Desktop cloud AI dynamic resource scheduling method for multiple office loads

Abstract

The invention belongs to the technical field of desktop cloud resource management, and particularly relates to a desktop cloud AI dynamic resource scheduling method for multiple office loads, which comprises the following specific steps: S1, constructing an office load feature library and AI model training: constructing a standardized office load feature library through offline collection of multi-dimensional features of multiple office loads of an enterprise; model training is carried out for load identification, demand prediction and priority; and S2, real-time load sensing and pre-judgment: collecting multi-dimensional operation data of the terminal in real time through a desktop cloud client Agent, and combining the trained AI model. Through a dynamic resource allocation strategy of pre-judgment, reservation, dynamic space complementing and redundant resource recovery and in combination with real-time state perception of the desktop cloud cluster resource pool, the method has the advantages that resources are flexibly adjusted along with periodic and sudden fluctuation of office loads, and idle or insufficient resources caused by static quota are avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of desktop cloud resource management, in particular to a desktop cloud AI dynamic resource scheduling method for multiple office loads. BACKGROUND

[0002] With the promotion of remote office and digital transformation, desktop cloud has become the core infrastructure of enterprise office due to its advantages of centralized management, safety and controllability, cross-terminal access, etc. The office load carried by the current enterprise desktop cloud presents diversity and dynamic characteristics: for example, from 9:00 to 10:00 in the morning, there are more real-time interactive loads such as video conferences and online approvals; from 12:00 to 13:00 in the afternoon, there are more light loads such as document editing and email processing; from 14:00 to 16:00 in the afternoon, there may be more computing-intensive loads such as data modeling and report generation.

[0003] However, the existing desktop cloud resource scheduling method has the following key problems, which are difficult to adapt to the needs of multiple office loads:

[0004] 1. Static resource allocation is rigid and cannot adapt to load fluctuations:

[0005] Existing solutions mostly use fixed quota mode (such as allocating 2-core CPU and 4GB memory to each desktop), without considering the time-varying changes of loads. For example: an Internet company has 1000 desktop cloud terminals, 600 employees simultaneously start video conferences during the morning rush hour (9:00-10:00), each video conference requires 3-core CPU support, but the static allocation is only 2-core, resulting in CPU usage exceeding 110%, causing video lag and sound delay; while during lunchtime (12:00-13:00), only 300 employees use light loads, and the desktop resource idle rate reaches 50%, resulting in serious resource waste.

[0006] 2. Load type recognition bias, low resource matching accuracy:

[0007] The existing scheduling system only uses CPU / memory usage as a single indicator to judge load demand, and cannot distinguish the differentiated resource demands of different load types. For example: a finance department employee simultaneously runs Excel big data report calculation (requires high memory and low network) and cloud document online collaboration (requires high network and low calculation), the existing system mistakenly judges the desktop as light load due to the detection of only 30% CPU usage, and does not supplement the memory resources in time, resulting in frequent Excel report calculation lag (swap exchange triggered by insufficient memory), while the network resources are idle.

[0008] 3. Resource adjustment lag, lack of prediction ability:

[0009] The existing scheduling relies on load-triggered adjustment (i.e. resources are allocated only after the load exceeds a threshold), and the adjustment delay is 3-5 minutes, which cannot meet the real-time office load demand. For example, a project team conducts 3D model rendering + online review at 15:00 every Friday, and this load needs to instantaneously increase GPU resources, but the existing system needs to wait until the GPU usage rate reaches 90% before starting the scheduling, resulting in a 10-minute rendering progress stagnation and a forced delay of the review meeting.

[0010] 4. Multiple load priority confusion, key business affected:

[0011] The existing scheme does not establish a priority mechanism for office loads, and when multiple types of loads are concurrent, non-critical loads may preempt critical resources. For example, a CEO remote board meeting (critical load) and an ordinary employee department file backup (non-critical load) occur simultaneously in a certain enterprise, and the existing system allocates storage resources in priority due to the high disk IO request volume of the file backup, resulting in frame dropping of the board meeting video due to insufficient disk IO, affecting the normal progress of the meeting.

