An AI-based MaaS platform intelligent scheduling settlement method and system
By employing an AI-based intelligent scheduling and billing method for the MaaS platform, and utilizing large language models and matching degree evaluation models, the problem of the MaaS platform's high barrier to entry for non-professional users and low efficiency in computing resource scheduling has been solved, achieving efficient and transparent computing resource management and billing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU SANLIANGSAN INFORMATION TECHNOLOGY CO LTD
- Filing Date
- 2026-04-21
- Publication Date
- 2026-05-29
Smart Images
Figure CN122115043A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of information technology, specifically to an AI-based intelligent scheduling and settlement method and system for a MaaS platform. Background Technology
[0002] With the rapid development of technologies such as artificial intelligence, big data, and high-performance computing, computing power has become a core resource supporting the development of the digital economy. However, problems such as uneven distribution of computing resources, supply-demand mismatch, low scheduling efficiency, and opaque billing mechanisms severely restrict the inclusiveness and efficiency of computing power services. To address this challenge, Model as a Service (MaaS) platforms have emerged, aiming to provide users with on-demand, elastic, and intelligent computing power delivery capabilities by uniformly scheduling multi-source heterogeneous computing resources.
[0003] Existing MaaS platforms typically rely on users to explicitly specify constraints such as computing power, execution time limits, and budgets through structured forms or API parameters, posing a high barrier to entry for non-professional users. Therefore, there is an urgent need for an intelligent scheduling method that supports natural language processing and dynamic billing to improve the usability and scheduling accuracy of MaaS platforms. Summary of the Invention
[0004] This specification describes an AI-based intelligent scheduling and settlement method and system for a MaaS platform through several embodiments.
[0005] Firstly, the embodiments of this specification provide an AI-based intelligent scheduling and settlement method for a MaaS platform, comprising the following steps:
[0006] Obtain a description of the user's computing power requirements in natural language;
[0007] Extract the semantics of the computing power requirement description to obtain task feature parameters, which include task type, computing power scale requirement, latency sensitivity, budget constraint, and execution time window;
[0008] Obtain channel description information for multiple idle computing power provision channels connected to the platform. The channel description information includes at least computing power scale, current load status, billing method, unit price, network latency, and compatible task type.
[0009] Based on the task feature parameters and the channel description information of the computing power provision channels, the matching degree between each computing power provision channel and the channel description of the computing power demand is calculated using a pre-trained matching degree evaluation model.
[0010] A computing power provision strategy is generated based on the channel matching degree, and the computing power provision strategy includes channel priority ranking for each task type;
[0011] Based on the computing power provision strategy and the actual idle status of the computing power provision channel, the user's computing power tasks are scheduled to be executed on the target computing power provision channel, and billing is completed based on the billing method of the computing power provision channel and the actual resource usage.
[0012] Secondly, embodiments of this specification provide an AI-based MaaS platform intelligent scheduling and settlement system, including:
[0013] The input module obtains the user's computing power requirements described in natural language;
[0014] The extraction module extracts the semantics of the computing power requirement description to obtain task feature parameters, which include task type, computing power scale requirement, latency sensitivity, budget constraint, and execution time window.
[0015] The access module obtains channel description information of multiple idle computing power provision channels connected to the platform. The channel description information includes at least computing power scale, current load status, billing method, unit price, network latency, and adapted task type.
[0016] The matching module calculates the matching degree between each computing power supply channel and the channel description information of the computing power supply channel based on the task feature parameters and the channel description information of the computing power supply channel using a pre-trained matching degree evaluation model.
[0017] The strategy module generates a computing power provision strategy based on the channel matching degree. The computing power provision strategy includes the channel priority ranking corresponding to each task type.
[0018] The scheduling module schedules the user's computing tasks to the target computing power supply channel for execution based on the computing power supply strategy and the actual idle status of the computing power supply channel, and completes the billing based on the billing method of the computing power supply channel and the actual resource usage.
[0019] Thirdly, embodiments of this specification provide an electronic device, including a processor and a memory;
[0020] The processor is connected to the memory;
[0021] The memory is used to store executable program code;
[0022] The processor runs a program corresponding to the executable program code stored in the memory to perform the method described in any of the above aspects.
[0023] Fourthly, embodiments of this specification provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the methods described in any of the above aspects.
[0024] Fifthly, embodiments of this specification provide a computer program product, including a computer program that, when executed by a processor, implements the methods described in any of the above aspects.
[0025] The beneficial effects of the technical solutions provided in some embodiments of this specification include at least the following:
[0026] In several embodiments of this specification, the AI-based MaaS platform intelligent scheduling and settlement method and system provide semantic understanding and structured extraction of users' natural language descriptions of computing power needs by introducing a large language model, thereby lowering the barrier to entry and improving the platform's usability. Utilizing a pre-trained matching degree evaluation model, and integrating task feature parameters with multi-dimensional channel description information, personalized intelligent matching of computing power channels is achieved, improving the rationality of resource scheduling and the success rate of task execution. The computing power provision strategy dynamically integrates channel matching degree, real-time idle status, budget constraints, and latency requirements, supporting priority ranking, task splitting, and fault rescheduling, enhancing the system's robustness and resilience, and helping to ensure service quality.
[0027] Other features and advantages of various embodiments of this specification will be further revealed in the following detailed description and accompanying drawings. Attached Figure Description
[0028] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0029] Figure 1 This is a schematic diagram of the intelligent scheduling and settlement of the MaaS platform provided in this manual.
