Cloud computing platform resource management method and system based on DPU

Abstract

The invention belongs to the field of computer system architecture, and discloses a DPU-based cloud computing platform resource management method and system.The method comprises the steps that firstly, by combining static resource data provided by cluster monitoring and dynamic microcosmic indexes collected by a DPU, resource availability evaluation and stability filtering are conducted on all computing nodes, and a preliminary candidate node set is constructed; then, a self-adaptive multi-target scoring function is adopted to comprehensively sort the candidate nodes to screen out target deployment nodes, and lightweight stability verification is carried out on the target deployment nodes before final deployment; by adopting the method, the resource utilization efficiency is optimized, the system resource consumption and the starting delay are obviously reduced, the operation and maintenance complexity and the resource fragmentation risk are reduced, the load balance and the node stability are improved, and meanwhile, the standardization degree of network interface configuration and the performance of the virtual hard disk technology are improved; on the premise that isolation and safety are guaranteed, the dynamic requirements of large-scale application deployment and complex service scenes are efficiently met.

Description

Technical Field

[0001] This invention belongs to the field of computer system architecture, specifically relating to a cloud computing platform resource management method and system based on DPU. Background Technology

[0002] In the current field of computer technology, virtualization technology has become a key means to improve the efficiency of computer resource utilization and optimize application deployment. Traditional virtual machine technology, by building a virtualization layer (hypervisor) on physical hardware, realizes the function of virtualizing physical resources into multiple independent virtual machines. Each virtual machine can run different operating systems and applications independently, isolated from each other, which improves computer efficiency to a certain extent. At the same time, container technology, as an innovative development direction in the field of virtualization, adopts a more lightweight virtualization approach. With the help of the containerization function of the operating system, such as Linux containers, it uses technologies such as namespaces, file system isolation, and resource control groups (cgroups) to achieve efficient isolation of applications and virtualization of the runtime environment. In addition, cloud-native technologies, relying on containerization, microservices, and other characteristics, have brought new ideas and methods to the construction, deployment, and expansion of applications, further promoting the widespread application of virtualization technology in cloud computing environments. The differences in network interface configurations among different users and virtual environments, as well as the role of virtual hard disk technology in ensuring data security and isolation, have also added new dimensions to the development of the entire virtualization technology system.

[0003] However, traditional virtual machine technology has revealed numerous problems in practical applications. On the one hand, its runtime consumes significant system resources, including disk space, memory, and processor performance. Since each virtual machine requires a complete copy of the operating system and applications, this undoubtedly incurs high overhead in resource utilization and startup time, reducing overall system performance and responsiveness. On the other hand, when facing large-scale application deployments and complex business scenarios, traditional manual deployment or static resource configuration methods struggle to meet rapidly changing business needs. This not only leads to cumbersome maintenance processes but also easily results in resource fragmentation and uneven node load, ultimately affecting service stability and reliability. While container technology and cloud-native technologies have improved the flexibility of resource utilization and application deployment to some extent, there is still room for improvement in the standardization of network interface configuration and the performance optimization of virtual hard disk technology.

[0004] In summary, existing cloud computing platform resource management methods, while ensuring application isolation and data security, result in high system resource consumption and low resource utilization efficiency, affecting the performance of virtual hard disk technology and making it difficult to meet the needs of increasingly complex business scenarios and large-scale application deployments. Summary of the Invention

[0005] This invention provides a cloud computing platform resource management method and system based on DPU. This method can effectively reduce system resource consumption, improve resource utilization efficiency, and enhance the performance of virtual hard disk technology while ensuring application isolation and data security, so as to meet the needs of increasingly complex business scenarios and large-scale application deployment.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A resource management method for a cloud computing platform based on DPU, comprising: Based on the static resource data provided by the cluster monitoring system and the dynamic micro-indicators collected by the DPU, resource availability assessment and stability filtering are performed on all computing nodes to construct a preliminary candidate node set. An adaptive multi-objective scoring function is used to comprehensively rank the candidate nodes in the initial candidate node set and select the target deployment node. Before the target deployment node is finally deployed, a lightweight stability check is performed on the target deployment node to enable resource management of the cloud computing platform.

[0007] Furthermore, based on the static resource data provided by the cluster monitoring system and the dynamic micro-indicators collected by the DPU, resource availability assessment and stability filtering are performed on all computing nodes to construct a preliminary candidate node set, including: Based on the resource requirements declared by the scheduled tasks in the static resource data provided by the cluster monitoring system, computing nodes whose remaining resources are clearly insufficient to run the task are initially eliminated. By combining the dynamic micro-indicators collected by the DPU, the real-time stability of the remaining computing nodes is evaluated and eliminated, and computing nodes that exhibit drastic resource fluctuations or stability below a preset threshold are filtered out at the current time or within a preset time. The computing nodes that pass resource availability assessment and stability filtering are used to construct an initial candidate node set.

