A multi-tenant agent collaborative system and method for optimizing cloud server resources

By using a collaborative optimization system of global agents and tenant agents, the problems of low resource utilization and information asymmetry in cloud computing are solved, enabling resource optimization decisions on the tenant side, improving resource utilization and data privacy and security, and adapting to large-scale cloud environments.

CN123123430APending Publication Date: 2026-11-17SHENZHEN POLYTECHNIC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611311012.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-27
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The problems of low resource utilization and difficulty in meeting the business needs of tenants in cloud computing are mainly due to resource optimization bottlenecks caused by information asymmetry and separation of decision-making power.

Method used

A multi-tenant agent collaborative system is constructed. Through the collaborative optimization system of global agent and tenant agent, a decentralized decision-making mechanism is realized. Tenant agent makes independent decisions based on local and global states and outputs resource optimization suggestions.

Benefits of technology

It improves the overall resource utilization of cloud server clusters, ensures the quality of tenant services, achieves data privacy and security on the tenant side, and supports horizontal scaling in large-scale cloud environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123430A_ABST
    Figure CN123123430A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-tenant agent collaborative system and method for cloud server resource optimization, belonging to the field of cloud computing resource optimization technology. It includes at least one global agent deployed in the cloud. Each global agent is used to obtain the global resource status of a corresponding cloud server cluster and broadcast the global resource status to multiple tenant agents within the group. Multiple tenant agents are deployed on the tenant side. Each tenant agent is used to obtain the local status information of its corresponding tenant, perform decentralized independent decision-making based on the local status information and the received global resource status, and output resource optimization suggestions. Each global agent manages its own set of tenant agents. The sets of tenant agents corresponding to different global agents do not overlap. Each tenant agent within a set only receives the global resource status issued by its parent global agent and carries out its own tenant's resource optimization decisions. The system maximizes the overall resource utilization of the cluster through a group coordination mechanism, achieving a balance between individual benefits and group efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a multi-tenant agent collaborative system and method for optimizing cloud server resources, belonging to the field of cloud computing resource optimization technology. Background Technology

[0002] As a pay-as-you-go infrastructure service model, cloud computing's inherent business nature dictates that resource utilization efficiency is key to a cloud service provider's core competitiveness. According to data from the China Academy of Information and Communications Technology's "2025 Cloud Computing Research White Paper," the global cloud computing market is projected to exceed 3 trillion RMB in 2025, with hardware resource investment accounting for 60-70% of data center operating costs. Research indicates that the average utilization rate of global cloud computing resources is only 20-30%, implying significant room for optimization. Every percentage point increase in resource utilization will bring substantial cost savings and profit growth. Therefore, maximizing the utilization efficiency of physical resources while ensuring tenant service quality (QoS) has become a core technical challenge continuously explored in the cloud computing field. Optimization of cloud server resources involves multiple dimensions: physical resource layer optimization, virtualization layer optimization, instance specification layer optimization, region / availability zone layer optimization, and time dimension optimization. However, the actual implementation of these optimization goals faces a fundamental contradiction: information asymmetry and separation of decision-making power exist between the decision-makers in resource optimization (the cloud platform's backend scheduling system) and the decision-makers in resource usage (tenants). Cloud platforms have a grasp of the overall resource status, but cannot force tenants to change their choices; tenants understand their own application needs, but lack awareness of the overall resource status. This contradiction constitutes the main bottleneck of existing resource optimization technologies.

[0003] In recent years, with breakthroughs in AI technology, mainstream cloud service providers have begun deploying Cloud Agent technology on the tenant side. According to Gartner's 2025 Cloud Technology Trends report, by 2028, 25% of enterprises will abandon some cloud technologies due to unrealistic expectations, highlighting the urgency of lowering the barriers to cloud service use and improving user experience. To reduce the technical barriers for tenants using cloud services and improve their experience, mainstream cloud service providers have begun deploying Cloud Agent technology on the tenant side. Cloud Agents typically exist in the form of intelligent assistants, chatbots, or embedded wizards, deployed in the cloud service console, command-line tool (CLI), or the tenant's local environment, acting as an intermediary between the tenant and the cloud platform. The core functions of Cloud Agents include: intent understanding and requirement parsing, knowledge base matching and recommendation, conversational interaction and dynamic adjustment, one-click deployment, and automated operation and maintenance.

[0004] Cloud server resource scheduling optimization is a fundamental technology in the field of cloud computing, and existing technologies mainly focus on the following aspects: 1. Centralized resource scheduling system Major cloud service providers deploy centralized resource scheduling and orchestration systems to manage the resource allocation of hundreds of thousands or even millions of physical servers within their data centers. Typical examples of such systems include: 1) OpenStack Nova Scheduler: the scheduling component of the open-source cloud computing platform OpenStack, employing a two-stage scheduling mechanism of filters and weights; 2) Google Borg / Kubernetes Scheduler: Google's internal large-scale cluster management system and the scheduler for its open-source version, Kubernetes; 3) VMware DRS (Distributed Resource Scheduler): the distributed resource scheduler for the VMware virtualization platform, which continuously monitors the resource usage of hosts within the cluster, automatically performs virtual machine hot migration, and achieves load balancing and defragmentation.