[0030] Figure 2 This is a schematic diagram of the intelligent scheduling and settlement method of the MaaS platform provided in this manual.
[0031] Figure 3 This is a schematic diagram of the method for obtaining task characteristic parameters provided in this specification.
[0032] Figure 4 This is a schematic diagram of the training matching evaluation model method provided in this manual.
[0033] Figure 5 This document provides a schematic diagram of the strategy and method flow for generating computing power as described in this specification.
[0034] Figure 6This document provides a flowchart illustrating the method for scheduling data to the target computing power.
[0035] Figure 7 This is a schematic diagram of the billing process provided in this manual.
[0036] Figure 8 This is a schematic diagram of the intelligent scheduling and settlement system of the MaaS platform provided in this manual.
[0037] Figure 9 This is a schematic diagram of the electronic device provided in this manual. Detailed Implementation
[0038] The technical solutions of the embodiments of this specification will be explained and described below with reference to the accompanying drawings. However, the following embodiments are only preferred embodiments of this specification and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of this specification.
[0039] The terms "first," "second," "third," etc., in the description, claims, and accompanying drawings are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.
[0040] In the following description, terms such as “inner,” “outer,” “upper,” “lower,” “left,” and “right” are used only to facilitate the description of the embodiments and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this specification.
[0041] All data involved in this application are information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions.
[0042] Before introducing the technical solutions described in this manual, the application scenarios and related technologies of the technical solutions will be introduced.
[0043] MaaS (Model as a Service) platforms are a business model and platform architecture that provides artificial intelligence models as standardized services to users via the cloud. Its core lies in encapsulating the complex technical aspects of model training, deployment, inference, and iterative maintenance. Users do not need to concern themselves with the underlying infrastructure and algorithm details; they can simply use APIs, web interfaces, or SDKs to invoke model capabilities to complete specific tasks, such as natural language processing, image recognition, and speech synthesis. The MaaS model significantly lowers the application threshold of AI technology, promoting the democratization and scaling of AI. Platforms typically provide pre-trained base models, industry-specific models, and corresponding fine-tuning tools, evaluation monitoring, and version management functions, forming a complete model lifecycle management ecosystem.
[0044] The operation of a MaaS platform heavily relies on stable and scalable computing power delivery channels, which form the physical foundation for model inference and training tasks. The key parameters of its computing resources are complex and diverse, primarily revolving around performance, specifications, and availability. Computational performance parameters are core, including the type (such as NVIDIA H100, AMD MI300X, Google TPU v5e), quantity, interconnect bandwidth (such as NVLink, InfiniBand), single-card memory capacity and bandwidth, and the CPU and system memory configuration supporting the chips. These directly determine the speed and scale of model processing (especially large language models). Storage parameters involve the IOPS and throughput of caches (such as NVMe SSDs for hot data), and the capacity and access latency of massive object storage or file storage used for model weights and training datasets. Network parameters are crucial, including internal network bandwidth and latency between instances (affecting distributed training efficiency), network egress bandwidth and stability for instances providing API services, and network security policies and isolation capabilities. In addition, resource availability and elasticity parameters, such as regional availability zone distribution, resource quotas, supply and scaling response time, as well as hardware redundancy and reliability indicators (such as SLA levels), are also key considerations when selecting a platform. Regarding billing methods, the computing power channels relied upon by MaaS platforms exhibit high flexibility and diversity. Pay-as-you-go billing is the most basic and flexible model, where users pay based on the actual computing power resources consumed, typically measured in "inference tokens," "GPU runtimes," or "data volume processed," suitable for tasks with large traffic fluctuations or temporary tasks.
[0045] When choosing a MaaS service, users need to spend a significant amount of time considering the computing power delivery channels. To improve the ease of use of MaaS, this manual provides an AI-based intelligent scheduling and settlement method and system for MaaS platforms. Please refer to the appendix. Figure 1The system obtains an unstructured description of the computing power requirements 10 described by the user in natural language. Then, it extracts the semantics of the computing power requirements description 10 and transforms it into structured task feature parameters 11. These parameters specifically include task type (e.g., model training or inference), computing power scale requirements (e.g., the number and model of GPUs required), latency sensitivity (requirements for task completion delay), budget constraints (the user's acceptable cost range), and execution time window (the time period during which the task needs to run).
[0046] While acquiring user needs, the platform obtains detailed descriptions of multiple available computing power provision channels. This information includes at least computing power scale, current load status, billing method (e.g., pay-as-you-go or reserved instances), unit price, network latency, and the task types suited to each channel. Based on the extracted task feature parameters 11 and the descriptions of all available computing power provision channels, a pre-trained matching degree evaluation model is used to calculate and output a quantified channel matching degree for each computing power provision channel, tailored to the current user needs. A specific computing power provision strategy 31 is generated based on the calculated channel matching degree. This strategy generates a corresponding, priority-ordered list of available channels for each possible task type in the user's task. According to the generated computing power provision strategy 31 and considering the real-time idle status feedback from each computing power provision channel, the user's computing power task is intelligently scheduled to the optimal target computing power provision channel for execution. Upon task completion, accurate billing and settlement are automatically completed based on the billing method of the target channel and the actual resource consumption of the task.
[0047] Specifically, this specification first provides an AI-based intelligent scheduling and settlement method for a MaaS platform. Please refer to the appendix. Figure 2 This includes the following steps:
[0048] Step S1) Obtain the user's computing power requirement description 10 in natural language.
