compiled on the resource intensity of different tasks. urgency of the task This study investigates the trends of historical average completion time (CI) across three dimensions and their impact on scheduling timeliness, constructing a correlation model between task attributes and scheduling priority. Combining actual scheduling performance, system feedback data, and expert experience, a weighted analysis is conducted on the impact of each attribute factor on task response speed, resource efficiency, and execution risk. This analysis references task scoring models in typical scheduling platforms, coefficient setting methods in cloud resource optimization scheduling frameworks, and evaluation criteria for task real-time requirements and resource constraint models.

[0103] Resource intensity, as the main source of scheduling pressure, should occupy a major weight;

[0104] The urgency of the task is directly related to service quality assurance and should be given high weight.

[0105] Historical completion time is a comprehensive reflection of task complexity and historical performance, and has some reference value but is highly volatile, so it is given relatively less weight.

[0106] The above coefficient combination has been verified by a large number of simulation experiments. It can effectively improve the rationality and efficiency of task scheduling. It can still maintain good scheduling performance when the competition for task resources is fierce or the system load fluctuates greatly, and has good general adaptability and stability.

[0107] S34. Prioritize and judge all tasks in the first set of tasks to be assigned, set a scheduling priority threshold Pth and a resource density threshold Zth, and compare and analyze the task scheduling priority coefficient PI with the scheduling priority threshold Pth to obtain the second evaluation results, including:

[0108] When the task scheduling priority coefficient PI is greater than or equal to the scheduling priority threshold Pth, it enters the first priority scheduling set. Then, the resource density ZR is compared and analyzed with the resource density threshold Zth. If the resource density ZR is greater than or equal to the resource density threshold Zth, resource allocation is immediately performed and it is included in the V1 scheduling scheme. If the resource density ZR is less than the resource density threshold Zth, a second warning instruction is triggered, and a second strategy is generated: dynamic resource downsizing technology is used to reduce the number of CPU cores, GPU memory, memory and bandwidth in the task request by 20% and reduce the precision by 15%. In the case of insufficient resources, scheduling is completed and it is included in the V1 scheduling scheme.

[0109] When the task scheduling priority coefficient PI is less than the scheduling priority threshold Pth, the current scheduling is suspended, and the task enters the next round of scheduling candidate set for recalculation until the task scheduling priority coefficient PI is greater than or equal to the scheduling priority threshold Pth.

[0110] The scheduling priority threshold Pth is set as follows: By modeling historical scheduling records of multiple typical tasks, the statistical correlation between the task scheduling priority coefficient PI distribution and its actual completion efficiency is extracted. Combined with attributes such as task latency tolerance, resource demand intensity, and average execution time, the characteristic critical region where tasks should be prioritized is identified. Referring to the scoring boundaries in multi-task priority scheduling algorithms, the urgency level classification standards in general task classification systems, and the empirical parameters of scheduling strategy optimization experts within the platform, this threshold is set to accurately reflect the urgency and benefit trade-offs of task scheduling, ensuring timely scheduling and resource allocation response for high-priority tasks.

[0111] The resource intensity threshold Zth is set as follows: By statistically analyzing a large number of task instances in the platform to understand the typical demand structure of various resources (CPU, GPU, memory, bandwidth, etc.), a normalized distribution curve of resource intensity ZR is constructed to identify the demand boundary between resource-intensive tasks and routine tasks. Combining the average scheduling margin, elastic allocation capability, and resource overload response limit of the cloud resource pool, and comprehensively investigating the degradation tolerance parameters and policy trigger points in mainstream resource management strategies, a resource intensity discrimination standard is formulated. This threshold aims to reasonably distinguish between high-resource-dependent tasks and adjustable tasks, improve resource allocation accuracy, and ensure the scheduling system's fine-grained response capability to complex resource combination tasks.

[0112] In this embodiment, by introducing a task scheduling priority coefficient PI and setting a scheduling priority threshold and a resource density threshold, the present invention can realize a refined hierarchical scheduling strategy for tasks driven by multiple factors. This not only ensures that critical tasks are executed first when resources are scarce, but also ensures scheduling continuity through dynamic resource downsizing technology in some resource-insufficient scenarios, thereby effectively improving the resource utilization efficiency and task completion rate of the cloud platform in a high-concurrency environment.