[0005] 2. Oversubscription Management Technology To further improve the utilization efficiency of physical resources, cloud service providers generally adopt super-resolution ratio technology, which allows the total amount of resources allocated to virtual instances to exceed the total amount of physical resources, while keeping the total amount of physical resources constant. This leverages the statistical characteristics of tenant actual utilization rates to achieve resource reuse. Existing super-resolution ratio management technologies mainly include: 1) Static super-resolution ratio setting: Pre-setting different super-resolution ratio thresholds based on instance type (e.g., CPU-sensitive, memory-sensitive); 2) Dynamic super-resolution ratio adjustment: Predicting the resource contention risk of physical hosts through machine learning models and dynamically adjusting the super-resolution ratio to achieve a balance between resource utilization and performance assurance; 3) Credit-based scheduling: Setting different super-resolution weights for tenants of different levels, with higher-level tenants enjoying lower super-resolution ratios and higher resource guarantees.

[0006] 3. Inventory Management and Capacity Forecasting Techniques Cloud service providers need to maintain adequate inventory levels for instances of various sizes to avoid both insufficient inventory preventing tenants from creating instances and excessive inventory leading to idle resources. Existing technologies mainly include: safety stock models: setting safety stock thresholds for each instance size based on statistical characteristics of historical demand data; time series forecasting: using time series forecasting models such as ARIMA and Prophet to predict the creation demand for each instance size over a future period; and replenishment strategy optimization: optimizing the timing and quantity of resource replenishment by considering factors such as hardware procurement cycles and new model launch plans.

[0007] In addition to the backend scheduling system, existing technologies have also developed resource recommendation technologies on the tenant side, aiming to help tenants make more reasonable specification choices and indirectly achieve resource optimization. These mainly include: specification recommendations based on historical behavior, guided prompts based on inventory status, and machine learning-driven intelligent recommendations. Summary of the Invention

[0008] The purpose of this invention is to address the aforementioned problems by providing a multi-tenant agent collaborative system and method for optimizing cloud server resources. This system constructs a two-level collaborative optimization system consisting of a global agent and multiple individual tenant agents. Under the premise of ensuring the SLA requirements of each tenant's business, it maximizes the overall resource utilization of the cluster through a group coordination mechanism, thereby achieving a balance between individual benefits and group efficiency.

[0009] The technical solution adopted in this invention is as follows: A multi-tenant agent collaboration system optimized for cloud server resources includes: At least one global agent is deployed in the cloud. Each global agent is used to obtain the global resource status of a corresponding cloud server cluster and broadcast the global resource status to multiple tenant agents in the group. Multiple tenant agents are deployed on the tenant side. Each tenant agent is used to obtain the local status information of the corresponding tenant, make decentralized independent decisions based on the local status information and the received global resource status, and output resource optimization suggestions. Each global agent governs a dedicated set of tenant agents. The sets of tenant agents corresponding to different global agents do not overlap. Within each set, tenant agents only receive the global resource status issued by their respective global agents and make resource optimization decisions for their own tenant.

[0010] The existing centralized scheduling architecture runs all optimization logic on the cloud platform backend. Tenant-side components only handle information display and demand collection, unable to participate in the decision-making process. The actual business perception of tenants is difficult to reflect in the resource optimization stage, easily leading to a mismatch between the optimization strategies output by the backend scheduling and the actual business needs of tenants. This solution uses a cloud-based global agent to collect and broadcast the global resource status of the cluster within the group. The tenant agent obtains the local status information of the corresponding tenant on the tenant side, and performs decentralized independent decision-making based on the local status information and the received global resource status. Furthermore, it outputs fine-grained resource optimization suggestions during interaction with the tenant environment. By decentralizing the decision-making execution to the tenant side, the tenant agent can simultaneously grasp both the tenant's private local status and the global resource status of the cluster. Tenant-side entities are no longer limited to information exchange intermediaries and can directly participate in the resource optimization decision-making process. Cluster resource demands can be broadcast to the tenant agent through the global agent, and tenant's own business demands can be incorporated into the decision-making process through local status information. This allows the overall optimization goals of the cluster and the tenant's business goals to be integrated at the tenant agent's decision-making stage, eliminating the isolation between the optimization system and the tenant. Each tenant agent makes independent decisions based solely on its own locally observed state and the global resource state issued by the global agent, eliminating the need for point-to-point information exchange between tenants and avoiding high-frequency real-time communication between them. The tenant's original business-related data is processed on the tenant side and is not transmitted to other tenant nodes, ensuring tenant data privacy and security from the underlying architecture. Furthermore, when dealing with multiple clusters, the isolated agent architecture supports parallel management of multiple clusters. There is no information exchange between agent groups corresponding to different clusters, and the state broadcasting and decision-making process within a single group will not interfere with other clusters, facilitating horizontal scaling in large-scale cloud environments. Moreover, the global agent can support a continuously growing number of tenant agents by broadcasting state according to groups, enabling the system to adapt to large-scale cloud deployment scenarios and possessing high scalability.

