Heterogeneous computing power collaborative adaptation method based on distributed pooling computing power

Abstract

The invention discloses a heterogeneous computing power collaborative adaptation method based on distributed pooling computing power, and relates to the technical field of heterogeneous computing power resource scheduling. According to the method, a basic data set is collected from a computing power heterogeneous platform, and a time sequence feature set and a scene feature set are obtained through standardization processing; the computing power demand trend of each heterogeneous cluster is accurately predicted based on a computing power prediction model composed of a basic time sequence prediction layer, a scene feature enhancement layer and an output fusion layer; setting a demand priority in combination with a computing power growth rate threshold and a scene label; a scheduling decision is generated through a coordinated scheduling model, and resource allocation is optimized by means of a dynamic adaptive adjustment mechanism; according to the method, refined representation of the heterogeneous computing power and accurate task matching are realized, the limitation of stiffness of an adaptation strategy of a traditional scheme is broken through, the adaptation accuracy of the heterogeneous computing power and the comprehensive utilization rate of resources are effectively improved, meanwhile, the performance fluctuation of core services is remarkably reduced, and dynamic balance of efficient utilization of the resources and stable operation of the services is achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of heterogeneous computing resource scheduling, and particularly relates to a heterogeneous computing resource collaborative adaptation method based on distributed pooled computing power. BACKGROUND

[0002] Under the background of digital economic development driven by cloud native and AI native, distributed pooled computing power has become a core technical mode supporting high computing power demand in complex scenarios. It aggregates heterogeneous computing power hardware such as CPU, GPU, NPU distributed in different physical locations and different ownership subjects through network technology to form a logically unified computing power resource pool, and realizes on-demand allocation and collaborative use of computing power resources. The existing heterogeneous computing power adaptation scheme mainly includes three types of open source adaptation scheme based on container orchestration, cross-platform computing power integration framework and distributed parallel computing framework. Among them, the distributed parallel computing framework divides complex tasks into subtasks and allocates them to different computing power heterogeneous platforms for parallel execution, combines a unified task scheduling and data synchronization mechanism, and realizes preliminary integration and use of computing power resources in big data analysis, AI computing and other scenarios. It is one of the core technical solutions in the field of distributed pooled computing power.

[0003] The existing distributed parallel computing framework still has significant technical defects due to the design of the technical architecture and the adaptation logic. On the one hand, the resource description model focuses on the basic parameters of traditional computing units, lacks fine characterization ability of heterogeneous hardware performance characteristics, and the resource allocation strategy and concurrency control logic are mostly statically configured, lacking a dynamic adjustment mechanism, resulting in a lack of heterogeneous computing power collaborative adaptation ability, and the inability to achieve precise matching of task characteristics and hardware capabilities, making it difficult to improve resource utilization. On the other hand, the scheduling system relies on a static topology model and a periodic load sampling mechanism, with a long sampling period, which cannot capture the dynamic fluctuations of computing power demand and the instantaneous changes of resource pool state in real time, and the scheduling model lacks adaptive adaptation ability to task types and business scenarios, and the cross-node and cross-cluster load balancing adjustment ability is weak, resulting in a lag in dynamic scheduling response, making it difficult to adapt to sudden computing power demand and cross-domain collaborative scenarios, and seriously affecting task execution efficiency and business operation stability.

[0004] To solve the above technical problems, the application provides a heterogeneous computing power collaborative adaptation method based on distributed pooled computing power. SUMMARY

[0005] The application aims to at least solve one of the technical problems in the prior art. To this end, the application provides a heterogeneous computing power collaborative adaptation method based on distributed pooled computing power.

[0006] To achieve the above purpose, the first aspect of the application provides a heterogeneous computing power collaborative adaptation method based on distributed pooled computing power, comprising: Basic datasets are collected from heterogeneous computing power platforms, and the basic datasets are standardized to obtain feature datasets. The basic datasets are used to analyze the performance of heterogeneous computing power platforms, while the feature datasets include time-series feature sets and scenario feature sets. The feature dataset is input into a pre-trained computing power prediction model to predict computing power demand trends; demand priorities are set based on computing power demand trends and scenario labels; where scenario labels are extracted from the basic dataset. Based on computing power demand trends, basic datasets, demand priorities, and a pre-trained coordination and scheduling model, scheduling decisions are generated; and computing power resources of heterogeneous computing power platforms are scheduled according to the scheduling decisions.

[0007] In one possible implementation, a basic dataset is collected from a heterogeneous computing platform, including: Set the sampling frequency; the sampling frequency can be set separately according to different data types; Basic datasets are collected from heterogeneous computing platforms according to the sampling frequency; the basic datasets include resource status data, task status data, and network link data.

[0008] In one possible implementation, the standardization process includes normalization, one-hot encoding, and sliding window statistics.

[0009] In one possible implementation, the platform architecture of a heterogeneous computing platform includes a computing resource pool; The computing power resource pool consists of several heterogeneous clusters; the heterogeneous clusters are built from computing power nodes of the same type; the computing power nodes include CPU nodes, GPU nodes and NPU nodes.