[0012] Based on the above, therefore, the application provides a desktop cloud AI dynamic resource scheduling method for multiple office loads. SUMMARY

[0013] To solve the above technical problems, the application provides the following technical solutions:

[0014] The desktop cloud AI dynamic resource scheduling method for multiple office loads includes the following specific steps:

[0015] S1, constructing an office load feature library and AI model training: through offline collection of multi-dimensional features of multiple office loads of an enterprise, a standardized office load feature library is constructed; at the same time, models are trained for load identification, demand prediction, and priority;

[0016] S2, real-time load sensing and prediction: through real-time collection of multi-dimensional running data of the terminal by the desktop cloud client Agent, and in combination with the trained AI model, the current office load type is first identified in real time, the differentiated resource demand of the load is accurately matched, and then based on user behavior habits and time period characteristics, the load peak and resource demand are predicted 15-30 minutes in advance;

[0017] S3, dynamic resource scheduling decision: based on the output load information, in combination with the real-time state of the desktop cloud cluster resource pool, a multi-objective optimization decision mechanism is constructed, the resource allocation weight is first determined through load priority sorting, and then a combination strategy of prediction reservation, dynamic filling, and threshold protection is adopted to realize accurate matching of resources and loads;

[0018] S4, resource execution and recycling: converting the allocation instructions of S3 into actual resource adjustment actions, on the one hand, dynamically expanding / contracting resources for the target terminal, and on the other hand, through real-time monitoring of load changes, quickly recycling redundant resources to the public resource pool when the load ends or demand decreases, and configuring an emergency resource pool to cope with prediction deviations;

[0019] S5, scheduling effect evaluation and AI model iteration: by collecting load performance indicators, resource matching degree and energy consumption data after scheduling, a scheduling efficiency scoring model is established, and based on the scoring results, the AI model is iterated offline in batches and fine-tuned online in real time, and a reinforcement learning algorithm is used to optimize the energy consumption performance balance strategy.

[0020] As a preferred scheme of the multi-office load-oriented desktop cloud AI dynamic resource scheduling method described in the application, wherein the specific steps of S1 are as follows:

[0021] S11, load feature collection: collecting enterprise historical office data;

[0022] S12, AI model training: first, train the load type classification model and the load demand prediction model, and then build the load priority model.

[0023] As a preferred scheme of the multi-office load-oriented desktop cloud AI dynamic resource scheduling method described in the application, wherein the enterprise historical office data includes load type labels, multi-dimensional feature parameters and resource demand labels.

[0024] As a preferred scheme of the multi-office load-oriented desktop cloud AI dynamic resource scheduling method described in the application, wherein the specific steps of S12 are as follows:

[0025] S121, training the load type classification model: using a lightweight CNN+Transformer model, inputting real-time feature parameters, and outputting load types;

[0026] S122, training the load demand prediction model: based on the LSTM time series prediction algorithm, combining user behavior habits and time period features, predicting load peaks and resource demand 15-30 minutes in advance;

[0027] S123, building the load priority model: based on the AHP hierarchical analysis method, setting priority weights.

[0028] As a preferred scheme of the multi-office load-oriented desktop cloud AI dynamic resource scheduling method described in the application, wherein the specific steps of S2 are as follows:

[0029] S21, real-time feature collection: through the desktop cloud client agent, multi-dimensional feature parameters of the terminal are collected every second, and user information and current running application are synchronized;

[0030] S22, real-time identification of load type: the collected feature parameters are input into a load type classification model, and a current load type is output;

[0031] S23, load demand prediction: the load type, user behavior and time period information are input into a load demand prediction model, and a resource demand prediction value in the next 15-30 minutes is output.

[0032] As a preferred scheme of the multi-office load-oriented desktop cloud AI dynamic resource scheduling method, the specific steps of S3 are as follows:

[0033] S31, resource pool state sensing: the total resource pool utilization rate of the desktop cloud cluster is monitored in real time, and idle resources are marked;

[0034] S32, priority adaptation: according to a load priority model, the current concurrent multi-load is prioritized;

[0035] S33, resource allocation algorithm: first, the predicted high-priority load peak value is reserved from the resource pool in advance by 5 minutes; then, if the actual load of a desktop exceeds the prediction value, the idle resources of the low-priority load are preferentially allocated; and then, a resource utilization rate safety threshold is set to avoid excessive resource occupation by a single load.

[0036] As a preferred scheme of the multi-office load-oriented desktop cloud AI dynamic resource scheduling method, the specific steps of S4 are as follows:

[0037] S41, resource allocation execution: instructions are sent to the resource management module to dynamically adjust the resource quota of the desktop terminal through KVM virtualization technology;

[0038] S42, real-time monitoring and adjustment: the resource usage state is detected every second, and if the load decreases, the excess resources are recovered within 30 seconds and released to the resource pool;

[0039] S43, exception backup: if the AI model prediction deviation leads to resource shortage, an emergency resource pool is started to ensure that the key load is not interrupted.