[0113] Example 7

[0114] This embodiment is an explanation based on Embodiment 6. Please refer to it. Figure 1 Specifically, step four includes:

[0115] After completing the allocation of execution resources and incorporating them into the V1 scheduling scheme, in order to avoid the trap of local optima and improve the global reliability of the scheduling scheme, the system introduces a multi-version comparison mechanism to evaluate the residuals of the currently generated scheduling scheme and compare it with the historical best version to determine whether version correction is needed.

[0116] S41. Extract the historical completion time CSJ and predicted completion time CYC of each task instance corresponding to the historical scheduling records within n time windows from the task completion performance comparison table, construct a version comparison dataset, and perform time alignment and record filtering based on task number;

[0117] S42. Extract the historical completion time CSJ and the predicted completion time CYC. After dimensionless processing, calculate the residual vector for each task in the current version scheduling scheme V1. The formula is as follows:

[0118]

[0119] In the formula, This indicates the actual completion time of the current task j. This represents the predicted completion time of the current task j.

[0120] The V1 scheduling scheme, as the first version of the resource scheduling structure generated in step three through the scheduling priority coefficient PI, already includes the task number, resource allocation status, estimated completion time CSJ, and predicted completion time CYC. During task execution, its actual completion time CSJ is recorded synchronously, and this execution record is written to the scheduling log item and stored in the historical scheduling record table. The "historical scheduling record table" data structure extracted in S41 originates from the log records after the execution of the V1 and earlier versions of the scheme in step three. The current version (i.e., the most recent V1 scheduling scheme) serves as the "residual calculation object." Its predicted completion time CYC value directly originates from the task's estimated completion time field calculated in step three, while CSJ is automatically generated by the system after the task is completed and written back to the record table, achieving a closed-loop comparison between prediction and reality. Therefore, the V1 scheduling scheme not only serves as a reference for the current scheduling execution but also constitutes the "evaluated object" in version comparison. Its structural fields directly participate in residual vector calculation and subsequent correction judgment, forming a strong correlation path between steps three and four, ensuring that the scheduling system has the ability to backtrack, correct, and optimize.

[0121] In this embodiment, by introducing a closed-loop comparison mechanism between historical scheduling records and the current version residual vector, the present invention can achieve refined error evaluation of the scheduling scheme after actual execution, effectively avoid the local optimum trap, and ensure the global optimality and execution reliability of the resource scheduling scheme under dynamic load environment through multi-version comparison.

[0122] Example 8

[0123] This embodiment is an explanation based on Embodiment 7. Please refer to it. Figure 1 Specifically, step four also includes:

[0124] S43. By retrieving the optimal version with the lowest mean residual vector from historical versions, set the mean residual value to be... Combined with the mean of the current version's residual vector, after dimensionless processing, the residual offset coefficient PYX is calculated as follows:

[0125]

[0126] In the formula, This represents the mean of the residual vector in the current version. This represents the mean residual of the historical best version;

[0127] S44. Set the residual tolerance threshold Xth, and compare and analyze the residual offset coefficient PYX with the residual tolerance threshold Xth to obtain the third evaluation results, including:

[0128] When the residual offset coefficient PYX ≤ the residual tolerance threshold Xth, it means that the current version of the scheduling effect is within the acceptable range, the system determines it as a valid scheduling version, and directly enters the scheduling sequence to be executed;

[0129] When the residual offset coefficient PYX > the residual tolerance threshold Xth, it indicates that the current version scheduling effect is not within the acceptable range. The system judges it as an invalid scheduling version, which poses a risk of system performance regression. This triggers the third warning instruction and generates the third strategy: start the reconstruction of the structure mapping based on the historical best version, adjust the current scheduling order and resource allocation structure, build a new scheduling scheme Vx, replace the current version cache and mark the correction status.

[0130] The residual tolerance threshold Xth is set as follows: By constructing a residual distribution model of the predicted and actual completion times of historical scheduling versions, the performance deviation of the scheduling system under different load fluctuations is analyzed. Combining the stable residual level of the historical optimal scheduling scheme with the convergence performance of the corrected version, an effective identification boundary for the residual offset trend is extracted. This threshold is jointly determined by the platform operation and maintenance team and scheduling experts, referencing task completion accuracy evaluation standards, scheduling system performance tolerance range setting specifications, and the fault tolerance standards of the scheduling system version control mechanism. This threshold effectively reflects the acceptable range of scheduling version offsets, ensuring the accuracy of the version comparison mechanism and preventing performance-degraded task versions from being mistakenly included in the execution sequence.