[0010] In one possible implementation, the trend of computing power demand is predicted, including: The computing power prediction model is invoked; the computing power prediction model is constructed based on the LSTM model and the gradient boosting tree model. The feature dataset is input into the computing power prediction model to predict the computing power demand trend of the heterogeneous computing power platform; the computing power demand trend includes the computing power demand of each heterogeneous cluster.

[0011] In one possible implementation, the computing power prediction model includes a basic time-series prediction layer, a scene feature enhancement layer, and an output fusion layer; The basic time series prediction layer is built on an LSTM model to predict computing power demand trends; the scene feature enhancement layer is built on a gradient boosting tree model to correct computing power demand trends. The output fusion layer is used to fuse the outputs of the basic time series prediction layer and the scene feature enhancement layer according to the fusion weight coefficient to obtain the computing power demand trend.

[0012] In one possible implementation, demand priorities are set based on computing power demand trends and scenario tags, including: Set a threshold for the computing power growth rate; The computing power growth rate threshold is used to match the computing power demand trend with the demand priority; when the demand priority and the scenario label are inconsistent, the scenario label is used to set the demand priority.

[0013] In one possible implementation, scheduling decisions are generated based on computing power demand trends, basic datasets, demand priorities, and a pre-trained coordination and scheduling model, including: Invoke the coordination and scheduling model; the coordination and scheduling model is built based on the DDPG algorithm; A state space is constructed based on computing power demand trends, basic datasets, and demand priorities. The state space is then input into the coordination and scheduling model to obtain action combinations. The action combination is validated against constraints to generate a scheduling decision; the scheduling decision includes path allocation, node selection, and task fragmentation ratio.

[0014] In one possible implementation, the coordination and scheduling model includes an input layer, a state coding layer, a core decision layer, and a decision output layer. The core decision layer makes decisions using MDP and DDPG algorithms.

[0015] In one possible implementation, the computing resources of a heterogeneous computing platform are scheduled according to scheduling decisions, including: Send scheduling decisions to heterogeneous computing power platforms; The heterogeneous computing platform generates node task instructions based on scheduling decisions and selects computing nodes as target nodes according to task allocation strategies; among which, task allocation strategies include proximity principle or load balancing. The node task instruction is sent to the target node, and the target node starts the task execution.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention collects basic datasets from heterogeneous computing power platforms, and obtains time-series feature sets and scene feature sets through standardization processing; relying on a computing power prediction model composed of a basic time-series prediction layer, a scene feature enhancement layer, and an output fusion layer, it accurately predicts the computing power demand trends of each heterogeneous cluster; it sets demand priorities by combining computing power growth rate thresholds and scene labels; and it generates scheduling decisions through a coordination and scheduling model, while optimizing resource allocation through a dynamic adaptation and adjustment mechanism; this invention achieves refined characterization of heterogeneous computing power and accurate task matching, breaking the limitations of rigid adaptation strategies in traditional solutions, effectively improving the accuracy of heterogeneous computing power adaptation and the overall utilization rate of resources, while significantly reducing the performance fluctuation of core businesses, achieving a dynamic balance between efficient resource utilization and stable business operation.

[0017] 2. This invention utilizes a computing power prediction model that integrates LSTM and gradient boosting tree models to predict computing power demand trends in advance and prioritize demands based on scenario labels. A state space is constructed based on computing power demand trends, demand priorities, and a basic dataset. A coordination scheduling model outputs the action combinations corresponding to the state space, and scheduling decisions are generated after constraint verification. Computing power resources are scheduled using proximity principles or load balancing task allocation strategies. This invention breaks free from the constraints of traditional static scheduling models, achieving advance prediction and real-time response to computing power demands. It significantly shortens scheduling response latency, effectively reduces node load differences, improves scheduling efficiency and load balancing, successfully adapts to sudden computing power demands and cross-domain collaborative scenarios, and promotes the transformation of distributed scheduling technology from passive response to proactive prediction-driven approaches. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method flow for the heterogeneous computing power collaborative adaptation method based on distributed pooling computing power in this invention. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Distributed pooled computing power, driven by both cloud-native and AI-native technologies in the digital economy, aggregates computing resources from heterogeneous hardware such as CPUs, GPUs, NPUs, and DPUs across different physical locations and ownership entities using network technology. This forms a logically unified and flexibly schedulable pool of computing resources. Based on a unified management and scheduling system, it enables on-demand allocation, collaborative operation, and efficient utilization of computing resources. Its core is breaking down the physical isolation and ownership barriers of computing resources, transforming dispersed computing power into a core production factor that can be scalably scheduled and used intensively, supporting the high computing power demands of complex scenarios such as AI large-scale model training, big data analysis, video processing, and intelligent recognition.

