A method for optimizing computing power configuration through software design

Abstract

The application discloses a kind of through software design optimization computing power configuration method and system, the configuration method includes: between model management platform and K8s computing power cluster, establish life cycle manager, identity and quota center, dynamic computing power controller, intelligent recycling agent;Model development whole life cycle state is bound with computing power resources, according to model stage transition automatically executes computing power application, rise or fall allocation or recycling;Adopt account-model-resource ternary binding mechanism, realize the association of user identity and model instance, underlying resource declaration;According to model complexity, provide the elastic quota of basic computing power package and extended computing power package, dynamically schedule by software strategy;Based on business activity, interactive state, system resource layer three-dimensional index carries out non-active determination, executes graceful automatic recycling.The extreme resource utilization rate is improved.

Description

Technical Field

[0001] This invention relates to the fields of financial data analysis, model management, and cloud computing resource scheduling, and in particular to a method for optimizing computing power configuration through software design. Background Technology

[0002] In the current digital transformation of the banking industry, machine learning and large-scale models are widely used, and model development generally adopts interactive environments such as Jupyter Notebook. Existing computing power management uses a static allocation and manual start / stop mode, which has the following drawbacks: The decoupling of resources from model lifecycles leads to long-term zombie notebook usage and significant waste of computing power. Recycling only based on connection / idle time may mistakenly kill long-running background tasks. The quota is limited and cannot meet the heavy computing demands of ordinary models and graph neural networks. Resources are tied to containers rather than users, making fine-grained scheduling and cost accounting difficult; Relying on manual release results in slow resource turnover and new tasks waiting in queues.

[0003] Existing open-source garbage collection solutions (such as the simple jupyterhub-idle-culler) typically rely solely on network connectivity or basic kernel activity to determine performance, making it difficult to distinguish between "genuine idle computing" and "long-running background operations." In the complex intranet environment of banks, simple automatic timeout garbage collection can easily mistakenly kill high-value training tasks that are currently running, leading to data loss and reduced trust from developers in the platform. Summary of the Invention

[0004] In view of the above problems, the present invention is proposed to provide a method for optimizing computing power configuration through software design to overcome or at least partially solve the above problems.

[0005] According to one aspect of the present invention, a method for optimizing computing power configuration through software design is provided, the configuration method comprising: Deploy an intelligent scheduling controller between the model management platform and the K8s computing power cluster, and establish a lifecycle manager, identity and quota center, dynamic computing power controller, and intelligent recycling agent; Bind the entire lifecycle status of model development to computing resources, and automatically execute computing resource requests, upgrades, downgrades, or recycling based on model stage transitions; A mandatory binding mechanism using account-model-resource triples is adopted to associate user identity with model instance and underlying resource declaration; Elastic quotas for basic and extended computing power packages are provided based on model complexity and dynamically scheduled through software strategies. Inactivity is determined based on three dimensions: business activities, interaction status, and system resource layer indicators, and graceful automatic recycling is performed.

[0006] Optionally, the entire model development lifecycle includes: data preparation, model building, model validation, model deployment, and model decommissioning; wherein: During the model building phase, a basic computing power package is automatically allocated and a Notebook container is launched. The model validation phase supports elastic task-based scaling. Switch to production-level quotas during the model deployment phase; The model is completely decoupled and its resources are released during the retirement phase.

[0007] Optionally, the account-model-resource triple binding includes: obtaining user accounts through OIDC or LDAP protocol, binding the accounts with model instances and K8s ResourceClaim objects one by one, and realizing resource ownership, usage statistics and roaming scheduling.

[0008] Optionally, the flexible quota mechanism includes: Basic computing power package: Default 8-core CPU, 16GB memory; Extended computing power package: 16-core CPU, 32GB memory or above configuration for graph models and large model fine-tuning; scheduled to the corresponding hardware nodes through K8s DRA dynamic resource allocation and DeviceClass mechanism.