[0008] Furthermore, in the preliminary candidate node set, the resource state vectors of the candidate nodes are specifically expressed as follows:

[0009] In the formula, This indicates the availability of a candidate node on resource type j, where j is 1 to n; The dynamic weight vector for each type of resource is obtained based on task resource requirements and system operation strategies, as follows:

[0010] In the formula, This represents the dynamic weight vector for each type of resource.

[0011] Furthermore, the adaptive multi-objective scoring function is used to comprehensively rank the candidate nodes in the initial candidate node set and filter out the target deployment nodes. The specific formula of the adaptive multi-objective scoring function is as follows:

[0012] In the formula, Indicates the first Resource class adaptive weights; Represents a node Then Availability of similar resources; This represents the short-term volatility of resources collected by the DPU, used to reflect stability. This represents the energy consumption index under unit resource scheduling, derived from the DPU energy efficiency model; This indicates the scheduling penalty for the node, reflecting its historical deployment failure rate or the number of task migrations. , , These represent different adjustment factors, used to balance resource stability, energy efficiency, and historical risk, respectively.

[0013] Furthermore, the step of employing an adaptive multi-objective scoring function to comprehensively rank the candidate nodes in the initial candidate node set and filter out the target deployment nodes includes: An adaptive multi-objective scoring function is used to score the candidate nodes in the initial candidate node set; The candidate nodes are sorted from highest to lowest score, and the candidate node with the highest score is selected as the target deployment node for the task.

[0014] Furthermore, when multiple candidate nodes with similar scores appear, a secondary comparison is performed on the multiple candidate nodes with similar scores based on secondary objectives to obtain the target deployment node; the secondary objectives include current load, service availability, and scheduling fairness.

[0015] Furthermore, before the final deployment of the target deployment node, a lightweight stability check is performed on the target deployment node to achieve resource management of the cloud computing platform, including: Before the final deployment of the target deployment node, a lightweight stability check is performed on the target deployment node; wherein, the lightweight stability check includes resource trend re-determination, scheduling congestion determination, and historical high-risk determination; The re-determination of resource trends includes: Based on the latest micro-resource data collected by the target node's DPU, the real-time changing trends of key resources are analyzed. If any adverse trend of any resource is detected to exceed the preset safety threshold, it is determined that the target deployment node has a future instability risk; otherwise, the target deployment node is determined to be normal. The scheduling congestion determination includes: Track the allocation of new tasks to all nodes within the current scheduling period. If it is detected that the preferred target node has been densely allocated multiple high-load tasks within the current period, or that resources are in a state of intense competition, it is determined that the target deployment node has a short-term scheduling congestion or resource contention overload risk; otherwise, the target deployment node is determined to be normal. The historical high-risk determination includes: Query the recent scheduling history of the target node. If the target deployment node has experienced multiple deployment failures, frequent task restarts, or task eviction events due to resource bottlenecks within a preset time period, the target deployment node is determined to be a historical high-risk node; otherwise, the target deployment node is determined to be a normal node.

[0016] Furthermore, after performing a lightweight stability check on the target deployment node before final deployment to achieve resource management of the cloud computing platform, the process also includes: If the target deployment node passes the lightweight stability check, then the task will be deployed to the target deployment node. If the target deployment node fails the lightweight stability check, the current target deployment node is abandoned, a suboptimal node is selected, and the lightweight stability check is performed again on the suboptimal node.

[0017] Furthermore, if the number of lightweight stability verification failures exceeds a preset number of consecutive failures, the following steps will be taken: The task scheduling operation is temporarily suspended, and the entire scheduling process is retried from the pre-selection stage after a period of time. Alternatively, trigger the automatic scheduling migration logic: attempt to migrate existing low-priority or migrateable tasks on the target node to free up resources for the target task; or after weighing the risks, select the node with the lowest risk among the previously failed nodes for deployment.

[0018] A cloud computing platform resource management system based on DPU, comprising: The node set construction module is used to perform resource availability assessment and stability filtering on all computing nodes based on static resource data provided by the cluster monitoring system and dynamic micro-indicators collected by the DPU, and to construct an initial candidate node set. The filtering module is used to comprehensively sort the candidate nodes in the initial candidate node set using an adaptive multi-objective scoring function, and filter out the target deployment nodes. The verification module is used to perform lightweight stability verification on the target deployment node before final deployment, so as to realize resource management of the cloud computing platform.

[0019] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a resource management method for cloud computing platforms based on Data Processing Units (DPUs). First, by combining static resource data provided by cluster monitoring with dynamic micro-indicators collected by the DPU, resource availability assessment and stability filtering are performed on all computing nodes to construct an initial candidate node set. Then, an adaptive multi-objective scoring function is used to comprehensively rank and select the target deployment nodes. Finally, a lightweight stability verification is performed on these nodes before final deployment. The dynamic indicator collection by the DPU provides real-time and accurate node resource status, compensating for the shortcomings of traditional static data and improving the comprehensiveness and timeliness of the assessment. The adaptive multi-objective scoring function integrates multiple optimization objectives to achieve intelligent ranking, avoiding biases from manual intervention. Lightweight verification ensures real-time node availability through rapid pre-checks, preventing deployment failures. This method optimizes resource utilization efficiency, significantly reduces system resource consumption and startup latency, reduces operational complexity and resource fragmentation risks, improves load balancing and node stability, and enhances the standardization of network interface configuration and the performance of virtual hard disk technology. It ensures efficient support for the dynamic needs of large-scale application deployments and complex business scenarios while maintaining isolation and security.