[0011] Compared to the existing resource optimization solutions mentioned in the background technology, this solution changes the traditional cloud resource optimization approach where decision-making power is completely centralized in the cloud backend. It deploys the decision-making unit on the tenant side, enabling tenant-side entities to have global resource awareness capabilities. It opens up the information path from the global cluster status to the tenant's local decision-making, and achieves collaborative resource optimization with tenant participation without large-scale modifications to the existing cloud platform's centralized scheduling components. It overcomes the technical bottleneck of information and decision-making power mismatch between the cloud platform and tenants through architectural repartitioning.

[0012] Also available are: The reward evaluation module is used to provide feedback on the decisions of each tenant agent. It adopts a distributed two-layer reward mechanism, which includes globally shared rewards and individual penalty items.

[0013] Because the global agent is only responsible for broadcasting the global resource status and lacks the ability to enforce the decision-making logic of tenant agents, tenant agents independently output resource optimization suggestions based on local information and the received global status. Simply relying on status broadcasting only achieves information exchange and cannot provide value guidance for decentralized individual decisions. Each tenant agent is prone to continuously biased towards its own local interests in decision-making, making it difficult to spontaneously converge towards the overall optimal direction of the cluster. The reward evaluation module can bridge the mapping relationship between the cluster's global benefits and the individual decisions of tenant agents, allowing the global optimization goals and the tenant's business assurance requirements to be jointly input into the tenant agent's decision-making process. Specifically, the globally shared reward is broadcast and distributed via the global agent, transforming cluster-level resource benefits into positive incentives perceptible to every tenant agent, while individual penalties are tied to each tenant's own situation to provide negative constraints. Without the reward evaluation module, the entire architecture can only distribute the global status to the tenant side. The decentralized decisions of each tenant agent lack a unified goal, and numerous independent individual decisions are prone to conflict, making it difficult to achieve a synergistic effect for the entire cluster. After the reward and evaluation module is overlaid with the original Agent architecture, the information distribution capability and incentive constraint capability complement each other. The cloud does not need to intervene in the tenant's internal decision-making process, which allows a large number of dispersed tenant Agents to spontaneously align with the cluster optimization goals in an autonomous decision-making mode, while maintaining the operational quality of each tenant's own business.

[0014] Existing technologies for multi-tenant resource collaborative optimization mostly rely on centralized systems to set unified constraint thresholds for control. Centralized systems need to collect business operation data from all tenants, which leads to significant data interaction overhead. Furthermore, centralized control is prone to single-point performance bottlenecks. This solution distributes incentive and constraint signals to each tenant agent through broadcasting and individual mapping. The cloud does not need to perform fine-grained forced corrections to the decision results of each tenant, reducing the computational and interaction pressure on cloud components. It can also adapt to multiple isolated global agents and tenant agent cluster architectures, enabling smooth horizontal scaling with the number of tenants.

[0015] Also available are: The policy constraint module is used to introduce a policy diversity regularization term to constrain the policy distribution of each tenant agent, so as to maintain the policy differences among the tenant agents.

[0016] When tenant agents make decentralized, independent decisions based on their local and global resource states, the lack of corresponding constraints can lead to convergence towards the same type of strategy that yields the best returns. A large number of tenant agents will output similar resource optimization suggestions and concentrate on adjusting resource configurations at the same time or on the same physical nodes, creating new resource demand hotspots within the cluster and weakening the optimization benefits of decentralized collaboration. The policy constraint module can impose diverse constraints on the policy space of tenant agents, increasing the differentiation in decisions among different tenant agents and encouraging tenant agents corresponding to different business types to form recommendation strategies that are distinct in both time and space dimensions. Each tenant agent possesses both local business awareness and global resource state awareness. The policy constraint module further guides the dispersed individual decisions on this basis, without disrupting the original global agent broadcast state and tenant agent autonomous decision-making infrastructure, nor interfering with the decision-making details within a single tenant agent. By leveraging differentiated recommendation strategies, resource demand hotspots within the cluster can be effectively dispersed, enabling spontaneous off-peak scheduling at the cluster level. This improves the overall cluster resource load distribution without deploying centralized forced scheduling commands, further enhancing the overall utilization of cluster resources, while retaining the ability of each tenant agent to output optimization suggestions adapted to its own business needs.

[0017] Furthermore, in real-world scenarios involving multi-tenant shared cluster resources, if the strategies of each agent converge (e.g., all choosing to scale up at the same time or all requesting the same type of resources), the reward feedback received by each agent will fluctuate drastically due to resource competition, causing a sharp drop in the signal-to-noise ratio of the reward signal, making it difficult for each agent to stably learn effective strategies. The policy constraint module, by forcibly maintaining decision diversity, naturally disperses the resource needs of each agent in both time and space. Computationally intensive tenant agents develop a preference for scaling up during periods of ample computing power, I / O-intensive tenant agents develop a preference for migrating during periods of idle storage bandwidth, and memory-intensive tenant agents develop a preference for deploying large memory nodes. This off-peak scheduling mechanism significantly reduces the resource competition faced by each agent during actual execution, and the reward feedback received by each agent more accurately reflects the quality of its own decisions rather than being overwhelmed by interference from other agents. Therefore, the reliability and convergence efficiency of the reward signal are significantly improved.