[0009] Optionally, the three-dimensional inactivity determination includes: Business activity layer: Detection of platform page operations, code submissions, and file saving records; Interactive state layer: Monitors Notebook front-end connectivity and kernel activity; System resource layer: Monitors the growth rate of Python kernel CPU time slices to determine whether CPU usage is suspended.

[0010] Optionally, the elegant automatic recycling includes: automatically saving Notebook state and variable snapshots, unregistering K8s resource declarations, releasing computing power to the resource pool, sending recycling notifications to users, and providing a one-click recovery entry.

[0011] Optionally, the smart recycling agent is deployed as a side-vehicle container within the Notebook Pod, continuously reporting activity signals and executing recycling commands.

[0012] Optionally, when the model moves from the construction phase to the verification / retirement phase, the system automatically triggers a reduction or release of computing power without manual intervention.

[0013] Optionally, the system can be configured with a hysteresis threshold and confirmation delay to avoid inactive judgment jitter and frequent recycling / restart.

[0014] This invention also provides a system for optimizing computing power configuration through software design, applying the method for optimizing computing power configuration through software design as described in any one of claims 19, characterized in that it includes: Lifecycle Manager: Maintains the model's phase state and triggers computing power events; Identity and Quota Center: Manages user accounts, computing power wallets, and tiered quota policies; Dynamic computing power controller: Implements resource declaration, scheduling, and allocation / degradation using Kubernetes Operator; Intelligent recycling agent: Performs 3D inactivity detection and graceful recycling.

[0015] This invention provides a method and system for optimizing computing power configuration through software design. The configuration method includes: deploying an intelligent scheduling controller between a model management platform and a Kubernetes computing power cluster; establishing a lifecycle manager, an identity and quota center, a dynamic computing power controller, and an intelligent recycling agent; binding the entire lifecycle status of model development with computing power resources; automatically executing computing power applications, upgrades, downgrades, or recycling based on model stage transitions; adopting a mandatory binding mechanism of account-model-resource triples to associate user identity with model instances and underlying resource declarations; providing elastic quotas for basic and extended computing power packages based on model complexity, and dynamically scheduling them through software strategies; and performing graceful automatic recycling based on three-dimensional indicators of business activities, interaction status, and system resource layer. This results in an extreme improvement in resource utilization.

[0016] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for optimizing computing power configuration through software design, as provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0020] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0022] like Figure 1 As shown, a method for optimizing computing power configuration through software design includes: Deploy an intelligent scheduling controller between the model management platform and the K8s computing power cluster, and establish a lifecycle manager, identity and quota center, dynamic computing power controller, and intelligent recycling agent; Bind the entire lifecycle status of model development to computing resources, and automatically execute computing resource requests, upgrades, downgrades, or recycling based on model stage transitions; A mandatory binding mechanism using account-model-resource triples is adopted to associate user identity with model instance and underlying resource declaration; Elastic quotas for basic and extended computing power packages are provided based on model complexity and dynamically scheduled through software strategies. Inactivity is determined based on three dimensions: business activities, interaction status, and system resource layer indicators, and graceful automatic recycling is performed.

[0023] This invention proposes a method for optimizing computing power configuration through software design, constructing an "intelligent scheduling controller" located between the model management platform and the computing power infrastructure (such as a Kubernetes cluster). This architecture coordinates business phases, user identities, and underlying resource allocation through an event-driven mechanism.

[0024] Logical architecture components, including: Lifecycle Manager: Responsible for maintaining the state changes of the model in MMP, such as the transition from "Data Preparation" to "Model Building".

[0025] Identity and Quota Center: Through account binding logic, determine the upper limit of resources that each builder can schedule and the quota that has been used.

[0026] Dynamic computing power controller: Implemented in K8s Operator mode, it listens for business events and directly operates the underlying computing resources.

[0027] Intelligent recycling agent: Deployed in a side-vehicle container inside the Notebook Pod or monitored via the Hub, continuously reporting multi-dimensional activity signals.