[0022] Distributed pooled computing power has the following characteristics: 1) Heterogeneity: The computing power resource pool includes various types and architectures of computing hardware, covering traditional CPUs, as well as GPUs with parallel computing advantages, NPUs for AI inference, and DPUs focused on data processing. These different hardware components exhibit significant differences in computing performance, applicable scenarios, and energy consumption; 2) The physical deployment of computing resources is decentralized, potentially distributed across different racks in the same data center, data centers in different cities, or even infrastructure of different cloud service providers, interconnected via network links; 3) Pooling integration: Through software-defined technologies, the distributed... Heterogeneous computing resources are abstracted into a unified logical resource pool, shielding the differences in underlying hardware and the dispersion of physical locations, and providing users with a one-stop computing power call interface; 4) The computing power allocation strategy is adjusted in real time according to the computing power requirements, priorities and resource pool status of business tasks, realizing the dynamic flow of computing power between different tasks and different scenarios, and improving resource utilization; 5) Heterogeneous computing power in the resource pool can work together according to task characteristics. For example, the CPU is responsible for logic control and task scheduling, the GPU is responsible for parallel computing intensive tasks, and the NPU is responsible for AI inference tasks. Through collaboration, the diversified computing power requirements of complex tasks can be met.

[0023] Distributed parallel computing frameworks (such as Apache Spark) are one of the core technologies in the field of distributed pooled computing power. Although they have achieved initial integration and efficient utilization of computing resources in scenarios such as big data analysis and AI computing through task decomposition and parallel execution models, they still have significant technical bottlenecks in key aspects such as heterogeneous collaboration, scheduling response, communication efficiency, fault tolerance and recovery, and cross-domain management due to limitations in technical architecture design and adaptation logic. The specific problems are analyzed in detail below: 1. Resource description models in distributed parallel computing frameworks often focus on basic parameters of traditional computing units such as CPUs (e.g., number of cores, memory capacity), lacking the ability to fine-tune the performance characteristics of heterogeneous hardware such as GPUs, NPUs, and DPUs. For example, the framework cannot accurately quantify the core differences in parallel computing efficiency, data processing latency, and energy consumption between heterogeneous chips with different architectures (e.g., Ascend 910 and NVIDIA A100) in AI training tasks, only distinguishing hardware based on simple labels such as "whether or not GPU is supported." This leads to an inability to achieve precise matching of "task characteristics - hardware capabilities" during task allocation. For instance, assigning high-parallelism AI inference tasks to CPU clusters or having GPUs, which excel at high-precision computing, perform simple data sorting tasks results in inefficient operation of high-performance computing units and serious resource waste.

[0024] 2. The resource allocation strategies and concurrency control logic of distributed parallel computing frameworks are mostly statically configured, unable to be dynamically adjusted according to real-time hardware status and task characteristics. On the one hand, the framework lacks real-time awareness and adaptive adjustment capabilities for dynamic changes in hardware status (such as a sudden increase in GPU load or excessive NPU temperature), still allocating resources according to the initial configuration, leading to an unbalanced state where some nodes are overloaded and others are idle. On the other hand, when the task type dynamically switches (such as switching from an I / O-intensive data reading task to a computationally sensitive model training task), the framework cannot adjust the resource adaptation strategy in time, still using the original concurrency and resource allocation ratio, resulting in a significant decrease in task execution efficiency. In addition, although some frameworks support resource overselling to improve utilization, the overselling ratio is a fixed value and cannot be dynamically adjusted according to business priority (such as core tasks versus non-core tasks) and hardware load status, which can easily cause fluctuations in the performance of core businesses.

[0025] Please see Figure 1 The first aspect of this invention provides a method for heterogeneous computing power collaborative adaptation based on distributed pooled computing power, comprising: collecting a basic dataset from a heterogeneous computing power platform; standardizing the basic dataset to obtain a feature dataset; wherein the basic dataset is used to analyze the performance of the heterogeneous computing power platform, and the feature dataset includes a time-series feature set and a scene feature set; inputting the feature dataset into a pre-trained computing power prediction model to predict computing power demand trends; setting demand priorities based on computing power demand trends and scene labels; generating scheduling decisions based on computing power demand trends, basic datasets, demand priorities, and a pre-trained coordination and scheduling model; and scheduling the computing power resources of the heterogeneous computing power platform according to the scheduling decisions.

[0026] When scheduling heterogeneous computing resources, the basic dataset of the heterogeneous computing platform can be collected by the real-time sensing module at a set collection frequency. The basic dataset includes the following: 1) Resource status data, which mainly refers to the operating parameters of the heterogeneous computing platform, including load, memory utilization, computing power utilization, temperature, energy consumption, data transmission rate, etc., to accurately characterize hardware performance and its operating status. 2) Task status data, which mainly refers to task parameters, including task type (such as IO-intensive, computing power-sensitive, memory-intensive, etc.), task priority, execution progress, remaining computing power, data size, etc. 3) Network link data, which mainly includes network indicators such as bandwidth utilization, transmission latency, packet loss rate, and link stability in cross-node, cross-cluster, and cross-cloud scenarios, covering key links in cross-domain communication.

[0027] When collecting basic datasets from heterogeneous computing power platforms, the platforms use incremental compression to transmit data to the real-time sensing module. For example, a lightweight private protocol based on UDP is used for data transmission, transmitting only the data that has changed since the last transmission instead of uploading the entire dataset. This avoids duplicate data and improves data collection and transmission efficiency.

[0028] The real-time sensing module and the heterogeneous computing platform engage in bidirectional data interaction and support closed-loop control.