[0020] Preferably, in this invention, the initial screening process involves: firstly, rigidly eliminating nodes with insufficient resources based on task requirements to quickly narrow down the scope; then, combining real-time DPU metrics to deeply evaluate node stability and filter out nodes with significant fluctuations or potential instability risks. This step effectively identifies and avoids nodes that, while meeting static resource requirements, may cause task failure due to performance fluctuations or sudden loads during actual operation. This ensures that the initial candidate node set not only has sufficient resources but also possesses the immediate stability required for operation, laying a solid foundation for subsequent accurate screening and reducing ineffective scheduling attempts.

[0021] Preferably, in this invention, the multi-dimensional resource availability (such as CPU, memory, disk, network, etc.) of nodes is uniformly quantified and expressed by defining resource state vectors and dynamic weight vectors. The dynamic weight vectors flexibly adjust the importance weights of various resources according to specific task requirements and system strategies. This abstract modeling makes the comparison of resource states between nodes standardized and computable, providing a structured input data foundation for the subsequent adaptive scoring function and supporting flexible strategy adjustments to adapt to different scenarios.

[0022] Preferably, in this invention, the adaptive scoring function is the core decision-making mechanism, and its formula design integrates multiple key objectives: resource availability is the foundation; short-term volatility collected by the DPU is introduced to directly assess the real-time stability risk of nodes; energy consumption indicators based on the DPU energy efficiency model are added to optimize platform energy efficiency; and the scheduling penalty term utilizes historical failure or migration records to avoid risky nodes. Stability, energy consumption, and risk preference are flexibly balanced through adjustment factors. This function intelligently generates a score reflecting the comprehensive value of nodes, ensuring that the selected target nodes not only match in terms of resources but also perform excellently in terms of stability, energy efficiency, and past reliability, achieving multi-objective joint optimization.

[0023] Preferably, in this invention, the sorting and filtering rules include sorting candidate nodes from high to low based on the scoring function results, and prioritizing the selection of the node with the highest score as the deployment target. This sorting mechanism based on quantitative scoring replaces subjective or simple rules (such as random or round-robin), ensuring that each scheduling decision tends to select the currently optimal node, maximizing the global benefits of resource allocation and the success rate of task deployment.

[0024] Preferably, in this invention, when nodes with similar scores are difficult to distinguish in terms of quality, a secondary objective (such as current load, service availability, or scheduling fairness) is introduced for a second comparison. This avoids local suboptimal selection or resource bias that may result from a single score, and can further select target nodes that better meet the overall health of the cluster (such as load balancing), service continuity requirements, or fair scheduling principles from among nodes with similar scores, thereby improving the overall efficiency and resilience of the cluster.

[0025] Preferably, in this invention, the lightweight stability verification includes three key checks: resource trend reassessment, scheduling congestion assessment, and historical high-risk assessment. Resource trend reassessment utilizes the latest DPU data to predict potential deterioration risks; scheduling congestion assessment identifies short-term resource contention overload on nodes; and historical high-risk assessment avoids frequently failing nodes based on recent records. These three layers of verification form a rapid and effective real-time risk assessment and interception network before final deployment, specifically capturing instantaneous risks or historical hidden dangers that are overlooked by traditional methods and emerge after scoring. This significantly reduces the probability of deploying tasks to nodes that are about to become unstable or have a high failure risk, thus significantly improving deployment success rate and operational stability.

[0026] Preferably, in this invention, the processing flow for verification failure includes: if the primary target node fails verification, it is automatically abandoned and a secondary node is selected for re-verification. This processing method ensures that the scheduling process can quickly and automatically degrade to the secondary solution when the primary node fails, avoiding scheduling stagnation or blind retries, maintaining the continuity and efficiency of the scheduling system, and reducing the overall scheduling delay caused by temporary problems of a single node.