[0028] Computing power binding method based on lifecycle stage The core step of this invention is to change the allocation logic of computing resources from "users can start at will" to "automatic readiness based on business stages".

[0029] Phase-sensitive resource allocation process When builders advance a new model project to the "Build Phase" on the model management platform, the system executes the following automated process: Identity verification: The controller obtains the account identifier of the currently logged-in user through the API and queries the number of computing power instances bound to that user's name.

[0030] Resource declaration: The system automatically creates a ResourceClaim object in Kubernetes.

[0031] According to the default strategy defined in the patent, this object defines a core resource requirement of 8c16g. Using Dynamic Resource Allocation (DRA) technology, the scheduler searches for compute nodes in the cluster that meet this requirement.

[0032] Environment provisioning: While resources are locked, the system automatically mounts the storage volumes (PVs) previously associated by the builders to ensure the consistency of the development environment.

[0033] Cross-stage computing power state transition As the model moves from the construction phase to subsequent phases, the computing power configuration will be automatically adjusted based on the business logic: Account binding and differentiated resource application mechanisms include: This invention not only solves the recycling problem, but also enables specialized support for complex model tasks through account binding.

[0034] Account and quota binding logic: Each developer account corresponds to a "computing power wallet" in the system. The system limits the total amount of each wallet across different projects through software logic. When a user needs to handle computationally intensive tasks such as graph models, they initiate a "specialized computing power application" on the platform.

[0035] Large-resource allocation scheme for graph models: When processing bank social relationship networks and anti-money laundering fund path analysis, the memory overhead of graph neural networks (GNN) far exceeds that of standard machine learning models.

[0036] To address these needs, this invention supports selecting "Resource-intensive Projects" during the application stage, and the system will allocate an advanced computing power package of 16 cores and 32 gigabytes. The controller will use the Kubernetes DeviceClass mechanism to schedule Pods to physical nodes equipped with larger caches and memory bandwidth, ensuring the efficiency of complex correlation analysis.

[0037] Intelligent detection and automatic recycling technologies are implemented, including: To address the problem of "occupying without using," this invention proposes a "hardware and software combined" detection mechanism.

[0038] The three-dimensional idle determination index means that the recycling engine does not rely solely on timeout, but comprehensively evaluates the following three dimensions: Business activity layer: Detect whether the builders have any page redirects, file saves, or code commit records on the model management platform within the last 2 hours.

[0039] Interaction State Layer: Monitors the activity level of the Notebook's front end.

[0040] System resource layer: Monitors the CPU time slice growth rate of Python or ipykernel processes within the container. If ( If the threshold is extremely small, then the kernel is considered to be suspended but not released.

[0041] Automatic recycling of execution paths, including: Once the system determines that the model has exceeded the construction phase or is in a state of deep idleness, it performs "graceful reclamation": Automatically save state: Triggers the Notebook's Autosave interface to persist unsaved variable states to object storage.

[0042] Resource unregistration: Call the K8s API to release the ResourceClaim and return the computing power of 8c16g or 16c32g to the global resource pool.

[0043] Notification mechanism: The bank will notify the construction personnel that the resources have been recovered through internal office tools (such as WeChat Work / email) and provide a "one-click recovery" link.

[0044] Beneficial effects: 1. Exceptional Resource Utilization: By forcibly binding computing power to the model's lifecycle, the "Notebook zombie" problem prevalent in the banking industry is effectively solved. Experimental data shows that after introducing this method, the overall turnover rate of the computing power pool can be increased by more than 60%, significantly alleviating the pain point of expensive and scarce computing power.

[0045] 2. Refined cost control capabilities: Based on the account binding mechanism, the bank can clearly track the computing power consumption of each department and project. Through a tiered strategy of "basic package (8c16g) + advanced package (16c32g)," resources are allocated on demand, avoiding over-configuration of computing power resources for simple tasks.

[0046] 3. Significantly reduce model management risks: Automated management throughout the entire lifecycle ensures that resources are released in a timely manner after the model is developed and it is transferred to the compliant monitoring / auditing stage, preventing unauthorized deployment in the development environment and meeting the regulatory requirements for closed-loop management of the entire model process.