[0029] The heterogeneous computing platform sends the data required for coordinating computing resources to the real-time sensing module, including at least: resource status data, such as computing node operating parameters (load rate, temperature, etc.), cluster resource utilization, and overall resource pool status, providing the sensing module with computing power supply-side data; task status data, such as the task type, priority, execution progress, and remaining computing power on each node / cluster, providing the sensing module with computing power demand-side data; and network status data, such as communication latency, bandwidth utilization, and packet loss rate between nodes, between clusters, and across regions, providing the sensing module with collaborative environment data.

[0030] The real-time perception module sends control commands for coordinating computing resources to the heterogeneous computing platform. It also sends the collected basic dataset to the predictive scheduling module, receives the generated computing resource scheduling decisions, and sends these decisions to the computing resource pool. The resource management component then schedules and coordinates the computing resources according to these decisions.

[0031] The hierarchical architecture of the heterogeneous computing platform consists of a computing resource pool, a heterogeneous cluster, and computing nodes.

[0032] Computing nodes are the core physical carriers of heterogeneous computing platforms. Each computing node is an independent server (including physical machines or cloud hosts) and is classified according to the type of computing hardware it carries. The main types of computing nodes include CPU nodes, GPU nodes, and NPU nodes. CPU nodes are equipped with various types of CPUs such as Intel Xeon, AMD EPYC, and Kunpeng, along with DDR memory and NVMe storage, providing general-purpose computing and collaborative auxiliary computing power. GPU nodes are equipped with GPUs such as NVIDIA A100, along with high-bandwidth video memory and efficient cooling components, providing high-performance parallel computing capabilities. NPU nodes are equipped with NPUs such as Ascend 910.

[0033] The distribution of computing nodes is cross-regional, and they can be deployed in different data centers and interconnected through high-speed networks.

[0034] Heterogeneous clusters integrate computing nodes into independent clusters based on their functions, including CPU clusters, GPU clusters, and NPU clusters. CPU clusters consist of CPU nodes from various data centers and are mainly used for I / O-intensive and collaborative auxiliary tasks; GPU clusters consist of GPU nodes from various data centers and are mainly used for computing power-sensitive tasks; NPU clusters consist of NPU nodes from various data centers and are mainly used for low-power computing scenarios and hybrid heterogeneous collaborative tasks.

[0035] Each heterogeneous cluster achieves unified management of nodes, resource isolation, and task scheduling through containerization technology and resource management components.

[0036] The computing resource pool is constructed by logically abstracting the hardware resources of CPU clusters, GPU clusters, and NPU clusters through resource management components. This abstracts away the differences in underlying hardware and geographical distribution, providing a unified description of resource parameters. These parameters primarily include computing power, memory / GPU memory capacity, and communication bandwidth.

[0037] In one example, a heterogeneous computing platform is executing a task, which is mainly handled by the NPU cluster for large-scale matrix operations, while the CPU cluster is responsible for data I / O and logical control.

[0038] The real-time sensing module sends data acquisition commands to the NPU cluster at a sampling frequency of 10ms, and sends the data acquisition commands through a private lightweight protocol based on UDP.

[0039] In a heterogeneous computing platform, after receiving a data acquisition command, the NPU node reads its hardware status, including resource status data and task status data. The resource status data includes NPU chip temperature, NPU computing power utilization, HBM memory usage, and power consumption. The task status data includes task ID, task type, remaining computational load, and current task progress. Simultaneously, the heterogeneous computing platform collects cluster resource utilization, remaining NPU computing power, remaining CPU computing power, and the network status between the NPU cluster and the CPU cluster, and sends these data to the real-time sensing module.

[0040] In heterogeneous computing platforms, NPU nodes, NPU clusters, CPU clusters, and computing resource pools upload data according to data acquisition commands, employing incremental compression to save bandwidth. Specifically, the current data is compared with the previous data, only the changed data is packaged, and then compressed before being uploaded to the real-time sensing module. Incremental compression is only performed if historical data exists; otherwise, all data must be uploaded.

[0041] The following tasks can be achieved through the prediction scheduling module: inputting the feature dataset into a pre-trained computing power prediction model to predict computing power demand trends; setting demand priorities based on computing power demand trends and scenario labels; generating scheduling decisions based on computing power demand trends, basic datasets, demand priorities, and a pre-trained coordination scheduling model; and scheduling computing power resources of heterogeneous computing power platforms according to the scheduling decisions.

[0042] In a preferred embodiment, predicting the computing power demand trend includes: calling a computing power prediction model; wherein the computing power prediction model is constructed through a pre-trained artificial intelligence model; and inputting a feature dataset into the computing power prediction model to predict the computing power demand trend of the heterogeneous computing power platform.

[0043] In a preferred embodiment, the computing power prediction model includes a basic time series prediction layer, a scene feature enhancement layer, and an output fusion layer. The basic time series prediction layer is built based on an LSTM model and is used to predict computing power demand trends. The scene feature enhancement layer is built based on a gradient boosting tree model and is used to correct computing power demand trends. The output fusion layer is used to fuse the output results of the basic time series prediction layer and the scene feature enhancement layer according to the fusion weight coefficient to obtain the computing power demand trend.