[0018] A multi-tenant agent collaboration method for optimizing cloud server resources includes: The global agent obtains the global resource status of a corresponding cloud server cluster and broadcasts the global resource status to the tenant agents within the same group; Each tenant agent obtains the local state information of the corresponding tenant, makes decentralized independent decisions based on the local state information and the received global resource state, and outputs fine-grained resource optimization suggestions during the interaction with the tenant environment.

[0019] The global agent is only responsible for state distribution and does not output mandatory scheduling instructions. The decision-making and calculation process is carried out on the side of each tenant. It can fully absorb the unique business operation details of the tenant and output fine-grained resource optimization suggestions that can fit the tenant's own business operation characteristics.

[0020] This solution utilizes a cloud-based global agent to collect and broadcast the cluster's global resource status within the group. Tenant agents obtain local status information for their respective tenants and, combining this local information with the received global resource status, make decentralized, independent decisions. Furthermore, it outputs fine-grained resource optimization suggestions during interactions with the tenant environment. By decentralizing decision-making to the tenant side, tenant agents can simultaneously grasp both their private local status and the cluster's global resource status. Tenant entities are no longer limited to information exchange intermediaries but can directly participate in the resource optimization decision-making process. Cluster resource demands can be broadcast to tenant agents via the global agent, while tenant business demands can be incorporated into the decision-making process through local status information. This allows the overall cluster optimization goals and tenant business goals to be integrated at the tenant agent's decision-making stage, eliminating the isolation between the optimization system and the tenant. Each tenant agent makes independent decisions based solely on its own observed local status and the global resource status disseminated by the global agent, eliminating the need for point-to-point information exchange between tenants and avoiding high-frequency real-time communication across tenants. Tenant-related raw business data is processed on the tenant side and is not transmitted to other tenant nodes, ensuring tenant data privacy and security from the underlying architecture. Furthermore, when dealing with multiple clusters, the multi-group, isolated agent architecture supports parallel management of multiple clusters. There is no information exchange between agent groups corresponding to different clusters, and the status broadcasting and decision-making process within a single group will not interfere with other clusters, facilitating horizontal scaling in large-scale cloud environments. Moreover, global agents can support a continuously growing number of tenant agents by completing status broadcasts according to groupings, enabling the system to adapt to large-scale cloud deployment scenarios and possessing high scalability.

[0021] Optionally, each tenant can deploy one or more tenant agents. Tenant agents from different tenants make independent decisions. When two or more tenant agents are deployed within the same tenant, these agents can collaboratively optimize to reach the optimal state as perceived by the user. Since tenant agents from different tenants are independent and do not involve the private data of other tenants, tenant data isolation and privacy are ensured. Multiple tenant agents autonomously collaborate and iterate, fine-tuning parameters within the tenant's closed environment. This allows for continuous refinement of resource configuration schemes from multiple dimensions, gradually converging to the optimal state of tenant resource operation and continuously aligning with the dynamic business operation requirements of the tenant. This collaborative optimization process runs entirely locally within the tenant, without transmitting tenant business data externally. This maintains the architecture's native data privacy protection capabilities and avoids additional cross-node communication overhead.

[0022] Alternatively, a distributed two-tier reward mechanism can be assigned to each tenant agent. The two-tier reward mechanism includes a globally shared reward and an individual penalty. The globally shared reward is determined based on the resource utilization of the cloud server cluster and is broadcast to all tenant agents. The individual penalty is determined based on the tenant's SLA default.

[0023] Globally shared rewards transform the overall cluster resource utilization into a shared feedback signal perceptible to all tenant agents. When the overall cluster resource utilization changes, all tenant agents receive the corresponding reward signal, guiding individual tenant agents to consider the overall cluster-level benefits in their local decision-making process and mitigating the tendency to solely pursue local tenant interests. Individual penalty items generate differentiated constraint signals based on SLA defaults during actual tenant operations, preventing tenant agents from sacrificing service quality by pursuing globally shared rewards. This establishes a balance between cluster resource optimization benefits and tenant service operation assurance. Both reward signals operate in a distributed manner on each tenant agent, eliminating the need for a centralized node to directly control or modify the tenant agent's internal decision-making logic. This maintains compatibility with the existing decentralized decision-making architecture of global agent broadcasting and tenant agent state. Globally shared rewards provide unified feedback to all tenant agents within the group, while individual penalty items are tied to the individual tenant's own service operation results. The two types of signals work together, allowing both overall cluster benefits and individual tenant service quality to simultaneously participate in the tenant agent's decision-making iteration process. Furthermore, this reward mechanism does not require tenants to exchange business data, does not introduce additional cross-tenant communication overhead, and can continue the data privacy protection capabilities and large-scale scalability of the underlying architecture.