[0040] As a preferred scheme of the multi-office load-oriented desktop cloud AI dynamic resource scheduling method, the specific steps of S5 are as follows:

[0041] S51, multi-dimensional effect index collection: after resource execution, load performance indicators, resource matching degree and energy consumption data are continuously collected;

[0042] S52, scheduling effect quantitative evaluation: first, build a scheduling efficiency score model, then identify abnormal scheduling;

[0043] S53, AI model iterative optimization: first, offline iteration, then online fine-tuning, and then energy consumption performance balance optimization.

[0044] As a preferred scheme of the multi-office load-oriented desktop cloud AI dynamic resource scheduling method, wherein: the specific steps of S52 are as follows:

[0045] S521, build a scheduling efficiency score model: adopt a weighted scoring method, set index weights, and output a comprehensive score of single scheduling;

[0046] S522, abnormal scheduling identification: when the score is less than 60, it is automatically marked as low-efficiency scheduling.

[0047] As a preferred scheme of the multi-office load-oriented desktop cloud AI dynamic resource scheduling method, wherein: the specific steps of S53 are as follows:

[0048] S531, offline iteration: at dawn every day, through batch processing of feedback data in the past 24 hours, retrain the load demand prediction model, and focus on optimizing the prediction parameters of low-efficiency scheduling scenarios;

[0049] S532, online fine-tuning: when detecting that the same type of load scheduling score is less than 70 for three consecutive times, trigger online fine-tuning of the model;

[0050] S533, energy consumption performance balance optimization: based on reinforcement learning algorithm, with scheduling score ≥80 and energy consumption reduction of 10% as the target, dynamically adjust the resource allocation strategy.

[0051] Compared with the prior art:

[0052] 1. Through the dynamic resource allocation strategy of pre-judgment, reservation, dynamic filling and redundant resource recycling, combined with real-time state perception of desktop cloud cluster resource pool, it has the advantages of realizing flexible adjustment of resources with office load periodical and sudden fluctuations, avoiding resource idling or shortage caused by static quota;

[0053] 2. By building an office load feature library covering multiple features, and matching a lightweight CNN+Transformer load type classification model, it has the advantages of accurately identifying different office load types, and then matching differentiated resource demands as needed, reducing resource mismatch caused by single index judgment;

[0054] 3. Through the load demand prediction model based on the LSTM time series prediction algorithm, combined with user behavior habits and period feature analysis, the advantages of realizing early prediction of office load peak and resource demand, promoting resource adjustment from load triggering to pre-judgment, and avoiding resource adjustment lag are realized;

[0055] 4. Through the load priority model constructed based on the AHP hierarchical analysis method, the concurrent office load is prioritized and resource allocation is adapted, which has the advantages of realizing key office load to obtain the required resources, and ensuring that the key business is not affected by the resource occupation of non-key load. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical scheme and advantages of the present application clearer, the embodiments of the present application will be described in further detail below.

[0057] The present application provides a desktop cloud AI dynamic resource scheduling method for multiple office loads, which includes the following specific steps:

[0058] S1, constructing an office load feature library and AI model training (offline stage): through offline collection of multi-dimensional features of enterprise multiple office loads, a standardized office load feature library is constructed; at the same time, model training is carried out for load identification, demand prediction and priority;

[0059] The specific steps of S1 are as follows:

[0060] S11, load feature collection: collecting enterprise historical office data;

[0061] The enterprise historical office data includes load type label, multi-dimensional feature parameter and resource demand label;

[0062] The load type label includes: document editing (T1), video conference (T2), data calculation (T3), 3D rendering (T4), file backup (T5) and other 12 common office loads;

[0063] The multi-dimensional feature parameters include: CPU / memory / GPU usage, network bandwidth / delay, disk IOPS, load duration, user department / position, time period, etc. 20 features;

[0064] The resource demand label includes: the optimal resource configuration of each type of load in different scenarios (such as T2 load needs 3-core CPU, 2GB memory and 4Mbps bandwidth when 10 people meet).