[0049] In practical applications, users often lack professional technical backgrounds and find it difficult to accurately fill out structured resource request forms. To address this, the front-end integrated platform interface allows users to input their computing power needs in a manner similar to everyday conversation. For example, when a research team initiates an AI model training task using the platform, they can directly enter in the text box: "We need to complete the training of a ResNet-50 image classification model within 48 hours. The dataset is approximately 100GB. We hope to keep the cost below 500 yuan and are not sensitive to latency." This statement constitutes a complete "Computing Power Requirement Description 10." The platform plays a crucial bridging role in this process. By shielding the complexity of the underlying infrastructure and using natural language input as the sole entry point, it lowers the user threshold and provides raw textual evidence for subsequent AI semantic analysis. Such descriptions can include explicit parameters (e.g., a budget of 500 yuan and a time window of 48 hours) or implicit task types (e.g., "training ResNet-50" corresponds to a deep learning training task) or resource scale (e.g., "100GB dataset" infers the required storage and GPU memory).
[0050] Step S2) Extract the semantics of the computing power requirement description 10 to obtain task feature parameters 11. The task feature parameters 11 include task type, computing power scale requirement, latency sensitivity, budget constraint and execution time window.
[0051] The natural language description input by the user is first sent to the semantic parsing module. This module relies on the integrated Large Language Model (LLM) to perform deep understanding of the text and reconstruct the user's true intent in a structured manner. Please refer to the appendix. Figure 3 The method for extracting the semantics of the computing power requirement description 10 and obtaining the task feature parameters 11 includes:
[0052] Step S21) Perform natural language processing on the computing power requirement description 10, and use the accessed large language model to identify key semantic elements. For example, when a user inputs "We plan to run a BERT fine-tuning task next week. The data volume is not large, and we hope to complete it as soon as possible, with a budget of no more than 800 yuan," the large language model will extract key semantic fragments such as "BERT fine-tuning" (task type), "not large data volume" (implying a small computing power scale), "complete as soon as possible" (high latency sensitivity), "budget no more than 800 yuan" (explicit budget constraint), and "next week" (execution time window). Although these elements are presented in an unstructured form, they already have the basis for being transformed into standard parameters.
[0053] Step S22) Maps the identified key semantic elements to the dimensions of predefined task feature parameters 11, which include task type, computing power requirements, latency sensitivity, budget constraints, and execution time window. The platform maintains a standardized template for task feature parameters 11. In the example above, "BERT fine-tuning" will be categorized as a "Natural Language Processing - Model Fine-tuning" task; "small data volume" combined with historical similar task data can be initially mapped to "low to medium GPU memory requirements (e.g., ≤16GB)"; "complete as soon as possible" corresponds to "high latency sensitivity"; "800 yuan" is directly entered into the budget constraint field; and "next week" is converted to a specific time range, such as "2026-02-10 00:00 to 2026-02-16 23:59".
[0054] Step S23) For parameters not explicitly mentioned, prompt the large language model to perform reasoning and completion based on contextual semantics and historical user behavior data, and assign confidence weights.
[0055] Not all parameters are explicitly stated in the user description. For example, a user might only input: "Train Stable Diffusion using 3 A100s for three days." Budget and latency sensitivity are not mentioned. In this case, the platform will guide the large language model to combine the context (e.g., "three days" implies a tolerable level of latency) and the user's past task records (e.g., historical tasks were mostly batch processing, with very few preemptive instances) to infer "latency sensitivity: low," and estimate the average market price under the current configuration to be approximately 1200 yuan, thus completing the "budget constraint: approximately 1200 yuan." Each inference result is accompanied by a confidence weight (e.g., 0.85) for risk assessment and priority adjustment in subsequent scheduling decisions.
[0056] Step S24) Standardize all extracted and inferred parameters to form a structured task feature parameter vector.
[0057] To support efficient matching and cost optimization of the subsequent scheduling engine, the extracted and inferred parameters need to be converted into a unified, computable numerical or enumerated form. First, normalization is performed based on a predefined parameter ontology. Task types are mapped to a predefined classification system; for example, "training Stable Diffusion" and "running a diffusion model" are uniformly classified as "Generative AI: Image Generation". Computing power requirements are converted into standardized computing units based on GPU model, quantity, and memory estimates (e.g., "3×A100@80GB": computing power score = 9.2 / 10). Latency sensitivity is divided into three levels—high (H), medium (M), and low (L)—based on semantic intensity and further encoded as numerical values (e.g., H=1.0, M=0.6, L=0.2). Budget constraints are uniformly expressed in Chinese Yuan (CNY), retaining the original values and confidence intervals (e.g., "approximately 1200 yuan ± 10%"). The execution time window is converted to a time period in ISO 8601 format (e.g., "next week": [2026-02-10T00:00:00, 2026-02-16T23:59:59]).
[0058] For example, a user enters on the platform, "I want to run a YOLOv5 object detection training program using two V100 processors, preferably within two days, and without spending too much money." After processing in the preceding steps, the initial results are: "Task type = YOLOv5 training," "Computing power = 2 × V100," "Time window ≈ 48 hours," "Budget is ambiguous," and "Latency is moderate." This information is then standardized as follows:
[0059] {
[0060] "task_type": "computer_vision-object_detection",
[0061] "compute_scale": 6.5, / / Normalized based on V100 performance benchmark
[0062] "latency_sensitivity": 0.6, / / "Preferably within two days" → Moderate sensitivity
[0063] "budget_cny": 600, / / Inference based on the average of similar historical tasks, confidence level 0.78
[0064] "time_window": ["2026-02-06T16:00:00", "2026-02-08T16:00:00"]
[0065] }
[0066] The platform successfully condenses diverse user expressions into consistent and operable scheduling inputs, providing data support for achieving accurate and efficient MaaS intelligent settlement.
