ML Model Repository Version Tree Traversal for IO Load Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
IT system operations managers face bottlenecks in storage IO load forecasting, which hinders agility in updating customer applications, as existing methods are inefficient in predicting IO loads for containers, leading to potential application performance failures.
Innovation Solution
A machine learning (ML) model repository is utilized, where ML models are searched and executed based on a container's version tree to predict IO loads, allowing for efficient selection and application of relevant models, even if no exact match is found, by traversing the version tree and applying models with sufficient accuracy thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional storage IO load forecasting methods are used, then application performance can be monitored, but the forecasting process becomes a bottleneck that prevents agile updates and reduces system productivity
Solution Approach 1:
The patent segments the monolithic forecasting process into multiple independent ML models, each specialized for specific container versions or version ranges. This segmentation allows parallel execution of multiple models rather than sequential processing, significantly reducing the time required for IO load forecasting while maintaining prediction accuracy across different container versions
Solution Approach 2:
The patent implements preliminary action by pre-training multiple ML models during the container image build process or in advance before deployment. The models are stored in a repository and automatically selected based on container version matching, eliminating the need for real-time model training or manual model selection during production operations
2Measurement precision
If multiple ML models are maintained for different container versions, then prediction accuracy for specific versions improves, but the complexity of model selection and management increases
Solution Approach 1:
The patent implements feedback mechanisms where the system automatically evaluates prediction accuracy metrics for each ML model and uses this feedback to refine model selection strategies. The feedback loop enables automatic model retraining, version updates, and repository maintenance, reducing manual intervention while improving prediction precision over time
Solution Approach 2:
The system implements self-service by automatically selecting appropriate ML models based on container version information without requiring manual intervention. The model selection process is automated through version tree traversal and matching algorithms that autonomously identify and execute the most suitable model for each container
3Measurement precision
If exact version matching is required for ML model selection, then prediction accuracy for matching versions is maximized, but the system cannot handle containers with unknown or new versions
Solution Approach 1:
The patent implements universality by creating ML models that can serve multiple container versions through version range specifications. A single model can be applied to multiple container versions within a specified range, allowing the system to handle both exact matches and approximate matches, thereby improving adaptability while maintaining reasonable prediction accuracy
Solution Approach 2:
The patent applies inversion by reversing the traditional approach: instead of requiring containers to exactly match model versions, the system inverts the matching logic to allow models to specify version ranges they support. This enables new container versions to be handled by finding the nearest compatible model in the version tree, improving system adaptability
Data Source
AI summary
Example implementations described herein are directed to searching ML models to facilitate IO load prediction. Even if there is no existing ML model applicable to the underlying container, example implementations described herein obtain the ML model for the nearest container. Through such example implementations, the underlying systems can reduce the total ML models maintained for the prediction of IO load and can enhance ML models with enough learning data for each ML model.