[0131] In this embodiment, by introducing a residual offset coefficient PYX and a historical best version comparison and analysis mechanism, the present invention can accurately identify the degree of execution deviation of the current scheduling scheme. If it exceeds the tolerance threshold, the structure mapping reconstruction strategy is automatically triggered to correct the invalid scheduling version in a timely manner, thereby ensuring the stability and global performance of the cloud platform scheduling system in a complex and fluctuating environment.

[0132] Example 9

[0133] This embodiment is an explanation based on Embodiment 8. Please refer to it. Figure 1 Specifically, step five includes:

[0134] After determining the final executable scheduling scheme and including it in the execution sequence, a dynamic adjustment and feedback optimization mechanism is entered to monitor and model the computing power usage in real time during the actual execution of the task, thereby improving the adaptability and stability of the scheduling system.

[0135] S51. Based on the execution status of each task in the pending scheduling sequence under the real operating environment, collect the changes in their core resource usage data over time, and construct a task resource usage feedback status set, including: the rate of change of resource utilization per unit time. and current standard deviation of resource usage .

[0136] Rate of change in resource utilization per unit time The sliding window sampling technique is used to sample the usage sequence of CPU, GPU, memory and bandwidth in real time, calculate the first difference of resource usage per unit time, and obtain the resource utilization change rate data during task execution to reflect the sudden trend of resource usage.

[0137] Current resource usage standard deviation The time window statistical analysis method is used to calculate the standard deviation of the utilization rate of each type of core resource (CPU / GPU / memory / bandwidth) within the current time window, and obtain the resource utilization fluctuation intensity index to measure the stability or dispersion of system operation.

[0138] In this embodiment, by introducing a real-time monitoring mechanism for the rate of change of resource utilization per unit time and the standard deviation of current resource utilization, the present invention can dynamically perceive the sudden changes and fluctuations in resource utilization during task operation, construct a highly timely task resource utilization feedback state set, thereby providing an accurate basis for subsequent scheduling adjustments and system stability control, and significantly improving the scheduling system's ability to respond quickly and adapt to operational anomalies.

[0139] Example 10

[0140] This embodiment is an explanation based on Embodiment 9. Please refer to it. Figure 1 Specifically, step five also includes:

[0141] S52, using the rate of change in resource utilization per unit time and current standard deviation of resource usage After dimensionless processing, the feedback rate of change Ft is calculated using the following formula:

[0142]

[0143] In the formula, This represents the fluctuation weighting adjustment coefficient;

[0144] Volatility Weighting Coefficient The approach involves long-term sampling of execution data for various typical tasks under different scheduling strategies and resource allocation modes. This allows for the extraction of the joint distribution characteristics of the rate of change in resource utilization per unit time and the standard deviation of the usage of various core resources. This comprehensive assessment evaluates the actual impact of different fluctuation patterns on the stability and response performance of the scheduling system. Combining task type characteristics, resource utilization sensitivity analysis results, and system instability early warning cases, a resource fluctuation impact model is constructed to quantify the weighted effect of fluctuations on the feedback rate of change. Referring to the feedback control sensitivity parameter configuration specifications of mainstream cloud scheduling platforms, resource load fluctuation suppression strategies, and the adjustment factor design methods of relevant intelligent scheduling optimization models, the fluctuation weight adjustment coefficient is determined collaboratively by system performance evaluation experts and the platform scheduling strategy team. The reasonable range of values ​​for this coefficient; The aim is to dynamically balance the impact of real-time mutation trends and resource fluctuation intensity on system feedback indicators, ensuring the feedback change rate. It can accurately reflect the stability and disturbance level of the current task execution status, providing a reliable basis for subsequent adaptive parameter adjustments and improving the scheduling system's response efficiency and robustness to load changes.