[0067] Step S3) Obtain channel description information 20 of multiple idle computing power provision channels accessed by the platform. The channel description information 20 includes at least computing power scale, current load status, billing method, unit price, network latency and adapted task type.
[0068] To achieve efficient and economical computing power scheduling, it is necessary to monitor the available resources of each computing power provider (such as public cloud vendors, edge computing nodes, university supercomputing centers, or private clusters) in real time. These providers register and connect to the platform in the form of "computing power channels," and their capabilities and status are continuously reported through standardized interfaces, forming a dynamically updated pool of channel description information. Each channel description information pool contains the following dimensions:
[0069] Computing power scale: refers to the total amount of computing resources that can be allocated at present, usually expressed as GPU model and number, CPU core count, memory capacity, etc. For example, an edge node reports "4×NVIDIA T4+64GB RAM", while a public cloud region provides "elastic A100 cluster, which can be expanded to a maximum of 20 cards".
[0070] Current load status: Reflects resource utilization and is used to determine whether it is truly "idle". For example, the current GPU utilization rate of a university supercomputing center is 35%, which is marked as "light load"; while a cloud region has reached 85% load due to a sudden traffic surge and is marked as "high load, only accepting high priority tasks".
[0071] Billing methods include per-second billing, hourly billing, spot instances, and reserved instances. For example, Channel A supports "pay-per-second", Channel B offers "50% off during off-peak hours", and Channel C only offers "reserved discounts for training tasks of 7 days or more".
[0072] Unit price: The current price is indicated in CNY / hour or CNY / GPU·hour. For example, "A100: ¥12.5 / GPU·h" and "T4: ¥3.2 / GPU·h".
[0073] Network latency: refers to the average round-trip time (RTT) from the user's location to the computing node, which is crucial for distributed training or real-time inference. For example, the latency for a user in East China to access a local cloud node is 8ms, while the latency to access a North American node is as high as 180ms.
[0074] Adapted Task Types: Based on hardware characteristics and software environment, declare the supported task categories. For example, one channel pre-installs PyTorch+NCCL and explicitly supports "distributed training"; another channel only configures TensorRT runtime and is only suitable for "inference tasks".
[0075] Step S4) Based on the task feature parameters 11 and the channel description information 20 of the computing power provision channels, calculate the channel matching degree between each computing power provision channel and the computing power demand description 10 using a pre-trained matching degree evaluation model.
[0076] Once the structured task feature parameter vector and the standardized description information of each computing power channel are ready, the platform calls a lightweight but high-precision pre-trained matching degree evaluation model, usually a multilayer perceptron or a fine-tuned Transformer encoder, to perform end-to-end compatibility and cost-effectiveness scoring for each channel and output the channel matching degree between 0 and 1.
[0077] Please refer to the appendix for details. Figure 4 Methods for training matching evaluation models include:
[0078] Step S41) Read the computing power scheduling record. The record includes the user's original natural language requirement description, the corresponding extracted task feature parameters 11, the channel description information 20 of the selected computing power provision channel, and the actual execution performance indicators. Read the historical computing power scheduling records from the platform's backend database. These records are the data foundation for model training. Each record contains the user's initial original natural language requirement description (e.g., "I need to use 4 A100 graphics cards to fine-tune a 7 billion parameter LLaMA model between 8 pm and 12 am tonight, with a budget of no more than 500 yuan"), the structured task feature parameters 11 automatically extracted by the platform based on the description, the complete description information of the computing power provision channel that the user ultimately selected (or the system automatically allocated) (including its computing power specifications, billing unit price, network latency at the time, etc.), and the actual performance indicators generated after the task was completed (e.g., whether it was successful, actual time consumed, resource utilization, final cost, etc.).
[0079] Step S42) Construct a training sample set based on the historical records. Each sample contains a task feature parameter vector 11, a channel description information vector 20, and a corresponding matching degree label. The matching degree label is a comprehensive adaptation score of the actual execution of the computing power providing channel. The comprehensive adaptation score is calculated by weighting the execution success rate, resource utilization rate, cost-effectiveness, and latency compliance rate.
[0080] Based on this rich historical data, a training sample set for supervised learning can be constructed. Specifically, each training sample consists of three parts: the first part is a task feature parameter vector (11 vectors), which is a structured numerical expression of user needs; the second part is a channel description information vector (20 vectors), which is a digital description of the characteristics of a specific computing power channel; and the third part is the matching degree label corresponding to the sample, a key true value. This label is not manually labeled, but is a "comprehensive fit score" calculated through an evaluation formula. The score comprehensively considers the actual effect of this matching, such as execution success rate (whether the task is successfully completed), resource utilization rate (whether the allocated computing power is used efficiently), cost-effectiveness (whether the actual expenditure is within budget and cost-effective), and latency compliance rate (whether it is completed within the expected time), and other indicators are calculated by weighting them through a pre-set weight. For example, a task that costs 450 yuan, is successfully completed within the specified time, and has a GPU utilization rate as high as 90% will have a high comprehensive fit score for "computing power channel A" under this requirement.
[0081] Step S43) Construct a machine learning or deep learning model as the architecture of the matching degree evaluation model. After fusing and encoding the task feature parameter vector 11 and the channel description information vector 20, input the model to predict the channel matching degree.