[0044] Architecture design of computing power prediction model: basic time series prediction layer and scene feature enhancement layer.

[0045] Basic Time Series Prediction Layer: The Time Series Prediction Model (LSTM) uses the computing power demand time series characteristics constructed based on time series data (such as resource usage trends, task growth rates, etc.) to predict future computing power demand trends.

[0046] Scene feature enhancement layer: Uses gradient boosting tree model to extract business scene features (such as e-commerce promotion cycle, weather warning period, business peak period identifier, cross-domain task execution plan, etc.) to correct the predicted computing power demand trend, so as to improve the prediction error of sudden computing power demand.

[0047] The training dataset required for training the computing power prediction model includes: resource status data, task status data, network status data, and scene feature data. Scene feature data consists of business scenario identifiers and time tags recorded by the real-time perception module. Business scenario identifiers include e-commerce promotions, weather warnings, and cross-domain tasks, while time tags include peak business periods and low-load periods in the early morning.

[0048] After cleaning and standardizing the basic dataset, feature extraction is performed to obtain a time-series feature set and a scenario feature set. The time-series feature set is used to represent the temporal variation pattern of computing power demand and is extracted from the time-series data corresponding to resource status data and task status data. The scenario feature set is used to characterize the impact of business scenarios on computing power demand and is extracted from scenario feature data, network / task attributes, and other data.

[0049] The time-series feature set is constructed based on a base dataset, namely resource status data, task status data, and network link data. This data is cleaned and then standardized. The standardization process includes normalization and one-hot encoding, with the aim of eliminating dimensions.

[0050] A sliding window is set up for each type of data in the basic data. Temporal feature codes are extracted from the sliding window. The temporal feature codes are feature vectors of fixed length. Taking GPU load rate as an example, the sliding window is 60 seconds, and a data point is taken every 10 seconds. Temporal feature codes such as load mean, maximum value, standard deviation, slope, number of peaks, and trend direction are extracted from the sliding window. The temporal feature codes are then concatenated with the original data to form a temporal feature vector.

[0051] The time-series feature set includes time-series feature vectors of computing nodes, heterogeneous clusters, and task queues. The task queue's time-series feature vector includes task queue length, task submission rate, and task type ratio, extracted from the task queues of the heterogeneous cluster. Simultaneously, each computing node also has its own corresponding task queue, which is primarily responsible for task execution scheduling.

[0052] When creating the scene feature set, scene labels, such as business scenario codes and time tag codes, are extracted from the scene feature data. Task attribute features, such as task type codes, task priority, remaining computational load, and data size, are extracted from the task status data. Simultaneously, network coordination features, such as cross-domain latency, bandwidth utilization, and packet loss rate, are extracted from the network link data. The scene feature set is then concatenated with the output of the basic time series prediction layer and used as the input to the scene feature enhancement layer.

[0053] When training the computing power prediction model, layered training is performed first, followed by cross-layer fusion optimization.

[0054] Layered training involves training the basic temporal prediction layer and the scene feature enhancement layer sequentially.

[0055] The basic time series prediction layer consists of an input layer, hidden layers, and an output layer. The input layer receives the time series feature set; the hidden layer is configured with three LSTM units, with the number of neurons in the three layers set to 128, 64, and 32 respectively. Dropout (dropout_rate=0.2) is used to prevent overfitting, and a BatchNorm layer is introduced to accelerate convergence; the output layer is a fully connected layer that outputs the computing power demand trend for a future period (e.g., 1-5 minutes). This computing power demand trend can be categorized according to node type (CPU, GPU, NPU), corresponding to the demand scale of the three types of heterogeneous computing power resources. It should be noted that the computing power demand trend is a predicted value of the computing power demand over a predicted future period.

[0056] The hyperparameters during training were set as follows: the optimizer was AdamW, the initial learning rate was 0.001, decayed by 10% every 5 epochs, the loss function was mean squared error (MSE), the number of training epochs was set to 50, and the batch size was set to 64. The temporal feature set was divided into training, validation, and test sets according to the proportions, and the basic temporal prediction layer was trained using the training, validation, and test sets.

[0057] The scene feature enhancement layer consists of an input layer, a dual-branch XGBoost structure (corresponding to classification and regression branches), and an output layer. The input layer receives the scene feature set and the computing power demand trend output by the basic time series prediction layer. The dual-branch XGBoost structure has 200 decision trees, a maximum tree depth of 8, and a learning rate of 0.05. It adopts a hybrid output mode of classification and planning, that is, it first classifies and determines the current scene type, and then adjusts the computing power demand trend for different scenes.

[0058] To adapt to the collaborative needs of heterogeneous computing power, the output fusion layer employs a dynamic weight allocation mechanism. A fusion weight coefficient is introduced during dynamic weight allocation. This coefficient represents the weight of the output data from the basic time-series prediction layer, while the weight of the output data from the scene feature enhancement layer is 1 minus the fusion weight coefficient. The basic time-series prediction layer and the scene feature enhancement layer are concatenated into an end-to-end model. The optimization objective is to minimize the MSE (Mean Separation of Expectations) between the final computing power demand trend and the actual computing power demand value. The optimized model yields the computing power prediction model.