[0027] Preferably, this invention provides two fallback strategies for the special case of multiple consecutive verification failures: pausing scheduling and retrying after resource conditions improve, suitable for temporary cluster overload; or activating automatic migration logic to free up resources by migrating low-priority tasks, or selecting the most reliable failed node for deployment under controllable risk conditions. These strategies effectively address scenarios of extreme cluster resource shortages or partial failures, striving to find feasible deployment solutions even in extreme situations, avoiding prolonged task blocking, enhancing the system's fault tolerance and resilience, and ensuring that critical services have the opportunity to obtain resources. Attached Figure Description

[0028] Figure 1 These are the three core stages of the container load balancing scoring mechanism provided in this embodiment of the invention; Figure 2 A flowchart illustrating a cloud computing platform resource management method based on a DPU, provided as an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a cloud computing platform resource management system based on DPU, provided for an embodiment of the present invention. Detailed Implementation

[0029] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0030] The technical terms used in this embodiment are explained below: DPU, short for Data Processing Unit, is a processor unit specifically designed to process data. In modern computing, massive amounts of data need to be processed, and the DPU plays a crucial role in this process. Compared to traditional central processing units (CPUs) and graphics processing units (GPUs), DPUs possess high data parallelism and optimization capabilities, enabling them to more effectively handle and accelerate various data-intensive tasks. In virtualization environments, placing swap space on the DPU can effectively improve the performance and reliability of swap operations, and it can migrate with virtual machines, thus meeting the swap space requirements of virtualization environments. Therefore, this invention aims to explore how to optimize container swap space in virtualization environments by utilizing the offloading function of the DPU.

[0031] The following is a detailed explanation of the resource management method for a cloud computing platform based on DPU provided in this embodiment, with reference to the accompanying drawings: like Figure 2 As shown, this embodiment provides a cloud computing platform resource management method based on DPU, including: Based on the static resource data provided by the cluster monitoring system and the dynamic micro-indicators collected by the DPU, resource availability assessment and stability filtering are performed on all computing nodes to construct a preliminary candidate node set. An adaptive multi-objective scoring function is used to comprehensively rank the candidate nodes in the initial candidate node set and select the target deployment node. Before the target deployment node is finally deployed, a lightweight stability check is performed on the target deployment node to enable resource management of the cloud computing platform.

[0032] For example, this embodiment provides a cloud computing platform resource management method based on DPU, which introduces an enhanced AMBR (Adaptive Multi-Resource Balancing and Adjustment) scheduling strategy. This strategy, through hardware and software collaboration with the DPU (Data Processing Unit), can make dynamic decisions based on micro-resource information (such as CPU / GPU fluctuation trends, network congestion index, etc.) collected in real time by the DPU. This AMBR scheduling strategy also combines Prometheus' multi-level indicators for joint monitoring.

[0033] The AMBR (Adaptive Multi-resource Balance and Regulation) algorithm is designed as a DPU-driven intelligent container load balancing algorithm. By deeply integrating with the real-time, fine-grained data acquisition capabilities of the Data Processing Unit (DPU), it aims to build a highly sensitive monitoring and response mechanism for the dynamics of Kubernetes cluster resources. This algorithm utilizes dynamic resource weight adjustment, multi-dimensional resource modeling, and real-time feedback from the DPU to overcome the shortcomings of traditional scheduling strategies in perceiving dynamic changes in node resources and achieving load balancing. Through these designs, the AMBR algorithm aims to achieve efficient adaptive scheduling and stable allocation of cluster resources under the DPUKube framework, thereby improving the system's load balancing capabilities and overall performance.

[0034] like Figure 1As shown, this embodiment provides a resource management method for a cloud computing platform based on DPU, applied to the AMBR scheduler. The AMBR scheduler employs a phased processing strategy, with the overall process divided into three key stages: pre-selection, optimization, and refinement. In the pre-selection stage, the system integrates static resource data provided by Prometheus with dynamic micro-indicators collected by the DPU to perform resource availability assessment and stability filtering on all computing nodes, constructing a preliminary candidate node set. In the optimization stage, the scheduler uses an adaptive multi-objective scoring function to comprehensively rank the candidate nodes, considering multiple dimensions such as remaining resource quantity, load fluctuation trends, energy efficiency, and historical scheduling performance, ensuring the selection of the target node with the highest overall scheduling value. In the refinement stage, the scheduling system further performs lightweight stability verification on the target nodes before deployment, effectively avoiding deployment failure or service jitter risks by detecting resource trend changes and scheduling congestion status.

[0035] In this embodiment, AMBR introduces an adaptive multi-objective scoring mechanism in the optimization stage, which can dynamically adjust resource weights according to the actual resource requests of the task and the cluster status, and integrate unstructured indicators such as energy efficiency and scheduling history feedback to achieve more intelligent comprehensive node ranking.

[0036] In this embodiment, AMBR constructs a resource state vector for each candidate node i.

[0037] in, This indicates the availability of the node on resource type j (such as CPU free time, free memory, disk IOPS, etc.). Simultaneously, the scheduler determines a dynamic weight vector for each resource type based on task resource requirements and system operation strategies. This vector supports real-time adjustment under different task or load strategies.

[0038] Based on this, the comprehensive scoring function is defined as shown in formula (1): (1) In the formula: Indicates the first Resource class adaptive weights; Represents a node Then Availability of similar resources; This represents the short-term volatility of resources collected by the DPU, reflecting stability; This represents the energy consumption index under unit resource scheduling, derived from the DPU energy efficiency model; This represents the scheduling penalty for the node, reflecting its historical deployment failure rate or the number of task migrations. , , These represent different adjustment factors, used to balance resource stability, energy efficiency, and historical risk, respectively.