[0065] S12, AI model training: first, train the load type classification model and load demand prediction model, then build the load priority model;

[0066] The specific steps of S12 are as follows:

[0067] S121, training a load type classification model: a lightweight CNN+Transformer model is used, input real-time feature parameters, and output load type (accuracy ≥98%);

[0068] S122, training a load demand prediction model: based on LSTM time series prediction algorithm, combined with user behavior habits (such as rendering model every Friday at 15:00) and time period characteristics, 15-30 minutes in advance to predict load peak and resource demand (error ≤10%);

[0069] S123, building a load priority model: based on AHP hierarchical analysis method, setting priority weight (such as CEO meeting T2 load priority = 0.9, ordinary file backup T5 load priority = 0.3).

[0070] S2, real-time load sensing and prediction (AI sensing layer): through the desktop cloud client Agent, real-time collection of terminal multi-dimensional running data, combined with the trained AI model, first real-time identification of current office load type, accurate matching of load differentiated resource demand, then based on user behavior habits and time period characteristics, 15-30 minutes in advance to predict load peak and resource demand;

[0071] The specific steps of S2 are as follows:

[0072] S21, real-time feature collection: through the desktop cloud client Agent, multi-dimensional feature parameters (CPU / memory / GPU / network / disk status) of the terminal are collected every second, and user information (department, position) and current running application (such as Zoom, Excel, Blender) are synchronized;

[0073] S22, real-time identification of load type: input the collected feature parameters into the load type classification model, and output the current load type (such as T2-video conference);

[0074] S23, load demand prediction: input the load type, user behavior, and time period information into the load demand prediction model, and output the resource demand prediction value in the next 15-30 minutes (such as "T2 load will reach peak in 5 minutes, requiring 4 core CPU, 3GB memory, 5Mbps bandwidth").

[0075] S3, dynamic resource scheduling decision (intelligent decision layer): based on the output load information, combined with the real-time state of the desktop cloud cluster resource pool, a multi-objective optimization decision mechanism is constructed, first through load priority sorting to determine resource allocation weight, then using the combination strategy of prediction reservation, dynamic filling and threshold protection, to realize the accurate matching of resources and load;

[0076] The specific steps of S3 are as follows:

[0077] S31, Resource Pool State Awareness: Real-time monitoring of the total resource pool (CPU / memory / GPU / storage / network) usage of the desktop cloud cluster, marking idle resources (such as unallocated CPU cores, idle memory blocks);

[0078] S32, Priority Adaptation: According to the load priority model, the current concurrent multi-load is prioritized (such as T2>T3>T1>T5);

[0079] S33, Resource Allocation Algorithm: First, reserve the corresponding resources from the resource pool 5 minutes in advance for the predicted high-priority load peak (such as T2 meeting); then, if the actual load of a desktop exceeds the predicted value (such as T3 data computing memory requirement increases from 4GB to 6GB), preferentially allocate from the idle resources of low-priority loads (such as reduce the memory allocation of T5 backup); then set the resource usage safety threshold (such as CPU≤85%, memory≤80%) to avoid excessive resource occupation by a single load.

[0080] S4, Resource Execution and Recovery (Resource Execution Layer): Convert the allocation instructions of S3 into actual resource adjustment actions, on the one hand, dynamically scale up / down resources for target terminals in real time, and on the other hand, through real-time monitoring of load changes, quickly recover redundant resources to the public resource pool when the load ends or demand decreases, and configure an emergency resource pool to cope with prediction deviations;

[0081] The specific steps of S4 are as follows:

[0082] S41, Resource Allocation Execution: Send instructions to the resource management module to dynamically adjust the resource quota of the desktop terminal through KVM virtualization technology (such as increasing the CPU of T2 meeting desktop from 2 cores to 4 cores);

[0083] S42, Real-time Monitoring and Adjustment: Detect resource usage status every second, if the load decreases (such as T2 meeting ends), recover the excess resources (such as reduce CPU to 2 cores) within 30 seconds and release them to the resource pool;

[0084] S43, Exception Backup: If the AI model prediction deviation leads to resource shortage, start the emergency resource pool (reserve 10% of the total resources) to ensure that the key load is not interrupted.