[0024] Alternatively, the distributed two-tier reward mechanism uses dynamic weights to weight and sum the globally shared reward and the individual penalty items to construct the individual reward function for each tenant agent. When cluster resources are strained or the load is uneven, the weight of the globally shared reward is increased; when resources are plentiful, the weight of the individual penalty item is increased. Under conditions of cluster resource strain or load imbalance, it proactively helps alleviate cluster load imbalance and resource scarcity, weakens the decision-making weight of single-tenant local interests, and quickly corrects the overall cluster resource operation status. When cluster resources are plentiful and stable, it allows each tenant agent to focus more on the service quality assurance of its own business, strictly avoids various SLA default risks, and finely adapts to the tenant's own business operation requirements. The adaptive adjustment mode of dynamic weights allows the two-tier reward mechanism to overcome the limitations of fixed parameter ratios, enabling the reward and penalty orientation to dynamically change with the cluster resource environment and adapt to the dynamic fluctuations in cloud cluster load. No changes are needed to the core agent decision-making logic and cluster interaction mechanism; the control of the group decision orientation can be completed simply through the dynamic iteration of the reward function. It simultaneously takes into account both the overall cluster resource optimization goal and the individual tenant business stability operation goal, achieving a dynamic balance between cluster load management and tenant SLA assurance.

[0025] Alternatively, an information theory-based policy diversity regularization term can be introduced to constrain the policy distribution of each tenant agent, maintaining policy diversity by maximizing the differences between the policy distributions of different tenant agents. The policy constraint module can impose diversity constraints on the policy space of tenant agents, increasing the decision-making differentiation among different tenant agents and encouraging tenant agents corresponding to different business types to form recommendation strategies that are distinct in both time and space dimensions. Each tenant agent possesses dual capabilities of local business awareness and global resource status awareness. Building upon this, the policy constraint module further guides the dispersed individual decisions with diversity, without disrupting the original global agent broadcast state and tenant agent autonomous decision-making infrastructure, nor interfering with the decision-making details within a single tenant agent. By leveraging differentiated recommendation strategies, resource demand hotspots within the cluster are effectively dispersed, promoting spontaneous peak-shifting scheduling at the cluster level. This improves the overall cluster resource load distribution without deploying centralized forced scheduling commands, further enhancing the overall utilization of cluster resources, while retaining the ability of each tenant agent to output optimization suggestions adapted to its own business needs.

[0026] Alternatively, each tenant agent may share an underlying basic feature perception network to extract general state representations, and each tenant agent may maintain an independent top-level policy output module; the policy diversity regularization term may be applied to the top-level policy output of each tenant agent.

[0027] It can balance the overall consistency of system cognition with the diversity of individual decisions. The underlying shared network improves the training efficiency and environmental adaptability of the multi-tenant agent system, avoids the problems of parameter redundancy and cognitive inconsistency caused by multiple independent perception networks, and the independent top-level policy module with regularization constraints can continuously widen the difference in decision output of different tenant agents, effectively avoiding the problem of convergence of group policies.

[0028] When combined with a reward mechanism, the unified perception at the bottom layer ensures that all tenant agents have a synchronized understanding of the cluster resource status. The differentiated policy constraints at the top layer, combined with the dynamic reward and punishment mechanism, can fully unleash the collaborative potential of multiple distributed agents under the premise of unified cluster optimization goals. This allows the system to have both the efficiency and stability of large-scale cluster deployment, as well as the optimization capabilities of differentiated peak scheduling and resource hotspot dispersion, achieving a two-way benefit of shared empowerment at the bottom layer and differentiated efficiency enhancement at the top layer.

[0029] Alternatively, the cloud server cluster can be divided into multiple optimization dimensions according to regions. Each global agent corresponds to a cloud server cluster in a single region. The optimization boundary of the global agent is limited to that region, and closed-loop optimization is only performed on the cloud servers and their associated resources in that region.

[0030] Cloud service providers have a massive tenant scale and total resource volume across the entire network. Cross-regional networks introduce unavoidable physical latency. If centralized optimization is performed on all cloud resources, the system will face state space dimensionality explosion and communication bottlenecks. The optimization latency of a centralized architecture will increase significantly, algorithm convergence difficulty will rise accordingly, and engineering implementation will incur high costs. This solution adopts a regionalized divide-and-conquer strategy. Each region's corresponding global agent independently completes the state collection, broadcasting, and collaborative optimization process within its group. The scope of optimization computation is constrained within its corresponding region. There is no centralized joint computation between different regions, which reduces the amount of cross-regional data interaction, avoids cross-regional network latency, and ensures the timeliness of the optimization process within a single region. This is consistent with the group management architecture of global agents and tenant agents. Tenant agents in the same region only receive the global resource status broadcast by the global agent in their local region. Agent groups in different regions are isolated from each other, and cross-regional decision coupling is avoided.

[0031] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. The multi-tenant agent collaborative system and method for cloud server resource optimization provided by this invention decentralizes decision-making capabilities to the tenant side. It relies on a global agent to broadcast the global cluster status within a region. Tenant agents combine the global resource status with their local status to conduct decentralized independent decisions, breaking down the inherent bottlenecks of information asymmetry and separation of decision-making power between the cloud platform and tenants. This changes the current situation where tenant-side components only handle information display and demand collection, aligning the overall cluster optimization goals with tenant business needs in the decision-making process. Simultaneously, each tenant agent completes decision-making calculations based solely on local observations, eliminating high-frequency real-time communication across tenants. Tenant private business data remains on the tenant side, ensuring tenant data privacy from the underlying architecture. The system can support horizontal scaling in large-scale tenant scenarios.