[0039] After node scoring is completed, the scheduler sorts candidate nodes from highest to lowest score and prioritizes the node with the highest score as the target deployment node for the task. If multiple nodes with similar scores are found, the system will further compare secondary indicators such as current node load, service availability, and scheduling fairness to avoid single-node scheduling concentration or system load skew.

[0040] The optimization phase identifies the target node with the highest score. However, in a dynamically changing cluster environment, node states exhibit significant timeliness and volatility. If the target node is experiencing increased resource load, concurrent scheduling congestion, or unstable micro-resources during scheduling execution, it may lead to container deployment failures, frequent task migrations, or even service jitter. Therefore, the AMBR scheduler introduces a lean verification phase before final deployment, performing lightweight stability checks and risk avoidance strategies on the target node to ensure that the scheduling results have higher robustness in real-world systems.

[0041] The core mechanisms of the optimization phase include the following three aspects: (1) Resource trend re-judgment mechanism: The scheduler analyzes the resource trend changes of the target node in the recent time window based on the high-frequency micro resource data collected by the DPU. For example, if the GPU temperature is detected to be rising rapidly, the CPU load slope is continuously positive, or the network traffic surges abnormally, it indicates that the node may be about to enter a high-pressure state. Once such a trend triggers the threshold, the system will mark the node as "unstable in the future" and suspend task deployment. (2) Scheduling congestion avoidance mechanism: The system tracks the scheduling activity of all nodes in each scheduling cycle. If the target node has been assigned multiple high-load tasks in the current cycle or is in a resource competition intensive stage, it is judged that the node has scheduling congestion risk. At this time, in order to avoid short-term resource overload and task squeezing, AMBR will prioritize the second highest-scoring but more stable alternative node and implement rollback scheduling. (3) Feedback-driven historical fault tolerance mechanism: The scheduler refers to the scheduling history of the node. If a node has recently experienced multiple Pod deployment failures, frequent restarts, or resource bottleneck events, its "scheduling credibility" will be reduced by the system. When scores are close, the scheduler will proactively avoid such "high-risk historical nodes" to improve the consistency and stability of scheduling results.

[0042] If the verification passes, the system will perform the final deployment operation, scheduling the task to the target node. If a high-risk factor is detected, an automatic rollback strategy will be triggered, selecting the next scoring candidate node and re-executing the verification logic. If several consecutive nodes have deployment risks, the task scheduling may be delayed or the automatic scheduling migration logic may be triggered.

[0043] To achieve the engineering deployment and practical performance verification of the AMBR scheduling strategy on the Kubernetes platform, this embodiment utilizes the Kubernetes official Scheduling Framework plugin mechanism to modularly integrate the AMBR algorithm, enabling customized control of critical paths such as "filtering—scoring—reservation—binding" in the scheduling process. Simultaneously, to achieve hardware-software awareness and collaboration, the system designs a data access mechanism that collaborates with Prometheus and the DPU, allowing the scheduler to acquire resource status in real time during execution and make adaptive decisions accordingly.

[0044] To achieve a hardware-software collaborative sensing mechanism, the system also designed the following data path interface scheme: Prometheus Metrics Collection: The AMBR plugin queries static metrics of nodes within a specified time window by calling the Prometheus HTTP API, which is suitable for information such as resource utilization, queue length, and scheduling latency; DPU Data Injection Module: Deploys a lightweight DPU Agent on each node side to collect micro-resource status in real time and write the results to a local temporary status cache, which is then read by the scheduling plugin via gRPC or local calls; Asynchronous data caching mechanism: To avoid the impact of frequent calls to external interfaces on performance during scheduling, the AMBR scheduler internally builds a time-sliding window-based index caching system to achieve a dynamic sensing structure with high-frequency reads and low-frequency updates.

[0045] Through the aforementioned plugin mechanism and data integration path, the AMBR scheduling framework achieves non-intrusive integration with the native Kubernetes scheduling system.

[0046] To further clarify the specific execution process of the AMBR scheduling algorithm, this embodiment presents a cloud computing platform resource management method based on DPU, namely the pseudocode description of the AMBR algorithm, as shown in Table 1.

[0047] Table 1 shows the AMBR (Adaptive Container Load Balancing) algorithm.

[0048] For example, this embodiment also provides the awareness requirements of the AMBR scheduling architecture, that is, it provides a multi-level monitoring system that integrates Prometheus and DPU, covering infrastructure, control components, container runtime and application services, etc., to realize the resource monitoring capability of hardware and software collaboration, and provide high-quality input data for subsequent scheduling processes.

[0049] First, a multi-level resource-aware architecture.