[0082] Construct a specific machine learning model (such as a gradient boosting tree) or deep learning model (such as a deep neural network) as the architecture for the matching degree evaluation model. In terms of model design, it is necessary to effectively fuse and encode the task feature vector representing the demand and the channel description vector representing the channel resources. For example, the two vectors can be concatenated first, and then input into a multi-layer fully connected network for processing, and finally output a predicted matching degree score.
[0083] Step S44) Optimize the model parameters by performing end-to-end training on the model by minimizing the loss function between the predicted matching degree and the label. After reaching the preset accuracy, the training is completed and the matching degree evaluation model is obtained.
[0084] The model training process is a standard end-to-end optimization process. By continuously inputting training samples into the model, the difference between the model's predicted matching degree and the "comprehensive fit score" label calculated based on real historical data (i.e., the value of the loss function) is calculated, and optimization algorithms such as backpropagation are used to minimize this difference, thereby gradually adjusting and optimizing the millions of parameters within the model. When the model's prediction accuracy on an independent validation dataset reaches a preset standard, for example, when the correlation coefficient between the predicted score and the actual score exceeds 0.9, the training is complete. At this point, an intelligent model capable of accurately assessing the matching degree between computing power requirements and channel resources is obtained, which can be used by the platform in subsequent intelligent scheduling.
[0085] Step S5) Generate a computing power provision strategy 31 based on the channel matching degree. The computing power provision strategy 31 includes channel priority sorting corresponding to each task type.
[0086] Please see the appendix Figure 5 The method for generating computing power provision strategy 31 based on the channel matching degree includes:
[0087] Step S51) Sort each computing power provision channel in descending order of its channel matching degree to form a candidate channel list. The platform sorts all connected computing power provision channels in descending order based on the previously calculated channel matching degree to form an initial candidate channel list. The channel with the highest matching degree, such as a GPU cluster that has demonstrated high cost-effectiveness and stability in similar historical tasks, will be ranked at the top of the list.
[0088] Step S52) Based on budget constraints and billing methods, perform cost feasibility verification on candidate channels and eliminate options that exceed the user's budget. The platform will introduce the user-set budget constraints as a hard screening condition. For example, if the user sets a budget limit of 200 yuan for a model training task, the platform will verify the cost calculated for each channel in the list based on the estimated usage time or computing power consumption. Suppose that a certain channel (such as Channel_A) has a high matching degree, but its estimated cost under the on-demand billing model is 250 yuan, which exceeds the budget, then the channel will be immediately removed from the candidate list. At the same time, complex billing methods (such as those including reserved discounts or bidding fluctuations) will also be included in cost simulation for accurate verification.
[0089] Step S53) Based on latency sensitivity and execution time window, remove channels that do not meet latency requirements and time availability. Further filtering is performed by combining the task's latency sensitivity with the specific execution time window. For an AI inference task requiring real-time response, the platform checks the current and predicted network latency and processing latency of each candidate channel. If the average latency of a channel (e.g., Channel_B) exceeds the task's tolerable threshold, it is excluded. Simultaneously, the platform compares the channel's resource availability commitment within the user-specified execution time window (e.g., 2023-10-01 20:00:00 to 2023-10-01 23:00:00). If the channel is already booked or has no available capacity during that period, it is also removed.
[0090] Step S54) The remaining computing power supply channels are selected as candidate channels. Based on the task type and the matching task type tags of the computing power supply channels, several candidate channels are selected for each task type. After the two rounds of filtering based on cost and constraints, the remaining channels constitute the final candidate pool after feasibility verification. The platform performs precise matching between task type and channel expertise. For example, a user task may simultaneously include both "large-scale distributed training" and "high-concurrency inference" subtypes. The platform will check the "matching task type" tags of each channel, assigning channels that excel in distributed_training (such as clusters equipped with high-speed InfiniBand networks) to training tasks and channels that excel in low-latency_inference (such as edge nodes) to inference tasks, thereby selecting the most suitable candidate channels for each specific task type.
[0091] Step S55) Generate a channel priority ranking for each task type based on the channel matching degree and idle load capacity of the candidate channels.
[0092] After identifying suitable candidate channels for each task type, the platform generates a final channel priority ranking. The dynamic idle load capacity of each channel is incorporated in real-time as a key trade-off. For example, even if Channel_C has a slightly lower match for a task than Channel_D, if Channel_C currently has more readily available computing power and can avoid queuing, the platform may prioritize Channel_C over Channel_D in the priority ranking.
[0093] Step S6) Based on the computing power provision strategy 31 and the actual idle status of the computing power provision channel, schedule the user's computing power tasks to the target computing power provision channel for execution, and complete the billing based on the billing method of the computing power provision channel and the actual resource usage. Dynamic decision-making and fault tolerance management are performed based on the generated computing power provision strategy 31 and real-time resource conditions.
[0094] Please refer to the appendix for details. Figure 6 The method for scheduling a user's computing power task to a target computing power supply channel for execution, based on the computing power supply strategy 31 and the actual idle status of the computing power supply channel, includes:
[0095] Step S611) Real-time query of the dynamic idle resource status of candidate channels during the task's scheduled execution period. The dynamic idle resource status includes the number of available computing units, bandwidth availability, and estimated queuing time. Initiate a real-time query to all candidate channels in the priority list to obtain the dynamic idle resource status during the task's scheduled execution period. Check if there are currently any available GPU instances, including detailed information such as the number of available computing units (e.g., 4 NVIDIA_A100 GPUs remaining), network bandwidth availability (e.g., 10Gbps), and the critical estimated queuing time for tasks. For example, a high-performance cluster Cluster_Alpha with a high matching degree might have queued tasks with an estimated queuing time as long as 2 hours, directly affecting the final selection of latency-sensitive tasks.