[0032] 2. The present invention provides a multi-tenant agent collaborative system and method for optimizing cloud server resources. It sets up a distributed two-layer reward mechanism and introduces dynamic weight adjustment. It transmits cluster resource utilization feedback signals through global shared rewards and constrains tenant SLA default risk with individual penalty items. It can also dynamically adjust the weights of the two types of reward and penalty items according to the cluster resource tension. When the cluster is under pressure, it strengthens the overall benefit orientation of the cluster and focuses on the quality assurance of tenant business when resources are abundant. It achieves multi-tenant agent goal alignment in a decentralized system and takes into account both the improvement of cluster resources and the stability of tenant business operation.

[0033] 3. The present invention provides a multi-tenant agent collaborative system and method for optimizing cloud server resources. It introduces a policy diversity regularization term based on information theory, and combines it with a hierarchical model structure that shares the top-level policy output with the underlying feature perception network. The regularization constraint only applies to the top-level policy output. While maintaining the consistency of environmental perception of all agents, it maximizes the difference in policy distribution among different tenant agents, suppresses the convergence of multi-tenant agent decisions, guides tenants of different business types to form resource optimization suggestions with spatiotemporal differences, effectively disperses cluster resource demand hotspots, and realizes spontaneous off-peak scheduling within a geographical area.

[0034] 4. The multi-tenant agent collaborative system and method for cloud server resource optimization provided by this invention adopts a localized divide-and-conquer closed-loop optimization approach. Each global agent corresponds to a single regional cluster and limits the local region optimization boundary. This avoids the state space explosion and cross-regional communication delay problems caused by centralized optimization of all resources, reduces the algorithm's computational scale, improves optimization timeliness and convergence ability, and reduces the complexity of engineering implementation. The localized divide-and-conquer architecture and agent grouping and isolation mechanism are mutually compatible, fully retaining all the advantages of the original system in terms of privacy protection and decentralized collaboration, and adapting to the actual deployment and operation requirements of ultra-large-scale cloud platforms. Attached Figure Description

[0035] Figure 1 This is an architectural diagram of the present invention. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings.

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0038] A multi-tenant agent collaborative system optimized for cloud server resources, characterized in that it includes: At least one global agent is deployed in the cloud. Each global agent is used to obtain the global resource status of a corresponding cloud server cluster and broadcast the global resource status to multiple tenant agents in the group. Multiple tenant agents are deployed on the tenant side. Each tenant agent is used to obtain the local status information of the corresponding tenant, make decentralized independent decisions based on the local status information and the received global resource status, and output resource optimization suggestions. Each global agent governs a dedicated set of tenant agents. The sets of tenant agents corresponding to different global agents do not overlap. Within each set, tenant agents only receive the global resource status issued by their respective global agents and make resource optimization decisions for their own tenant.

[0039] Specifically, for a single cluster, a global agent is deployed at the upper layer. Therefore, when multiple clusters exist, a separate global agent should be selected for each cluster. Each global agent manages its own exclusive set of tenant agents, and the sets of tenant agents corresponding to different global agents do not overlap. Each tenant agent only corresponds to and receives the cluster's global resource status issued by its assigned global agent, avoiding decision conflicts among multiple global agents and effectively saving cross-cluster communication costs. The global agent is responsible for continuously monitoring the overall cluster resource utilization, load distribution of each availability zone, resource fragmentation, and supply and demand fluctuations, and periodically broadcasts global resource status summaries and guidance signals to each tenant agent through an asynchronous communication mechanism. The asynchronous communication mechanism is only a preferred option; other communication mechanisms can also be used. Each tenant agent is used to finely perceive the local private status of its tenant, such as instance utilization, SLA achievement rate (e.g., latency, throughput), historical operation trajectory, and business load cycle. During interaction with the tenant environment, it outputs fine-grained resource optimization suggestions, including instance specification scaling up / down, idle resource reclamation and release, dynamic migration across availability zones, and elastic scaling.

[0040] As another specific implementation method, it also includes: The reward evaluation module is used to provide feedback on the decisions of each tenant agent. It adopts a distributed two-layer reward mechanism, which includes globally shared rewards and individual penalty items.

[0041] In this embodiment, the reward evaluation module is deployed in the cloud or a centralized evaluation server, communicating with each global agent and tenant agent. The global shared reward is calculated by the global agent based on the improvement in cluster resource utilization, load balancing, and defragmentation benefits, and is broadcast to all tenant agents, thus naturally incentivizing each agent to pursue global resource efficiency while optimizing its own performance. The individual penalty is designed individually for each tenant agent, applying negative feedback when the tenant defaults on its SLA (including exceeding response latency limits, decreased availability, or task queuing overflow). The final individual reward function output by the reward evaluation module is constructed by dynamically weighting and summing the global shared reward and the individual penalty, and is fed back to each tenant agent for policy updates.