[0050] The Kubernetes monitoring system is mainly divided into the following four perception layers: Infrastructure layer: Monitor the core resource status of physical hosts or virtual nodes, including CPU load, memory usage, disk I / O and network bandwidth, as the basis for the scheduler to identify node resource bottlenecks; Control plane component layer: Focus on the performance metrics of key components such as kube-apiserver, etcd, and kube-scheduler, such as API response time, scheduling latency, and configuration consistency, to determine the robustness of the scheduling chain; Container and Pod Layer: Collects fine-grained metrics such as container running status, resource consumption, and abnormal events to support container-level load analysis and deployment strategy formulation; Business service layer: Assess the operational status from the perspective of service call quality (such as latency, error rate, and throughput) to assist the system in making QoS assurance and SLA compliance judgments.

[0051] The data from these levels are aggregated into the scheduling and perception center through a unified monitoring channel, providing a basis for AMBR's decision-making.

[0052] Second, the collaborative sensing mechanism between Prometheus and DPU.

[0053] To balance the comprehensiveness and timeliness of the indicators, this system adopts a resource awareness strategy that combines hardware and software: Prometheus: As a mainstream open-source monitoring solution for Kubernetes, it is primarily responsible for collecting low- to medium-frequency metrics and is suitable for static resource assessment and trend analysis. Through its automatic service discovery mechanism, it can cover objects such as nodes, containers, services, and the control plane, and extends the collection scope through various Exporters.

[0054] DPU (Data Processing Unit): As a hardware-side auxiliary module, it is deployed on each node to collect micro-resource status changes in a high-frequency manner, such as CPU load slope, GPU temperature trend, network burst traffic, etc., which is suitable for identifying sudden bottlenecks and predicting scheduling risks.

[0055] Both integrate the indicator streams through the scheduler's data fusion engine to form multi-dimensional information, including resource sufficiency, volatility, stability, and predictability, which is used to drive scheduling decisions.

[0056] Third, the path for indicator collection and data visualization.

[0057] In actual deployment, Prometheus achieves closed-loop monitoring through the following components: Node Exporter: collects host-level resources; cAdvisor: collects container-level metrics; kube-state-metrics: provides K8s object state data; business metrics client (such as Python / Go SDK): embeds applications to collect service-level performance.

[0058] The collected data is aggregated using PromQL and then visualized using Grafana, supporting fault tracing and trend insights. Meanwhile, the data collected by the DPU is pushed to the scheduling awareness module in real time via a lightweight agent, avoiding delays and forming a mechanism for "fusion and awareness of hot and cold metrics".

[0059] Fourth, the evolution of perception targets and monitoring for scheduling.

[0060] To meet the real-time and accuracy requirements of the AMBR scheduling strategy, the system proposes the following objectives for the monitoring module: improve the frequency of indicator refresh and the ability to predict resource heat; optimize the monitoring response to sudden anomalies (such as resource fluctuations and local network congestion); support hierarchical data collection and adaptive sampling to avoid monitoring itself becoming a burden on the system; and integrate AI anomaly detection and root cause analysis to help the scheduler avoid potential risk nodes.

[0061] Therefore, the cloud computing platform resource management system based on DPU provided in this embodiment has the following advantages: The AMBR scheduling strategy outperforms the native Kubernetes scheduler in several key metrics, including: significantly improving Pod startup success rate and reducing scheduling failure rate; increasing the overall utilization of CPU, GPU, and other resources; shortening Pod startup time; and effectively mitigating the problem of uneven resource distribution among nodes. The collaborative optimization of various factors in the scoring mechanism, along with the DPU-driven risk avoidance mechanism, jointly support the improved stability of the scheduler.

[0062] With the support of hardware and software collaborative perception, multi-objective modeling and scheduling feedback mechanisms, the AMBR algorithm effectively improves the resource matching efficiency, system stability and operating performance of Kubernetes clusters, providing a theoretical basis and engineering feasibility for building intelligent scheduling systems for complex load environments.

[0063] like Figure 3As shown, this embodiment also provides a cloud computing platform resource management system based on DPU, including: a node set construction module, used to perform resource availability assessment and stability filtering on all computing nodes based on static resource data provided by the cluster monitoring system and dynamic micro-indicators collected by DPU, and construct a preliminary candidate node set; a screening module, used to use an adaptive multi-objective scoring function to comprehensively sort the candidate nodes in the preliminary candidate node set and screen out the target deployment node; and a verification module, used to perform lightweight stability verification on the target deployment node before final deployment, so as to realize resource management of the cloud computing platform.

[0064] The present invention also provides a DPU-based cloud computing platform resource management device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the DPU-based cloud computing platform resource management method.

[0065] When the processor executes the computer program, it implements the above-mentioned steps for resource management of the cloud computing platform based on DPU, such as: based on the static resource data provided by the cluster monitoring system and the dynamic micro-indicators collected by the DPU, performing resource availability assessment and stability filtering on all computing nodes to construct a preliminary candidate node set; using an adaptive multi-objective scoring function to comprehensively rank the candidate nodes in the preliminary candidate node set to select the target deployment node; and before the final deployment of the target deployment node, performing a lightweight stability verification on the target deployment node to realize resource management of the cloud computing platform.