[0059] The computing power prediction model incrementally updates its parameters at set time intervals (e.g., every 5 minutes) based on the deviation between actual computing power demand and computing power demand trends. This ensures that the computing power prediction model adapts to the dynamic changes of different business scenarios and maintains the stability of its long-term prediction accuracy.

[0060] In the implementation of incremental updates, the basic dataset for the most recent 5 minutes is collected every 5 minutes through the real-time perception module. This basic dataset is then processed and added to the training set. A rolling window strategy is used to delete the oldest expired samples in the training set. During the incremental update process, the underlying parameters of the basic temporal prediction layer and the feature enhancement layer are fixed. Only the fully connected layer, the fusion weight coefficients, and the XGBoost learning rate are fine-tuned. The fine-tuned fully connected layer is connected to the outputs of the basic temporal prediction layer and the feature enhancement layer, and then to the final computational power demand trend.

[0061] In one example, the real-time sensing module collects basic datasets at a set collection frequency. After preprocessing, these basic datasets are input into the computing power prediction model. The basic time series prediction layer in the computing power prediction model converts various types of data into time series feature vectors based on the constructed time window (such as the past 60 seconds). The time series feature vectors include the sliding window mean, trend slope, growth rate in the last 10 seconds, frequency of peak occurrence, volatility, etc.

[0062] In a preferred embodiment, setting demand priorities based on computing power demand trends and scenario tags includes: setting a computing power growth rate threshold; matching demand priorities to computing power demand trends based on the computing power growth rate threshold; and setting demand priorities using scenario tags when demand priorities and scenario tags are inconsistent.

[0063] The computing power demand trends are categorized into three types and prioritized accordingly, providing a clear basis for subsequent scheduling and adaptation. Sudden demand (highest priority): Scenarios where computing power demand surges in a short period, such as promotional peaks or emergency alerts, require advance resource reservation; Stable demand (medium priority): Stable computing power demand under normal business scenarios, requiring balanced resource allocation; Decreasing demand (lower priority): Scenarios where business is wrapping up or traffic is declining, allowing for optimization of resource utilization (e.g., increasing the overselling ratio).

[0064] Demand tiering is determined based on the computing power demand trend and scenario labels from the computing power prediction model, and a computing power growth rate threshold is set for each tier. The computing power demand trend is compared with the computing power growth rate threshold to initially determine the corresponding demand type. This is then verified using scenario labels. If the verification results are inconsistent, the demand type is set according to the scenario labels, and a priority is assigned to the demand type.

[0065] In one example: GPU demand increases by 25% / minute (≥20% threshold) + scene tag "e-commerce promotion" → judged as "sudden demand"; CPU demand increases by 35% / minute (≥30% threshold) + scene tag "daily business" → corrected to "stable demand", eliminating misjudgments caused by hardware fluctuations.

[0066] In a preferred embodiment, a scheduling decision is generated based on computing power demand trends, basic datasets, demand priorities, and a pre-trained coordination and scheduling model. This includes: invoking the coordination and scheduling model; constructing a state space based on computing power demand trends, basic datasets, and demand priorities; inputting the state space into the coordination and scheduling model to obtain action combinations; and generating a scheduling decision after constraint verification of the action combinations. The scheduling decision includes path allocation, node selection, and task sharding ratio.

[0067] In a preferred embodiment, the coordination and scheduling model includes an input layer, a state coding layer, a core decision layer, and a decision output layer, with the core decision layer implementing decisions through MDP and DDPG algorithms.

[0068] The state space of the coordinated scheduling model consists of a preprocessed feature dataset, computing power demand trend, demand priority, and discounted rewards for past N historical actions. The reward function adopts a combination of immediate rewards (such as scheduling delay and load balancing) and discounted rewards (such as resource utilization), with a discount factor γ=0.9, to guide the agent to learn the globally optimal strategy.

[0069] The DDPG algorithm in the coordinated scheduling model outputs a corresponding action space based on the state space. This action space, after being verified by action constraints, forms the scheduling decision. Action constraint verification mainly includes delay constraints, load constraints, adaptation constraints, and resource constraints.

[0070] The training process of the coordination and scheduling model is described in detail below.

[0071] Set scheduling objectives: Establish a multi-objective optimization function that balances low latency, high balance, and high utilization.

[0072] Scheduling objective 1: Scheduling response latency is less than 50ms. Scheduling response latency refers to the total time taken from the generation of a requirement or the submission of a task to the completion of the allocation process.

[0073] Scheduling objective 2: The load difference between each node is less than 20%. The node load is calculated based on resource status awareness data. This objective is to avoid node overload or idleness.

[0074] Scheduling objective 3: The overall utilization rate of heterogeneous computing resources is no less than 85%, which is achieved through dynamic adaptation and adjustment.

[0075] Learning-driven scheduling strategy: The scheduling process is modeled as a Markov decision process, a coordinated scheduling model is constructed, and the scheduling strategy is freely optimized through real-time interaction between the coordinated scheduling model and the computing resource pool, task queue, and network links.