[0145] S53. Set a feedback threshold Fth, and compare the feedback change rate Ft with the feedback threshold Fth to obtain the fourth evaluation results, including:

[0146] When the feedback change rate Ft ≤ the feedback threshold Fth, it indicates that the current scheduling execution status is stable and no adjustment is needed; continuous monitoring is required.

[0147] When the feedback change rate Ft > the feedback threshold Fth, it indicates that the current scheduling execution state is unstable, triggering the fourth warning instruction and generating the fourth strategy: adopting adaptive scheduling parameter update technology to dynamically adjust the resource allocation ratio, priority evaluation coefficient and version replacement threshold parameters in the next round of task scheduling, and increasing the sampling frequency by 10% and the window sliding amplitude by 8% to realize the scheduling system's timely response to changes in task load and self-learning optimization.

[0148] The feedback threshold Fth is set as follows: A multidimensional distribution model of the feedback change rate Ft is constructed by performing statistical regression analysis on the fluctuation characteristics of the resource utilization change rate and standard deviation under different operating scenarios during task execution, extracting the boundary index between stable and unstable system operation. Combining the response rate requirements of the adaptive scheduling mechanism to load disturbances, the timeliness parameters of the feedback sampling mechanism, and the dynamic fluctuation tolerance range in typical scheduling anomaly scenarios, a feedback stability judgment standard is set. This threshold aims to promptly identify nonlinear fluctuations or resource mutation risks during task operation, triggering adaptive adjustments to scheduling parameters, and enabling the scheduling system to achieve self-learning response and continuous optimization to external load changes.

[0149] In this embodiment, by introducing the feedback change rate Ft and its comparison mechanism with the feedback threshold Fth, the present invention can identify in real time whether the scheduling execution status is within a stable range. Once abnormal fluctuations in resource usage are detected, an adaptive parameter update strategy can be triggered to achieve dynamic optimization and adjustment of resource allocation ratio, scheduling priority coefficient and version correction parameters, significantly improving the scheduling system's response speed and self-learning ability to runtime load fluctuations, and enhancing the stability and evolution of the overall scheduling strategy.

[0150] The threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value, it is acceptable.

[0151] The above formulas are all derived from software simulation using a large amount of data, and are selected to be close to the actual values. The coefficients in the formulas are set by those skilled in the art based on the actual situation. 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 changes 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 computing power allocation method based on dynamic windows and multi-version correction, characterized in that, Includes the following steps: Step 1: Monitor the set of tasks to be scheduled and the status of platform resources in the cloud platform resource scheduling system in real time; collect task request frequency (PL) and resource occupancy growth rate coefficient. The data includes: resource utilization fluctuation value ZB, resource demand vector D for each task, total available resources A, maximum latency tolerance L, task submission time Tj, historical completion time CSJ and predicted completion time CYC for each task instance; after data processing, the data is aggregated into the scheduling state set S. Step 2: Based on the scheduling state set S, construct a time window Wt, and extract the task request frequency PL and resource consumption growth rate coefficient within the time window. Calculate and obtain the dynamic window index factor DTY, and compare it with the window stability threshold Dth to determine whether the current window is stable and effective. If it is stable and effective, generate the first set of tasks to be assigned; if it is unstable and ineffective, give a strategy. Step 3: Extract the resource requirement vector of each task in the first set of tasks to be assigned through the task attribute mapping table, construct the resource density ZR, task urgency index UI, and historical average completion time CI, calculate the task scheduling priority coefficient PI, and compare it with the scheduling priority threshold Pth. When PI≥Pth, it enters the first priority scheduling set. Then compare the resource density ZR with the resource density threshold Zth. If ZR≥Zth, execute resource allocation immediately and include it in the V1 scheduling scheme; if ZR<Zth, a strategy is given. When PI<Pth, the current scheduling is suspended and it enters the next round of scheduling candidate set. Step 4: Extract the historical completion time (CSJ) and predicted completion time (CYC) of each task instance from the historical scheduling records within the n time windows in the task completion performance comparison table, and calculate the residual vector. Further calculate and obtain the residual offset coefficient PYX, and compare and analyze it with the residual tolerance threshold Xth to determine whether the current version scheduling effect is within the acceptable range. If it is, it will directly enter the scheduling sequence to be executed; otherwise, a strategy will be given. Step 5: Based on the execution status of each task in the pending scheduling sequence under the real running environment, collect the changes in its core resource usage data over time, construct a task resource usage feedback status set, calculate the feedback change rate Ft, and compare it with the feedback threshold Fth to determine whether the current scheduling execution status is stable. If it is unstable, a strategy is given.