[0066] Exemplarily, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing preset functions, the instruction segments describing the execution process of the computer program in the DPU-based cloud computing platform resource management device. For example, the computer program can be divided into a node set construction module, a filtering module, and a verification module; the specific functions of each module are as follows: the node set construction module is used to perform resource availability assessment and stability filtering on all computing nodes based on static resource data provided by the cluster monitoring system and dynamic micro-indicators collected by the DPU, constructing a preliminary candidate node set; the filtering module is used to use an adaptive multi-objective scoring function to comprehensively rank the candidate nodes in the preliminary candidate node set, filtering to obtain the target deployment node; the verification module is used to perform lightweight stability verification on the target deployment node before final deployment, to achieve resource management of the cloud computing platform.

[0067] The DPU-based cloud computing platform resource management device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The DPU-based cloud computing platform resource management device may include, but is not limited to, processors and memory. Those skilled in the art will understand that the above are examples of DPU-based cloud computing platform resource management devices and do not constitute a limitation on DPU-based cloud computing platform resource management devices. It may include more components than described above, or combine certain components, or different components. For example, the DPU-based cloud computing platform resource management device may also include input / output devices, network access devices, buses, etc.

[0068] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or any conventional processor. This processor is the control center of the DPU-based cloud computing platform resource management system, connecting various parts of the DPU-based cloud computing platform resource management device via various interfaces and lines.

[0069] The memory can be used to store the computer program and / or modules. The processor implements various functions of the DPU-based cloud computing platform resource management device by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory.

[0070] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function (such as sound playback, image playback, etc.). The data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD cards), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0071] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the DPU-based cloud computing platform resource management method.

[0072] If the modules / units integrated by the DPU-based cloud computing platform resource management system are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0073] Based on this understanding, the present invention can implement all or part of the processes in the above-mentioned DPU-based cloud computing platform resource management method, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-mentioned DPU-based cloud computing platform resource management method. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or a preset intermediate form, etc.

[0074] The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0075] It should be noted that the content contained in the computer-readable storage medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable storage medium does not include electrical carrier signals and telecommunication signals.

[0076] This invention provides a cloud computing platform resource management method based on DPU, which has the following advantages compared with existing scheduling and resource management methods: This invention performs initial node screening by integrating static cluster resource data with real-time, fine-grained metrics collected by the Data Processing Unit (DPU), constructing a candidate set of nodes with sufficient resources and meeting stability standards. Then, an adaptive multi-objective scoring function (integrating resource availability, real-time volatility, energy efficiency metrics, and historical risk penalties) is used to intelligently rank the candidate nodes, prioritizing the target node with the best overall score. Finally, a lightweight real-time verification is implemented before final deployment (covering three lines of defense: resource trend deterioration warning, short-term scheduling congestion detection, and historical high-frequency fault backtracking). The fine-grained real-time monitoring data provided by the DPU compensates for the lag in traditional static monitoring, making resource assessment more accurate and comprehensive. The multi-objective scoring model intelligently balances multiple dimensions such as resource utilization, node stability, energy efficiency, and historical reliability through dynamic weights and adjustment factors. The lightweight verification intercepts instantaneous risks and potential failure points that arise after scoring in the final decision-making stage. This approach significantly reduces resource consumption and task startup latency on cloud computing platforms, effectively alleviates resource fragmentation and uneven load distribution, and improves virtual hard disk performance and network configuration standardization. While ensuring application isolation and security, it greatly enhances resource utilization efficiency, deployment success rate, and service stability, enabling the platform to efficiently support the elastic needs of complex business scenarios and the agile deployment of large-scale applications.

[0077] The above embodiments are merely one of the implementation methods for achieving the technical solution of the present invention. The scope of protection claimed by the present invention is not limited to this embodiment, but also includes any variations, substitutions and other implementation methods that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A resource management method for a cloud computing platform based on DPU, characterized in that, include: Based on the static resource data provided by the cluster monitoring system and the dynamic micro-indicators collected by the DPU, resource availability assessment and stability filtering are performed on all computing nodes to construct a preliminary candidate node set. An adaptive multi-objective scoring function is used to comprehensively rank the candidate nodes in the initial candidate node set and select the target deployment node. Before the target deployment node is finally deployed, a lightweight stability check is performed on the target deployment node to enable resource management of the cloud computing platform.

2. The resource management method for a cloud computing platform based on DPU according to claim 1, characterized in that, Based on the static resource data provided by the cluster monitoring system and the dynamic micro-indicators collected by the DPU, resource availability assessment and stability filtering are performed on all computing nodes to construct a preliminary candidate node set, including: Based on the resource requirements declared by the scheduled tasks in the static resource data provided by the cluster monitoring system, computing nodes whose remaining resources are clearly insufficient to run the task are initially eliminated. By combining the dynamic micro-indicators collected by the DPU, the real-time stability of the remaining computing nodes is evaluated and eliminated, and computing nodes that exhibit drastic resource fluctuations or stability below a preset threshold are filtered out at the current time or within a preset time. The computing nodes that pass resource availability assessment and stability filtering are used to construct an initial candidate node set.