[0076] State space: contains multi-dimensional information such as real-time status of resource pool, task queue status, network link status, and computing power demand trends; Action space: including task allocation path selection (local node / cross-domain node / cloud node), node selection (low load / near link / highly adaptable node), task sharding ratio, etc.; Reward function: The core reward factors are scheduling response latency, load balancing, resource utilization, and task execution efficiency, which guide the agent to learn the globally optimal scheduling strategy.

[0077] The coordinated scheduling model is based on an input layer → state encoding layer → core decision layer (MDP+DDPG) → decision output layer. It employs a three-stage training process: offline pre-training → online fine-tuning → closed-loop iterative optimization, achieving deep adaptation between the structure and training logic. In the offline pre-training stage, the state encoding layer and core decision layer are trained jointly. Based on a large amount of historical sample data, MDP quintuples are constructed to optimize dimensionality reduction and decision parameters. The convergence condition is that the average reward fluctuation on the validation set is ≤3% and the dimensionality reduction information retention rate is ≥90%. In the online fine-tuning stage, the underlying parameters of the state encoding layer are fixed, and the top-level structure of the core decision layer is fine-tuned to adapt to real-time input data and business scenarios. In the closed-loop iterative stage, MDP parameters are updated based on feedback from the entire system. An ε-greedy strategy is used to balance exploration and stability, and a full-link evaluation and parameter rollback mechanism is executed every 24 hours. The training process introduces an experience replay pool and a target network synchronization strategy to ensure model training stability, while matching the decision output layer verification rules to ensure the scheduling objectives are achieved.

[0078] The input layer receives the basic dataset (updated every 10ms) from the real-time perception module and the computing power demand trend from the prediction scheduling module. The basic dataset includes the status of the computing power resource pool, the status of the task queue, and the status of the network link. The computing power demand prediction data includes the computing power demand trend and demand level labels.

[0079] The state encoding layer consists of two lightweight MLP layers without complex convolutional or recurrent structures. The first layer has 64 neurons with ReLU activation; the second layer has 32 neurons with ReLU activation. The state encoding layer is used for data dimensionality reduction, reducing model inference time.

[0080] The core decision-making layer uses the MDP quintuple as its logical framework and employs the DDPG algorithm for decision-making, taking into account both discrete and continuous action spaces. The MDP quintuple is defined as follows: , where is the state space, is the action space, is the transition probability, is the reward function, and is the discount factor. The transition probability P is fitted using historical data statistics, and the discount factor γ is fixed at 0.9. The DDPG algorithm initializes the Actor network with a He normal distribution and the Critic network with a Xavier normal distribution. The pre-training epochs are set to 100, the batch size to 128, and the optimizer is AdamW (learning rate 0.001, weight decay 1e-5). The loss function is a combination of Actor policy loss and Critic temporal difference loss. The pre-training convergence condition is that the average reward on the validation set fluctuates by ≤3% over 10 consecutive epochs.

[0081] After the coordination and scheduling model is trained, the basic dataset, demand priority, and computing power demand trend are input into the model, and the corresponding action combination is output. This action combination is a ternary combination of path allocation, node selection, and task sharding ratio. The decision output layer selects the action combination with the highest probability as the scheduling decision. Before outputting this scheduling decision, the action combination needs to be validated according to the scheduling objective; only after the validation is passed can the scheduling decision be output.

[0082] In a preferred embodiment, scheduling the computing resources of the heterogeneous computing platform according to the scheduling decision includes: sending the scheduling decision to the heterogeneous computing platform; the heterogeneous computing platform generating node task instructions according to the scheduling decision, and selecting computing nodes as target nodes according to the task allocation strategy; wherein the task allocation strategy includes proximity principle or load balancing; sending the node task instructions to the target node, and the target node starting task execution.

[0083] The coordination and scheduling model uses the RPC protocol to send cluster-level scheduling decisions to the resource management component of the heterogeneous computing platform. The resource management component breaks down the scheduling decisions into node-level task allocation instructions and selects the target node based on node load and link latency. After the task is executed, the real-time sensing module collects the node-level status within 10ms and feeds it back, updating the state space of the coordination and scheduling model to form a closed-loop control.

[0084] In one example, assume the scheduling decision is: allocate the recommended task (task ID: Rec-20260001) to the GPU cluster, using a local scheduling strategy, without task sharding, and with priority P1. The resource management component of the heterogeneous computing platform breaks down the scheduling decision into node-level instructions. Based on the "proximity allocation + load balancing" strategy, it selects the node with the lowest load rate and link latency below 60%. If GPU-02 has a load of 48% and a latency of 6ms; GPU-01 has a load of 52% and a latency of 7ms; and GPU-05 has a load of 55% and a latency of 6.5ms, then GPU-02 is ultimately selected as the target node. The resource management component issues a task allocation instruction to the target node GPU-02, and the target node GPU-02 starts task execution upon receiving the instruction.

[0085] It should be noted that the execution of scheduling decisions is achieved through a hierarchical scheduling mechanism, which is suitable for single cluster pre-cross-domain collaborative scenarios.