[0096] Step S612) Based on the channel priority ranking and real-time idle status of the candidate channels, select the candidate channels that meet the task characteristic parameter 11 and whose resources can be allocated immediately as the target channels. Considering the pre-calculated priority ranking of each channel and the just-acquired real-time idle status, a final scheduling decision is made. The core logic is to select channels that not only meet all the task characteristic parameters but also provide immediately or fastest-allocating resources as the target channels. For example, for an interactive data analysis task that needs to start immediately, even if Cluster_Beta is ranked second, if the first-ranked Cluster_Alpha is in a queue, and Cluster_Beta has immediately available resources, then the platform will prioritize scheduling the task to Cluster_Beta for execution.
[0097] Step S613) Sends task deployment instructions to the selected target channel and establishes a task execution monitoring session to continuously collect the actual resource usage, execution progress, and abnormal events of the computing power task. After determining the target channel, the platform will send precise task deployment instructions to the channel through standard API interfaces, including container images, startup commands, environment variables, and data mount points. At the same time, the platform will establish a persistent task execution monitoring session to continuously pull or receive actual resource usage data (such as GPU utilization and memory consumption), execution progress logs, and any possible abnormal events (such as node failures and program crashes) from the target channel.
[0098] (Step S614) During the execution of a computing task, if a channel failure occurs or performance deviates significantly from expectations, a rescheduling mechanism is triggered to reallocate the remaining computing tasks based on the current candidate channels. If the monitoring session detects a hardware failure, network interruption, or performance metrics significantly deviating from expected thresholds during the execution of a computing task, the platform will automatically trigger the rescheduling mechanism. At this time, the platform will quickly reassign the unfinished tasks to another available and suitable channel based on the latest list of idle candidate channels, for example, migrating from the failed Cluster_Beta to the now idle Cluster_Alpha, thereby ensuring the final completion of user tasks and achieving highly reliable computing service delivery.
[0099] Please see the appendix Figure 7 The billing methods based on the computing power provision channel and actual resource usage include:
[0100] Step S621) Obtain the actual resource type, usage, and usage duration data from the task execution monitoring session. The platform will initiate an automated and precise billing process. Extract precise actual consumption data from the established task execution monitoring session, including the specific resource type used (e.g., NVIDIA_V100 instance, 100GB object storage capacity), actual usage (e.g., 3.5 GPU hours), and usage duration accurate to the second (e.g., the task ran from 14:02:33 on 2023-11-01 to 16:30:15 on 2023-11-01).
[0101] Step S622) Based on the billing method declared by the target computing power providing channel, automatically match the corresponding billing model. The billing method includes billing based on computing power usage, billing based on usage duration, or billing based on usage duration in different time periods. The platform automatically calls the corresponding billing model for calculation based on the pre-declared billing method of the target computing power providing channel. For example, if channel CloudProvider_X uses billing based on vCPU and memory usage, the platform applies its unit price based on the extracted usage data; if channel CloudProvider_Y uses simple billing based on instance usage duration, the platform directly calculates the billing by multiplying the duration by the hourly rate; for more complex billing models, such as time-based billing, the platform will enter a more detailed processing logic.
[0102] (Step S623) If it is a time-segmented billing mode, the cost is calculated in segments based on the actual execution time of the task and the time-segment price list published by the channel.
[0103] When encountering a time-segmented billing model, the platform will calculate the cost in segments based on the actual start and end timestamps of the task and the dynamic time-segment price list published by the channel. For example, if a certain computing power channel charges 2 yuan / hour during weekday daytime (09:00-18:00) and 1 yuan / hour during nighttime and weekends, and a task runs from 17:30 on Friday afternoon to 01:30 on Saturday morning, the platform will automatically split the usage time into three billing segments: Friday 17:30-18:00 (daytime rate), Friday 18:00-24:00 (nighttime rate), and Saturday 00:00-01:30 (nighttime rate), and calculate the cost for each segment separately before adding them up.
[0104] Step S624) Compare the calculated cost with the user's budget to generate a cost breakdown and bill, and support aggregated settlement by task granularity or user account granularity.
[0105] After calculating the cost, the platform compares the total cost with the budget set when the user submitted the task to ensure that there was no budget overrun (or that the cost was within the pre-authorized range). Then, it automatically generates a detailed cost breakdown and bill. This bill can be presented at the level of an individual task or by aggregating all task costs by user account period (such as monthly) to form a unified settlement statement.
[0106] Step S625) Synchronize the settlement results to the platform's financial system and provide users with auditable billing vouchers. At the same time, provide feedback on the billing data for this billing to optimize the subsequent matching degree evaluation model.
[0107] Billing results (32) will be automatically synchronized to the enterprise's internal financial system or payment gateway via API to complete deductions or accounting. Simultaneously, the platform will provide users with an auditable billing voucher containing the task ID, resource usage details, timestamps, and billing formulas, ensuring transparency in the billing process. Actual task costs, resource utilization efficiency, and other billing-related data will be structured and stored in a historical record database, serving as key training data for subsequent optimization of the matching evaluation model. This allows the model to more accurately learn the balance between cost-effectiveness and user budget, achieving continuous optimization of billing and scheduling.