[0085] S5, Dispatching Effect Evaluation and AI Model Iteration: Collect the load performance indicators, resource matching degree, and energy consumption data after dispatching, establish a dispatching efficiency scoring model, and based on the scoring results, offline batch iteration and online real-time fine-tuning of the AI model, while combining reinforcement learning algorithm to optimize energy consumption performance balance strategy;

[0086] The specific steps of S5 are as follows:

[0087] S51, multi-dimensional effect indicator collection: after resource execution, continuously collect load performance indicators, resource matching degree, and energy consumption data; the load performance indicators include: video conference stall rate, data calculation completion time, 3D rendering frame rate, and other 15 business experience indicators; the resource matching degree includes: deviation rate of predicted resource demand and actual demand (such as the deviation of the predicted value of 4 cores of CPU and the actual peak value of 4.2 cores); the energy consumption data includes: cluster PUE value, power consumption fluctuation during resource pool dynamic adjustment (such as instantaneous power consumption when GPU frequency is raised and lowered);

[0088] S52, scheduling effect quantitative evaluation: first, a scheduling efficiency score model is constructed, and then abnormal scheduling is identified;

[0089] The specific steps of S52 are as follows:

[0090] S521, constructing a scheduling efficiency score model: using a weighted scoring method, setting index weights (such as business experience accounting for 40%, resource matching degree accounting for 35%, and energy consumption accounting for 25%), and outputting a comprehensive score (0-100 points) of a single scheduling;

[0091] S522, abnormal scheduling identification: when the score is lower than 60 points, it is automatically marked as low-efficiency scheduling (such as a T4 rendering caused by insufficient GPU allocation leading to a 50% increase in rendering time beyond expectation).

[0092] S53, AI model iterative optimization: first, offline iteration is performed, then online fine-tuning is performed, and then energy consumption performance balance optimization is performed;

[0093] The specific steps of S53 are as follows:

[0094] S531, offline iteration: at dawn every day, feedback data of the previous 24 hours is processed in batches, the load demand prediction model is retrained, and the prediction parameters of low-efficiency scheduling scenarios are optimized (such as the GPU memory prediction error of a certain department 3D rendering being reduced from 15% to 8%);

[0095] S532, online fine-tuning: when it is detected that the same type of load scheduling score is lower than 70 points for three consecutive times, online fine-tuning of the model is triggered (such as the network bandwidth prediction value of a T2 conference being continuously low, and the network weight coefficient of the prediction model is automatically increased by 12%);

[0096] S533, energy consumption performance balance optimization: based on a reinforcement learning algorithm, with scheduling score ≥ 80 points and energy consumption reduction of 10% as the target, dynamically adjusting resource allocation strategies (such as reducing CPU frequency by 20% during non-peak periods to save power consumption while ensuring business experience).

[0097] Although the present application has been described with reference to the embodiments above, various changes and modifications can be suggested to one skilled in the art, and it is intended that the present application encompass such changes and modifications as fall within the scope of the appended claims. Particularly, each feature disclosed in the description and / or the claims can be used in the combination with each of the features disclosed in the description and / or the claims, unless specifically stated otherwise. Therefore, the present application is not intended to be limited to the particular embodiments disclosed in the description and / or the claims.

Claims

1. A desktop cloud AI dynamic resource scheduling method for multiple office workloads, characterized in that, The specific steps are as follows: S1, Building an Office Load Feature Library and AI Model Training: By collecting multi-dimensional features of multiple office loads from enterprises offline, a standardized office load feature library is built; at the same time, models are trained for load identification, demand prediction, and prioritization. S2, Real-time Load Awareness and Prediction: The desktop cloud client Agent collects multi-dimensional operating data of the terminal in real time and combines it with the trained AI model to first identify the current office load type in real time and accurately match the differentiated resource demands of the load. Then, based on user behavior habits and time period characteristics, it predicts the load peak and resource demand 15-30 minutes in advance. S3, Dynamic Resource Scheduling Decision: Based on the output load information and combined with the real-time status of the desktop cloud cluster resource pool, a multi-objective optimization decision mechanism is constructed. First, the resource allocation weight is determined by load priority sorting. Then, a combination of predictive reservation, dynamic supplementation and threshold protection strategies are adopted to achieve accurate matching of resources and load. S4, Resource Execution and Reclamation: Transforms the allocation instructions of S3 into actual resource adjustment actions. On the one hand, it dynamically expands / shrinks resources for the target terminal in real time. On the other hand, by monitoring load changes in real time, it quickly reclaims redundant resources to the public resource pool when the load ends or demand decreases. At the same time, it configures an emergency resource pool to deal with prediction deviations. S5, Scheduling effect evaluation and AI model iteration: By collecting load performance indicators, resource matching degree and energy consumption data after scheduling, a scheduling efficiency scoring model is established. Based on the scoring results, the AI ​​model is iterated offline in batches and fine-tuned online in real time. At the same time, the energy consumption performance balancing strategy is optimized by combining reinforcement learning algorithms.