[0047] 4. Superior User Experience and Fairness: The automated "application-ready" mechanism reduces waiting time for manual approval. With inefficient resource usage significantly reduced, project developers can access available computing power more quickly, creating a fair and efficient research environment.

[0048] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing computing power configuration through software design, characterized in that, The configuration method includes: Deploy an intelligent scheduling controller between the model management platform and the K8s computing power cluster, and establish a lifecycle manager, identity and quota center, dynamic computing power controller, and intelligent recycling agent; Bind the entire lifecycle status of model development to computing resources, and automatically execute computing resource requests, upgrades, downgrades, or recycling based on model stage transitions; A mandatory binding mechanism using account-model-resource triples is adopted to associate user identity with model instance and underlying resource declaration; Elastic quotas for basic and extended computing power packages are provided based on model complexity and dynamically scheduled through software strategies. Inactivity is determined based on three dimensions: business activities, interaction status, and system resource layer indicators, and graceful automatic recycling is performed.

2. The method for optimizing computing power configuration through software design according to claim 1, characterized in that, The entire model development lifecycle includes: data preparation, model building, model validation, model deployment, and model decommissioning; among which: During the model building phase, a basic computing power package is automatically allocated and a Notebook container is launched. The model validation phase supports elastic task-based scaling. Switch to production-level quotas during the model deployment phase; The model is completely decoupled and its resources are released during the retirement phase.

3. The method for optimizing computing power configuration through software design according to claim 1, characterized in that, The account-model-resource triplet binding includes: obtaining user accounts through OIDC or LDAP protocol, binding the accounts with model instances and K8sResourceClaim objects one by one, and realizing resource ownership, usage statistics and roaming scheduling.

4. The method for optimizing computing power configuration through software design according to claim 1, characterized in that, The flexible quota mechanism includes: Basic computing power package: Default 8-core CPU, 16GB memory; Extended computing power package: 16-core CPU, 32GB memory or above configuration for graph models and large model fine-tuning; scheduled to the corresponding hardware nodes through K8sDRA dynamic resource allocation and DeviceClass mechanism.

5. The method for optimizing computing power configuration through software design according to claim 1, characterized in that, The three-dimensional inactivity determination includes: Business activity layer: Detection of platform page operations, code submissions, and file saving records; Interactive state layer: Monitors Notebook front-end connectivity and kernel activity; System resource layer: Monitors the growth rate of Python kernel CPU time slices to determine whether CPU usage is suspended.

6. The method for optimizing computing power configuration through software design according to claim 1, characterized in that, The elegant automatic recycling includes: automatically saving Notebook state and variable snapshots, unregistering K8s resource declarations, releasing computing power to the resource pool, sending recycling notifications to users, and providing a one-click recovery entry.

7. The method for optimizing computing power configuration through software design according to claim 1, characterized in that, The intelligent recycling agent is deployed as a side-vehicle container within the Notebook Pod, continuously reporting activity signals and executing recycling commands.

8. The method for optimizing computing power configuration through software design according to claim 2, characterized in that, When the model moves from the construction phase to the verification / retirement phase, the system automatically triggers a reduction or release of computing power without manual intervention.

9. The method for optimizing computing power configuration through software design according to claim 1, characterized in that, The system sets hysteresis thresholds and confirmation delays to avoid inactive judgment jitter and frequent recycling / restart.

10. A system for optimizing computing power configuration through software design, employing the method for optimizing computing power configuration through software design as described in any one of claims 1-9, characterized in that, include: Lifecycle Manager: Maintains the model's phase state and triggers computing power events; Identity and Quota Center: Manages user accounts, computing power wallets, and tiered quota policies; Dynamic computing power controller: Implements resource declaration, scheduling, and allocation / degradation using Kubernetes Operator; Intelligent recycling agent: Performs 3D inactivity detection and graceful recycling.

Images