2. The computing power allocation method based on dynamic windows and multi-version correction according to claim 1, characterized in that, Step one includes: S11. Real-time monitoring of the set of tasks to be scheduled and the status of platform resources in the cloud platform resource scheduling system; S111. Statistically analyze the number of computing task instances in the pending task set that enters the cloud platform resource scheduling system, use the sliding time window statistical method to analyze the number of new requests per unit time, obtain the task request frequency PL, and establish a task arrival time series record. S112. Monitor the historical usage of various computing resources on the platform, including: CPU cores, GPU memory capacity, RAM, TPU units, bandwidth channels, and disks; use a multi-period moving average method to analyze the resource usage growth trend per unit time and obtain the resource usage growth rate coefficient. And establish a time series record of resource changes; S113. Collect real-time utilization rate change data of each computing node in the resource usage status of the platform, use the standard deviation statistical method to analyze the fluctuation range of resource usage, obtain the resource utilization rate fluctuation value ZB, and establish a resource fluctuation time series record. S114. Extract the structural parameters of the computation task instances in the task set to be scheduled, and use the task attribute vector construction method to obtain the resource requirement vector D, total available resources A, maximum latency tolerance L and task submission time Tj for each task, and establish a task attribute mapping table. S115. Retrieve historical scheduling records within the last n time windows of the platform resource usage status, use the task trajectory alignment method to obtain the historical completion time CSJ and predicted completion time CYC of each task instance, and establish a task completion performance comparison table. S12. Construct a scheduling state set S, and process the collected parameters by noise reduction, anomaly removal, normalization, time series calibration and missing data filling. Summarize the processed data into the scheduling state set S, and simultaneously generate a resource usage statistics table and a task execution record table.

3. The computing power allocation method based on dynamic windows and multi-version correction according to claim 2, characterized in that, Step two includes: S21. Based on the scheduling state set S, construct a time window Wt, and extract the task request frequency PL and resource consumption growth rate coefficient within the time window. After dimensionless processing, the dynamic window exponent factor DTY is calculated and obtained.

4. The computing power allocation method based on dynamic windows and multi-version correction according to claim 3, characterized in that, Step two also includes: S22. Based on the judgment of window validity, a window stability threshold Dth is set, and the dynamic window exponent factor DTY is compared and analyzed with the window stability threshold Dth to obtain the first evaluation result, including: When the dynamic window exponent factor DTY is greater than or equal to the window stability threshold Dth, it indicates that the current window is stable and valid, and the first set of tasks to be assigned is generated. When the dynamic window index factor DTY < the window stability threshold Dth, it indicates that the current window is unstable and invalid, the task load and resource status are unstable, and there is a risk that the system's scheduling efficiency will be affected. This triggers the first warning instruction and generates the first strategy: increase the resource monitoring frequency by 20% to detect abnormal fluctuations in system load by monitoring resource changes more frequently; queue or postpone the scheduling of current task requests and reallocate them when resources are sufficient; expand the dynamic window and adjust the time range of the dynamic window to cope with changes in task requests; recalculate until the dynamic window index factor DTY ≥ the window stability threshold Dth.

5. The computing power allocation method based on dynamic windows and multi-version correction according to claim 4, characterized in that, Step three includes: S31. Extract the resource requirement vector of each task in the first task set to be assigned through the task attribute mapping table, obtain the proportion of the total available resources, and construct the resource density ZR after dimensionless processing. S32. Extract the maximum delay tolerance L and the time Tn when the scheduling request is received for each task through the task attribute mapping table. Combine the task submission time Tj with the dimensionless processing, and calculate the task urgency index UI.