2. The desktop cloud AI dynamic resource scheduling method for multiple office workloads according to claim 1, characterized in that, The specific steps of S1 are as follows: S11, Load Characteristics Collection: Collect historical office data of the enterprise; S12, AI Model Training: First, train the load type classification model and the load demand prediction model, and then build the load priority model.

3. The desktop cloud AI dynamic resource scheduling method for multiple office workloads according to claim 2, characterized in that, The enterprise's historical office data includes load type tags, multidimensional feature parameters, and resource requirement tags.

4. The desktop cloud AI dynamic resource scheduling method for multiple office workloads according to claim 2, characterized in that, The specific steps of S12 are as follows: S121, Training the load type classification model: Using a lightweight CNN+Transformer model, inputting real-time feature parameters, and outputting load type; S122, Training the load demand prediction model: Based on the LSTM time series prediction algorithm, combined with user behavior habits and time period characteristics, predict the load peak and resource demand 15-30 minutes in advance. S123, Construct a load priority model: Based on the AHP (Analytical Hierarchy Process), set priority weights.

5. The desktop cloud AI dynamic resource scheduling method for multiple office workloads according to claim 1, characterized in that, The specific steps of S2 are as follows: S21, Real-time Feature Acquisition: Through the desktop cloud client Agent, multi-dimensional feature parameters of the terminal are collected every second, and user information is synchronized with the currently running application; S22, Real-time load type identification: Input the collected feature parameters into the load type classification model and output the current load type; S23, Load Demand Forecast: Input load type, user behavior, and time period information into the load demand forecast model, and output the resource demand forecast value for the next 15-30 minutes.

6. The desktop cloud AI dynamic resource scheduling method for multiple office workloads according to claim 1, characterized in that, The specific steps of S3 are as follows: S31, Resource Pool Status Awareness: Real-time monitoring of the total resource pool utilization of the desktop cloud cluster and marking idle resources; S32, Priority Adaptation: Based on the load priority model, prioritize the multiple concurrent loads. S33, Resource Allocation Algorithm: First, for the predicted high-priority load peak, reserve corresponding resources from the resource pool 5 minutes in advance; then, if the actual load of a desktop exceeds the predicted value, prioritize allocating resources from the idle resources of low-priority loads; then set a resource utilization safety threshold to avoid excessive resource contention by a single load.

7. The desktop cloud AI dynamic resource scheduling method for multiple office workloads according to claim 1, characterized in that, The specific steps of S4 are as follows: S41, Resource Allocation Execution: Sends instructions to the resource management module to dynamically adjust the resource quota of the desktop terminal through KVM virtualization technology; S42, Real-time monitoring and adjustment: Detects resource usage status every second. If the load decreases, it will reclaim excess resources within 30 seconds and release them to the resource pool. S43, Abnormal Backup: If AI model prediction errors lead to insufficient resources, activate the emergency resource pool to ensure that critical loads are not interrupted.

8. The desktop cloud AI dynamic resource scheduling method for multiple office workloads according to claim 1, characterized in that, The specific steps of S5 are as follows: S51, Multi-dimensional performance metrics collection: After resource execution, continuously collect load performance metrics, resource matching degree, and energy consumption data; S52, Quantitative evaluation of scheduling effectiveness: First, construct a scheduling performance scoring model, and then identify abnormal scheduling. S53, AI model iterative optimization: first, offline iteration is performed, then online fine-tuning is performed, and finally, energy consumption and performance balance optimization is performed.

9. The desktop cloud AI dynamic resource scheduling method for multiple office workloads according to claim 8, characterized in that, The specific steps of S52 are as follows: S521, Construct a scheduling performance scoring model: Use a weighted scoring method, set the weights of the indicators, and output a comprehensive score for a single scheduling operation; S522, Abnormal Scheduling Identification: When the score is below 60, it is automatically marked as inefficient scheduling.

10. The desktop cloud AI dynamic resource scheduling method for multiple office workloads according to claim 8, characterized in that, The specific steps of S53 are as follows: S531, Offline Iteration: Every day at midnight, the load demand prediction model is retrained by batch processing the feedback data from the previous 24 hours, with a focus on optimizing the prediction parameters for inefficient scheduling scenarios. S532, Online Fine-tuning: When the same type of load scheduling score is detected to be below 70 points for 3 consecutive times, online fine-tuning of the model is triggered; S533, Energy Consumption Performance Balance Optimization: Based on reinforcement learning algorithm, with the goal of scheduling score ≥80 points and energy consumption reduced by 10%, the resource allocation strategy is dynamically adjusted.