[0108] An AI-based MaaS platform intelligent scheduling and settlement system; please refer to the appendix. Figure 8 ,include:
[0109] Input module 100 obtains the computing power demand description 10 described by the user in natural language;
[0110] The extraction module 200 extracts the semantics of the computing power requirement description 10 to obtain task feature parameters 11, which include task type, computing power scale requirement, latency sensitivity, budget constraint and execution time window.
[0111] The access module 300 obtains channel description information 20 of multiple idle computing power provision channels accessed by the platform. The channel description information 20 includes at least computing power scale, current load status, billing method, unit price, network latency and adapted task type.
[0112] The matching module 400 calculates the channel matching degree between each computing power supply channel and the computing power demand description 10 based on the task feature parameters 11 and the channel description information 20 of the computing power supply channel using a pre-trained matching degree evaluation model.
[0113] The strategy module 500 generates a computing power provision strategy 31 based on the channel matching degree. The computing power provision strategy 31 includes the channel priority sorting corresponding to each task type.
[0114] The scheduling module 600 schedules the user's computing power tasks to the target computing power supply channel for execution based on the computing power supply strategy 31 and the actual idle status of the computing power supply channel, and completes the billing based on the billing method of the computing power supply channel and the actual resource usage.
[0115] Please see Figure 9 The diagram shown is a structural schematic of an electronic device provided in an embodiment of this specification.
[0116] like Figure 9 As shown, the electronic device 1100 may include: at least one processor 1101, at least one network interface 1104, a user interface 1103, a memory 1105, and at least one communication bus 1102. The communication bus 1102 can be used to connect and communicate with the various components mentioned above. The user interface 1103 may include buttons, and optionally may include standard wired or wireless interfaces. The network interface 1104 may include, but is not limited to, a Bluetooth module, an NFC module, or a Wi-Fi module. The processor 1101 may include one or more processing cores. The processor 1101 connects to various parts within the electronic device 1100 using various interfaces and lines, and performs various functions of the routing device and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 1105, and by calling data stored in the memory 1105. Optionally, the processor 1101 may be implemented using at least one hardware form of DSP, FPGA, or PLA. The processor 1101 may integrate one or more combinations of CPU, GPU, and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content that the display screen needs to show; and the modem is used for wireless communication.
[0117] It is understandable that the aforementioned modem may not be integrated into the processor 1101, but may be implemented using a separate chip.
[0118] The memory 1105 may include RAM or ROM. Optionally, the memory 1105 may include a non-transitory computer-readable medium. The memory 1105 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 1105 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 1105 may also be at least one storage device located remotely from the aforementioned processor 1101. As a computer storage medium, the memory 1105 may include an operating system, a network communication module, a user interface module, and application programs. The processor 1101 may be used to call the application programs stored in the memory 1105 and execute the methods in the above-described embodiments.
[0119] This specification also provides a computer-readable storage medium storing instructions that, when executed on a computer or processor, cause the computer or processor to perform multiple steps as described in the above embodiments. If the constituent modules of the above-described electronic device are implemented as software functional units and sold or used as independent products, they can be stored in the computer-readable storage medium.
[0120] This specification also provides a computer program product, including a computer program that, when executed by a processor, implements the multiple steps described in the above embodiments.
[0121] Where there is no conflict, the technical features in this embodiment and implementation scheme can be combined arbitrarily.
[0122] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes multiple computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this specification are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center integrating multiple available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital versatile discs (DVDs)), or semiconductor media (e.g., solid-state drives (SSDs)).
[0123] When implemented through hardware or firmware, the aforementioned method flow is programmed into the hardware circuit to obtain the corresponding hardware circuit structure and achieve the corresponding function. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit, whose logic function is determined by the user programming the device. Designers can program a digital system onto a PLD themselves, eliminating the need for chip manufacturers to design and fabricate dedicated integrated circuit chips. Furthermore, nowadays, instead of manually fabricating integrated circuit chips, this programming is mostly implemented using "logic compiler" software, similar to the software compiler used in program development. The original code before compilation must also be written in a specific programming language, called a Hardware Description Language (HDL). There is not just one HDL, but many. Those skilled in the art should understand that by simply performing some logic programming on the method flow using one of the aforementioned hardware description languages and programming it into an integrated circuit, the hardware circuit implementing the logic method flow can be easily obtained.
[0124] The embodiments described above are merely preferred embodiments of this specification and are not intended to limit the scope of this specification. Any modifications and improvements made by those skilled in the art to the technical solutions of this specification without departing from the spirit of this specification should fall within the protection scope defined by the claims of this specification.
Claims
1. An AI-based intelligent scheduling and settlement method for a MaaS platform, characterized in that, Includes the following steps: Obtain a description of the user's computing power requirements in natural language; Extract the semantics of the computing power requirement description to obtain task feature parameters, which include task type, computing power scale requirement, latency sensitivity, budget constraint, and execution time window; Obtain channel description information for multiple idle computing power provision channels connected to the platform. The channel description information includes at least computing power scale, current load status, billing method, unit price, network latency, and compatible task type. Based on the task feature parameters and the channel description information of the computing power provision channels, the matching degree between each computing power provision channel and the channel description of the computing power demand is calculated using a pre-trained matching degree evaluation model. A computing power provision strategy is generated based on the channel matching degree, and the computing power provision strategy includes channel priority ranking for each task type; Based on the computing power provision strategy and the actual idle status of the computing power provision channel, the user's computing power tasks are scheduled to be executed on the target computing power provision channel, and billing is completed based on the billing method of the computing power provision channel and the actual resource usage.