3. The resource management method for a cloud computing platform based on DPU according to claim 2, characterized in that, The resource state vectors of the candidate nodes in the initial candidate node set are specifically expressed as follows: In the formula, This indicates the availability of a candidate node on resource type j, where j is 1 to n; The dynamic weight vector for each type of resource is obtained based on task resource requirements and system operation strategies, as follows: In the formula, This represents the dynamic weight vector for each type of resource.

4. The resource management method for a cloud computing platform based on DPU according to claim 1, characterized in that, The adaptive multi-objective scoring function is used to comprehensively rank the candidate nodes in the initial candidate node set and filter out the target deployment nodes. The specific formula of the adaptive multi-objective scoring function is as follows: In the formula, Indicates the first Resource class adaptive weights; Represents a node Then Availability of similar resources; This represents the short-term volatility of resources collected by the DPU, used to reflect stability. This represents the energy consumption index under unit resource scheduling, derived from the DPU energy efficiency model; This indicates the scheduling penalty for the node, reflecting its historical deployment failure rate or the number of task migrations. , , These represent different adjustment factors, used to balance resource stability, energy efficiency, and historical risk, respectively.

5. The resource management method for a cloud computing platform based on DPU according to claim 1, characterized in that, The process employs an adaptive multi-objective scoring function to comprehensively rank the candidate nodes in the initial candidate node set and filter out the target deployment nodes, including: An adaptive multi-objective scoring function is used to score the candidate nodes in the initial candidate node set; The candidate nodes are sorted from highest to lowest score, and the candidate node with the highest score is selected as the target deployment node for the task.

6. The resource management method for a cloud computing platform based on DPU according to claim 5, characterized in that, When multiple candidate nodes with similar scores appear, a secondary comparison is performed on these candidate nodes based on secondary objectives to obtain the target deployment node; the secondary objectives include current load, service availability, and scheduling fairness.

7. The resource management method for a cloud computing platform based on DPU according to claim 1, characterized in that, Before final deployment on the target deployment node, a lightweight stability check is performed on the target deployment node to achieve resource management of the cloud computing platform, including: Before the final deployment of the target deployment node, a lightweight stability check is performed on the target deployment node; wherein, the lightweight stability check includes resource trend re-determination, scheduling congestion determination, and historical high-risk determination; The re-determination of resource trends includes: Based on the latest micro-resource data collected by the target node's DPU, the real-time changing trends of key resources are analyzed. If any adverse trend of any resource is detected to exceed the preset safety threshold, it is determined that the target deployment node has a future instability risk; otherwise, the target deployment node is determined to be normal. The scheduling congestion determination includes: Track the allocation of new tasks to all nodes within the current scheduling period. If it is detected that the preferred target node has been densely allocated multiple high-load tasks within the current period, or that resources are in a state of intense competition, it is determined that the target deployment node has a short-term scheduling congestion or resource contention overload risk; otherwise, the target deployment node is determined to be normal. The historical high-risk determination includes: Query the recent scheduling history of the target node. If the target deployment node has experienced multiple deployment failures, frequent task restarts, or task eviction events due to resource bottlenecks within a preset time period, the target deployment node is determined to be a historical high-risk node; otherwise, the target deployment node is determined to be a normal node.

8. The resource management method for a cloud computing platform based on DPU according to claim 1, characterized in that, Before the final deployment of the target deployment node, a lightweight stability check is performed on the target deployment node to achieve resource management of the cloud computing platform. This also includes: If the target deployment node passes the lightweight stability check, then the task will be deployed to the target deployment node. If the target deployment node fails the lightweight stability check, the current target deployment node is abandoned, a suboptimal node is selected, and the lightweight stability check is performed again on the suboptimal node.

9. A resource management method for a cloud computing platform based on DPU according to claim 8, characterized in that, If the number of lightweight stability verification failures exceeds the preset number of consecutive failures, the following actions will be taken: The task scheduling operation is temporarily suspended, and the entire scheduling process is retried from the pre-selection stage after a period of time. Alternatively, trigger the automatic scheduling migration logic: attempt to migrate the existing low-priority or migrateable tasks on the target node to free up resources for the target task; Alternatively, after weighing the risks, choose the node with the lowest risk among the previously failed verification nodes for deployment.

10. A cloud computing platform resource management system based on DPU, characterized in that, include: The node set construction module is used to perform resource availability assessment and stability filtering on all computing nodes based on static resource data provided by the cluster monitoring system and dynamic micro-indicators collected by the DPU, and to construct an initial candidate node set. The filtering module is used to comprehensively sort the candidate nodes in the initial candidate node set using an adaptive multi-objective scoring function, and filter out the target deployment nodes. The verification module is used to perform lightweight stability verification on the target deployment node before final deployment, so as to realize resource management of the cloud computing platform.

Images