[0042] As another specific implementation method, it also includes: The policy constraint module is used to introduce a policy diversity regularization term to constrain the policy distribution of each tenant agent, so as to maintain the policy differences among the tenant agents.

[0043] In this embodiment, the policy constraint module is deployed on the training server or cloud coordination node and connected to the policy network of each tenant agent. To prevent resource request hotspots and homogeneous recommendations caused by policy convergence during multi-agent collaborative training, the policy constraint module explicitly introduces an information theory-based policy diversity regularization term into the training objective. Specifically, all tenant agents share a basic feature perception network at the bottom layer to extract general state representations, but each maintains an independent policy output module at the top layer. The policy diversity regularization term acts on the top-level policy output of each tenant agent, constraining the policy distribution of each agent to maintain differences by maximizing the KL divergence or negative mutual information between the policy distributions of different tenant agents, thus preserving individual decision preferences. This regularization term encourages tenant agents of different business types (such as compute-intensive, I / O-intensive, memory-intensive, and network-intensive) to form differentiated recommendation strategies in time and space, effectively dispersing resource demand hotspots and achieving staggered scheduling at the cluster level. During training, the policy constraint module adds a regularization term to the loss function and controls the regularization strength through hyperparameters. When the policy distributions of each agent are detected to be converging, the regularization constraint is automatically strengthened.

[0044] A multi-tenant agent collaboration method for optimizing cloud server resources includes: The global agent obtains the global resource status of a corresponding cloud server cluster and broadcasts the global resource status to the tenant agents within the same group; Each tenant agent obtains the local state information of the corresponding tenant, makes decentralized independent decisions based on the local state information and the received global resource state, and outputs fine-grained resource optimization suggestions during the interaction with the tenant environment.

[0045] The global resource status acquired by the global agent includes overall cluster resource utilization, load distribution across availability zones, resource fragmentation, and supply-demand fluctuation trends. The global agent periodically broadcasts a global resource status summary and guidance signals to each tenant agent using an asynchronous communication mechanism. The broadcast content is compressed to include only the global macro-level features required for tenant decision-making, without involving other tenants' private data, thus ensuring tenant data isolation and privacy. Upon receiving the global status summary, each tenant agent fuses and encodes the global features with locally observed statuses such as instance utilization, SLA achievement rate, historical operation patterns, and business load cycles to form a complete status representation vector, which is then mapped to resource optimization actions via a policy network. During interaction with the tenant environment, the tenant agent continuously outputs fine-grained resource optimization suggestions through an interactive feedback loop, observing environmental feedback and updating its local policy after each decision.

[0046] As another specific implementation, each tenant deploys one or more tenant agents; the tenant agents of different tenants make independent decisions; when the same tenant deploys two or more tenant agents, the tenant agents of the same tenant can be coordinated and optimized to reach the optimal state for the user.

[0047] In this embodiment, when multiple tenant agents are deployed for the same tenant, each tenant agent can be responsible for optimization decisions for different business types (such as online business, offline batch processing, and different components in the microservice chain) or different resource dimensions (computing resources, storage resources, and network resources) of the tenant.

[0048] As another specific implementation, a distributed two-tier reward mechanism is assigned to each tenant agent. The two-tier reward mechanism includes a globally shared reward and an individual penalty. The globally shared reward is determined based on the resource utilization rate of the cloud server cluster and is broadcast to all tenant agents. The individual penalty is determined based on the tenant's SLA default.

[0049] In this embodiment, the calculation of the globally shared reward in the distributed two-layer reward mechanism is performed by the global agent. Specifically, at the end of each broadcast cycle, the global agent collects the current resource utilization data of the cluster, calculates the utilization improvement compared to the previous cycle, the Gini coefficient or variance of the load distribution of each availability zone, and the amount of schedulable resources released due to fragmentation. The weighted sum of these three factors is used as the globally shared reward value, which is then distributed to all tenant agents within its jurisdiction via the broadcast channel. The individual penalty item is calculated by the tenant agent itself or the tenant-side monitoring module. When the SLA indicators (response latency, throughput, availability, task queue length) of a tenant are detected to exceed the corresponding default threshold, a negative reward value is generated based on the duration and degree of deviation of the default. If multiple SLA defaults occur simultaneously, the penalty items are accumulated or the maximum value is taken as the comprehensive penalty. Each tenant agent receives a comprehensive feedback signal composed of the globally shared reward and the individual penalty item after each decision, and updates the parameters of its own policy network accordingly.

[0050] As another specific implementation, the distributed two-layer reward mechanism uses dynamic weights to perform a weighted summation of the global shared reward and the individual penalty item to construct the individual reward function for each tenant agent; when cluster resources are scarce or the load is skewed, the weight of the global shared reward is increased, and when resources are abundant, the weight of the individual penalty item is increased.

[0051] As another specific implementation, an information theory-based policy diversity regularization term is introduced to constrain the policy distribution of each tenant agent, thereby maintaining policy diversity by maximizing the differences between the policy distributions of different tenant agents.

