ML Temperature Forecasting for Storage Objects via Sub-Object Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learning-based approaches for forecasting the future activity of storage objects in storage systems are computationally expensive due to high memory footprint and CPU overhead, making them prohibitive for systems with millions of objects.
Innovation Solution
The method involves dividing storage objects into sub-objects and using a first machine learning model to forecast temperature, with a second model determining temperature for each sub-object using a limited subset of IO features, and projecting this temperature onto each sub-object, thereby reducing computational resources required.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are used to forecast storage object temperature, then forecasting accuracy is improved, but computational cost and memory footprint increase
Solution Approach 1:
The patent segments the storage object into multiple sub-objects and applies different machine learning model complexities to different segments. Hot sub-objects use full ML models for high accuracy, while cold sub-objects use simplified models or statistical methods, reducing overall computational cost while maintaining accuracy for important objects.
Solution Approach 2:
The patent applies local quality by using different forecasting approaches based on the temperature and importance of specific sub-objects. High-value hot sub-objects receive sophisticated ML modeling with full feature sets, while less important cold sub-objects use simpler methods, optimizing the balance between accuracy and computational resources.
2Measurement precision
If machine learning models process all storage objects, then forecasting accuracy is improved, but CPU overhead becomes prohibitive
Solution Approach 1:
The patent divides the storage object into sub-objects and processes them with different levels of ML model complexity. This segmentation allows the system to maintain high forecasting accuracy for critical sub-objects while using lightweight processing for others, thereby improving overall system throughput without sacrificing essential accuracy.
Solution Approach 2:
The patent applies partial action by using full machine learning processing only for hot and warm sub-objects that require accurate forecasting, while using simplified statistical methods for cold sub-objects. This selective approach maintains necessary accuracy while significantly reducing CPU overhead and improving system productivity.
3Measurement precision
If full machine learning models are applied to all sub-objects, then temperature forecast accuracy is improved, but memory footprint increases
Solution Approach 1:
The patent segments sub-objects by temperature category and applies different model complexities accordingly. Hot sub-objects use comprehensive ML models with full feature sets requiring more memory, while cold sub-objects use simplified models with reduced feature sets, optimizing memory footprint while maintaining forecast accuracy where needed.
Solution Approach 2:
The patent applies local quality by matching model complexity to the specific needs of each sub-object segment. Critical hot sub-objects receive high-accuracy models with appropriate memory allocation, while less important cold sub-objects use lightweight models, thereby optimizing the overall memory footprint while maintaining necessary forecast accuracy.
Data Source
AI summary
A method, computer program product, and computing system for forecasting a temperature of a storage object of a storage system using a first machine learning model and a plurality of input/output (IO) features. The storage object may be divided into a plurality of storage sub-objects. A temperature may be determined for each storage sub-object with a subset of the plurality of IO features using a second machine learning model. A portion of the temperature of the storage object may be projected onto the temperature of each of the plurality of storage sub-objects based upon, at least in part, the temperature determined for each storage sub-object and the temperature determined for each storage object.