[0086] Local scheduling: For tasks within a single cluster, a "nearest allocation + load balancing" strategy is adopted, prioritizing the allocation of tasks to nodes with low network latency (≤10ms) and load rates below 60%; Cross-domain scheduling: A hierarchical collaboration model is established between regional schedulers and a global scheduler. The regional scheduler is responsible for local resource scheduling, while the global scheduler coordinates cross-cloud and cross-data center resources. When local resources cannot meet the demand, the global scheduler quickly calls on idle computing power from neighboring regions based on network status awareness data and optimal path algorithms. The cross-domain scheduling response time is ≤100ms.

[0087] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments.

[0088] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any other combination thereof. When implemented using a software program, it can be implemented entirely or partially in the form of a computer program product. This computer program product includes one or more 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 application 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 from one computer-readable storage medium to another. For example, 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 containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0089] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A heterogeneous computing power collaborative adaptation method based on distributed pooled computing power, characterized in that, The method comprises the following steps: Collecting a basic data set from a heterogeneous computing platform, and performing standardization processing on the basic data set to obtain a feature data set; wherein the basic data set is used to analyze the performance of the heterogeneous computing platform, and the feature data set comprises a time series feature set and a scene feature set; Inputting the feature data set into a pre-trained computing power prediction model to predict a computing power demand trend; and setting a demand priority according to the computing power demand trend and a scene label; wherein the scene label is extracted from the basic data set; Generating a scheduling decision based on the computing power demand trend, the basic data set, the demand priority and a pre-trained coordination and scheduling model; and scheduling computing power resources of the heterogeneous computing platform according to the scheduling decision; Generating a scheduling decision based on the computing power demand trend, the basic data set, the demand priority and a pre-trained coordination and scheduling model, comprising: Calling a coordination and scheduling model; wherein the coordination and scheduling model is constructed based on a DDPG algorithm; Constructing a state space based on the computing power demand trend, the basic data set and the demand priority, and inputting the state space into the coordination and scheduling model to obtain an action combination; The action combination generates a scheduling decision after constraint verification; wherein the scheduling decision comprises an allocation path, node selection and task fragmentation ratio.

2. The heterogeneous computing power collaborative adaptation method based on distributed pooled computing power according to claim 1, characterized in that, Collecting a basic data set from the heterogeneous computing platform, comprising: Setting a sampling frequency; wherein the sampling frequency can be set according to different data types; Collecting a basic data set from the heterogeneous computing platform according to the sampling frequency; wherein the basic data set comprises resource state data, task state data and network link data.

3. The heterogeneous computing power collaborative adaptation method based on distributed pooled computing power according to claim 1, characterized in that, The standardization processing comprises normalization processing, one-hot encoding processing and sliding window statistics.

4. The heterogeneous computing power collaborative adaptation method based on distributed pooled computing power according to claim 1, characterized in that, The platform architecture of the heterogeneous computing platform comprises a computing power resource pool; The computing power resource pool is composed of a plurality of heterogeneous clusters; the heterogeneous clusters are constructed by computing power nodes of the same type; the computing power nodes comprise CPU nodes, GPU nodes and NPU nodes.

5. The heterogeneous computing power collaborative adaptation method based on distributed pooled computing power according to claim 1, characterized in that, Predicting a computing power demand trend, comprising: Calling a computing power prediction model; wherein the computing power prediction model is constructed based on an LSTM model and a gradient boosting tree model; Inputting the feature data set into the computing power prediction model to predict a computing power demand trend of the heterogeneous computing platform; wherein the computing power demand trend comprises computing power demands of each heterogeneous cluster.

6. The heterogeneous computing power collaborative adaptation method based on distributed pooled computing power according to claim 5, characterized in that, The computing power prediction model comprises a basic time series prediction layer, a scene feature enhancement layer and an output fusion layer; The basic time series prediction layer is constructed based on an LSTM model and is used to predict a computing power demand trend; The scene feature enhancement layer is constructed based on a gradient boosting tree model and is used to correct the computing power demand trend; The output fusion layer is used to fuse the output results of the basic time series prediction layer and the scene feature enhancement layer according to a fusion weight coefficient to obtain a computing power demand trend.

7. The heterogeneous computing power collaborative adaptation method based on distributed pooled computing power according to claim 5, characterized in that, Setting a demand priority according to a computing power demand trend and a scene label, comprising: Setting a computing power growth rate threshold; Matching the demand priority to the computing power demand trend based on the computing power growth rate threshold; and setting the demand priority by using the scene label when the demand priority and the scene label are inconsistent.

8. The heterogeneous computing power collaborative adaptation method based on distributed pooled computing power according to claim 1, characterized in that, The coordination scheduling model comprises an input layer, a state coding layer, a core decision layer and a decision output layer, and the core decision layer realizes decision-making through MDP and DDPG algorithms.

9. The heterogeneous computing power collaborative adaptation method based on distributed pooled computing power according to claim 1, characterized in that, According to the scheduling decision, the computing power resources of the computing power heterogeneous platform are scheduled, which comprises: sending the scheduling decision to the computing power heterogeneous platform; According to the scheduling decision, the computing power heterogeneous platform generates a node task instruction and selects a target node as a target node according to a task allocation strategy; wherein the task allocation strategy comprises a proximity principle or load balancing; The node task instruction is sent to the target node, and the target node starts task execution.

Images