[0052] In another specific implementation, each tenant agent shares a basic feature perception network to extract general state representations, and each tenant agent maintains an independent top-level policy output module; the policy diversity regularization term applies to the top-level policy output of each tenant agent. Specifically, each tenant agent shares the parameters of a shared neural network; the shared parameters do not require original data, and current training frameworks such as PyTorch can support the sharing and synchronous updating of network parameters.

[0053] As another specific implementation method, the cloud server cluster is divided into multiple optimization dimensions according to the region. Each global agent corresponds to a cloud server cluster in a single region. The optimization boundary of the global agent is limited to the region, and closed-loop optimization is only performed on the cloud servers and their associated resources in the region.

[0054] In this embodiment, the cloud server cluster is divided into multiple optimization dimensions based on geographical location (e.g., region, province, city). Each region independently deploys a global agent, and there is no hierarchical relationship between the global agents; they run in parallel. Each global agent only perceives and manages the resource status of the cloud server cluster within its own region. The global resource status summary and guidance signals it broadcasts only cover the area within that region. The optimization boundary is clearly limited to that region, and cross-regional resources are not uniformly scheduled to avoid cross-regional network latency and data compliance risks. Tenant agents within each region only belong to and receive signals from the global agent of that region. When a tenant's business is deployed across multiple regions, different tenant agent instances correspond to each region. Each instance independently receives signals from the global agent of its own region and makes independent decisions, indirectly coordinating with each other through tenant-level upper-layer business policies.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. The invention extends to any new features or combinations disclosed in this specification, and any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention. It is obvious to those skilled in the art that the invention is not limited to the details of the above exemplary embodiments, and that the detailed technical features not disclosed in these embodiments can be understood by those skilled in the art according to specific circumstances. This disclosure does not specifically limit the embodiments in this regard.

Claims

1. A multi-tenant agent collaborative system optimized for cloud server resources, characterized in that: It includes at least one global agent, deployed in the cloud. Each global agent is used to obtain the global resource status of a corresponding cloud server cluster and broadcast the global resource status to multiple tenant agents in the group. Multiple tenant agents are deployed on the tenant side. Each tenant agent is used to obtain the local status information of the corresponding tenant, make decentralized independent decisions based on the local status information and the received global resource status, and output resource optimization suggestions. Each global agent governs a dedicated set of tenant agents. The sets of tenant agents corresponding to different global agents do not overlap. Within each set, tenant agents only receive the global resource status issued by their respective global agents and make resource optimization decisions for their own tenant.

2. The system as described in claim 1, characterized in that: Also includes: The reward evaluation module is used to provide feedback on the decisions of each tenant agent. It adopts a distributed two-layer reward mechanism, which includes globally shared rewards and individual penalty items.

3. The system as described in claim 1, characterized in that: Also includes: The policy constraint module is used to introduce a policy diversity regularization term to constrain the policy distribution of each tenant agent, so as to maintain the policy differences among the tenant agents.

4. A multi-tenant agent collaboration method for optimizing cloud server resources, characterized in that: Using the system of claim 1, the method includes: The global agent obtains the global resource status of a corresponding cloud server cluster and broadcasts the global resource status to the tenant agents within the same group; Each tenant agent obtains the local state information of the corresponding tenant, makes decentralized independent decisions based on the local state information and the received global resource state, and outputs fine-grained resource optimization suggestions during the interaction with the tenant environment.

5. The method as described in claim 4, characterized in that: Each tenant deploys one or more tenant agents; the tenant agents of different tenants make independent decisions; when the same tenant deploys two or more tenant agents, the tenant agents of the same tenant can be coordinated and optimized to reach the optimal state for the user.

6. The method as described in claim 4, characterized in that: Each tenant agent is assigned a distributed two-tier reward mechanism, which includes a globally shared reward and an individual penalty. The globally shared reward is determined based on the resource utilization of the cloud server cluster and is broadcast to all tenant agents. The individual penalty is determined based on the tenant's SLA default.

7. The method as described in claim 6, characterized in that: The distributed two-layer reward mechanism uses dynamic weights to perform a weighted summation of the global shared reward and the individual penalty item to construct the individual reward function for each tenant agent; when cluster resources are scarce or the load is skewed, the weight of the global shared reward is increased, and when resources are abundant, the weight of the individual penalty item is increased.

8. The method as described in claim 4, characterized in that: An information theory-based policy diversity regularization term is introduced to constrain the policy distribution of each tenant agent, thereby maintaining policy diversity by maximizing the differences between the policy distributions of different tenant agents.

9. The method as described in claim 8, characterized in that: Each tenant agent shares an underlying basic feature perception network to extract general state representations, and each tenant agent maintains an independent top-level policy output module; the policy diversity regularization term is applied to the top-level policy output of each tenant agent.

10. The method as described in claim 4, characterized in that: The cloud server cluster is divided into multiple optimization dimensions according to the region. Each global agent corresponds to a cloud server cluster in a single region. The optimization boundary of the global agent is limited to that region, and closed-loop optimization is only performed on the cloud servers and their associated resources in that region.