6. The computing power allocation method based on dynamic windows and multi-version correction according to claim 5, characterized in that, Step three also includes: S33. By constructing the resource intensity ZR and task urgency index UI, and combining them with the corresponding historical average completion time CI, after dimensionless processing, the task scheduling priority coefficient PI is calculated and obtained. S34. Prioritize and judge all tasks in the first set of tasks to be assigned, set a scheduling priority threshold Pth and a resource density threshold Zth, and compare and analyze the task scheduling priority coefficient PI with the scheduling priority threshold Pth to obtain the second evaluation results, including: When the task scheduling priority coefficient PI is greater than or equal to the scheduling priority threshold Pth, it enters the first priority scheduling set. Then, the resource density ZR is compared and analyzed with the resource density threshold Zth. If the resource density ZR is greater than or equal to the resource density threshold Zth, resource allocation is immediately performed and it is included in the V1 scheduling scheme. If the resource density ZR is less than the resource density threshold Zth, a second warning instruction is triggered, and a second strategy is generated: dynamic resource downsizing technology is used to reduce the number of CPU cores, GPU memory, memory and bandwidth in the task request by 20% and reduce the precision by 15%. In the case of insufficient resources, scheduling is completed and it is included in the V1 scheduling scheme. When the task scheduling priority coefficient PI is less than the scheduling priority threshold Pth, the current scheduling is suspended, and the task enters the next round of scheduling candidate set for recalculation until the task scheduling priority coefficient PI is greater than or equal to the scheduling priority threshold Pth.

7. The computing power allocation method based on dynamic windows and multi-version correction according to claim 6, characterized in that, Step four includes: S41. Extract the historical completion time CSJ and predicted completion time CYC of each task instance corresponding to the historical scheduling records within n time windows from the task completion performance comparison table, construct a version comparison dataset, and perform time alignment and record filtering based on task number; S42. Extract the historical completion time CSJ and the predicted completion time CYC. After dimensionless processing, calculate the residual vector for each task in the current version scheduling scheme V1. .

8. The computing power allocation method based on dynamic windows and multi-version correction according to claim 7, characterized in that, Step four also includes: S43. By retrieving the optimal version with the lowest mean residual vector from historical versions, set the mean residual value to be... And combined with the mean of the residual vector in the current version, after dimensionless processing, the residual offset coefficient PYX is calculated and obtained; S44. Set the residual tolerance threshold Xth, and compare and analyze the residual offset coefficient PYX with the residual tolerance threshold Xth to obtain the third evaluation results, including: When the residual offset coefficient PYX ≤ the residual tolerance threshold Xth, it means that the current version of the scheduling effect is within the acceptable range, the system determines it as a valid scheduling version, and directly enters the scheduling sequence to be executed; When the residual offset coefficient PYX > the residual tolerance threshold Xth, it indicates that the current version scheduling effect is not within the acceptable range. The system judges it as an invalid scheduling version, which poses a risk of system performance regression. This triggers the third warning instruction and generates the third strategy: start the reconstruction of the structure mapping based on the historical best version, adjust the current scheduling order and resource allocation structure, build a new scheduling scheme Vx, replace the current version cache and mark the correction status.

9. A computing power allocation method based on dynamic windows and multi-version correction according to claim 8, characterized in that, Step five includes: S51. Based on the execution status of each task in the pending scheduling sequence under the real operating environment, collect the changes in their core resource usage data over time, and construct a task resource usage feedback status set, including: the rate of change of resource utilization per unit time. and current standard deviation of resource usage .

10. A computing power allocation method based on dynamic windows and multi-version correction according to claim 9, characterized in that, Step five also includes: S52, through the rate of change in resource utilization per unit time and current standard deviation of resource usage After dimensionless processing, the feedback rate of change Ft is calculated and obtained; S53. Set a feedback threshold Fth, and compare the feedback change rate Ft with the feedback threshold Fth to obtain the fourth evaluation results, including: When the feedback change rate Ft ≤ the feedback threshold Fth, it indicates that the current scheduling execution status is stable and no adjustment is needed; continuous monitoring is required. When the feedback change rate Ft > the feedback threshold Fth, it indicates that the current scheduling execution state is unstable, triggering the fourth warning instruction and generating the fourth strategy: adopting adaptive scheduling parameter update technology to dynamically adjust the resource allocation ratio, priority evaluation coefficient and version replacement threshold parameters in the next round of task scheduling, and increasing the sampling frequency by 10% and the window sliding amplitude by 8% to realize the scheduling system's timely response to changes in task load and self-learning optimization.

Images