2. The AI-based intelligent scheduling and settlement method for a MaaS platform according to claim 1, characterized in that, Methods for extracting the semantics of the computing power requirement description and obtaining task feature parameters include: Natural language processing is performed on the computing power requirement description, and key semantic elements are identified using the accessed large language model; The identified key semantic elements are mapped to predefined task feature parameters, including task type, computing power requirements, latency sensitivity, budget constraints, and execution time window. For parameters not explicitly mentioned, the large language model is prompted to perform reasoning and completion based on contextual semantics and historical user behavior data, and to assign confidence weights. All extracted and inferred parameters are standardized to form a structured task feature parameter vector.
3. The AI-based intelligent scheduling and settlement method for a MaaS platform according to claim 1, characterized in that, Methods for training matching evaluation models include: Read the computing power scheduling record, which includes the user's original natural language demand description, the corresponding extracted task feature parameters, the channel description information of the selected computing power provision channel, and the actual execution performance indicators; A training sample set is constructed based on the historical records. Each sample contains a task feature parameter vector, a channel description information vector, and a corresponding matching degree label. The matching degree label is a comprehensive adaptation score of the actual execution of the computing power channel. The comprehensive adaptation score is calculated by weighting the execution success rate, resource utilization rate, cost-effectiveness, and latency compliance rate. Construct a machine learning or deep learning model as the architecture for the matching degree evaluation model. After fusing and encoding the task feature parameter vector and the channel description information vector, input the model to predict the channel matching degree. The model is trained end-to-end to optimize its parameters by minimizing the loss function between the predicted matching degree and the label. Once the preset accuracy is achieved, the training is completed, and a matching degree evaluation model is obtained.
4. The AI-based intelligent scheduling and settlement method for a MaaS platform according to claim 1, characterized in that, The method for generating a computing power provision strategy based on the channel matching degree includes: The computing power delivery channels are sorted in descending order of their channel matching degree to form a candidate channel list; Based on budget constraints and billing methods, the cost feasibility of candidate channels is verified, and options that exceed the user's budget are eliminated. Based on latency sensitivity and execution time window, remove channels that do not meet latency requirements and time availability; The remaining computing power supply channels are used as candidate channels. Based on the task type and the matching task type tag of the computing power supply channel, several candidate channels are selected for each task type. Based on the channel matching degree and idle load capacity of the candidate channels, a channel priority ranking is generated for each task type.
5. The AI-based intelligent scheduling and settlement method for a MaaS platform according to claim 1, characterized in that, The method for scheduling a user's computing power tasks to a target computing power delivery channel for execution, based on the aforementioned computing power delivery strategy and the actual idle status of the computing power delivery channel, includes: Real-time query of the dynamic idle resource status of candidate channels during the task plan execution period, the dynamic idle resource status including the number of available computing power units, bandwidth margin and estimated queuing time; Based on the channel priority ranking and real-time idle status of the candidate channels, select the candidate channels that meet the task characteristic parameters and whose resources can be allocated in real time as the target channels. Send task deployment instructions to the selected target channels and establish a task execution monitoring session to continuously collect the actual resource usage, execution progress and abnormal events of computing power tasks; If a channel failure or performance deviates significantly from expectations during the execution of a computing task, a rescheduling mechanism is triggered to reallocate the remaining computing tasks based on the current candidate channels.
6. The AI-based intelligent scheduling and settlement method for a MaaS platform according to claim 1, characterized in that, The methods for billing based on the computing power provision channel and actual resource usage include: Obtain data on the actual resource types, usage, and duration of use from the task execution monitoring session; Based on the billing method declared by the target computing power providing channel, the corresponding billing model is automatically matched. The billing method includes billing based on computing power usage, billing based on usage time, or billing based on usage time in different time periods. If it is a time-segmented billing model, the fee will be calculated in segments based on the actual execution time of the task and the time-segment price list published by the channel. The calculated costs are compared with the user's budget to generate a detailed cost breakdown and bill, and aggregated settlement is supported by task granularity or user account granularity. The settlement results are synchronized to the platform's financial system, and auditable billing vouchers are provided to users. At the same time, the billing data is fed back to optimize the subsequent matching degree evaluation model.
7. An AI-based MaaS platform intelligent scheduling and settlement system, characterized in that, include: The input module obtains the user's computing power requirements described in natural language; The extraction module extracts the semantics of the computing power requirement description to obtain task feature parameters, which include task type, computing power scale requirement, latency sensitivity, budget constraint, and execution time window. The access module obtains channel description information of multiple idle computing power provision channels connected to the platform. The channel description information includes at least computing power scale, current load status, billing method, unit price, network latency, and adapted task type. The matching module calculates the matching degree between each computing power supply channel and the channel description information of the computing power supply channel based on the task feature parameters and the channel description information of the computing power supply channel using a pre-trained matching degree evaluation model. The strategy module generates a computing power provision strategy based on the channel matching degree. The computing power provision strategy includes the channel priority ranking corresponding to each task type. The scheduling module schedules the user's computing tasks to the target computing power supply channel for execution based on the computing power supply strategy and the actual idle status of the computing power supply channel, and completes the billing based on the billing method of the computing power supply channel and the actual resource usage.
8. An electronic device, characterized in that, Including the processor and memory; The processor is connected to the memory; The memory is used to store executable program code; The processor runs a program corresponding to the executable program code stored in the memory to perform the method as described in any one of claims 1-6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.