Dynamic Metric Prediction Using Inflection-Based Model Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting dynamic metrics in computer systems, such as power and performance, are ad-hoc and lack generalizability, leading to unreliable predictions that do not effectively adapt to changing workloads.
Innovation Solution
A method that dynamically chooses an objective function and prediction model from a pool of available options, using past values to locate points of inflection and solve equations for model parameters, allowing for accurate prediction of future metric values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If ad-hoc prediction methods are used for specific metrics, then prediction can be performed for that specific metric, but the method does not generalize to other metrics or unseen data
Solution Approach 1:
The patent creates a universal prediction framework that can handle multiple different metrics (power, performance, temperature, reliability) through a single principled approach. The objective function and prediction model are designed to be metric-agnostic, allowing the same methodology to be applied across different computer system metrics without requiring metric-specific customization, thereby achieving both generalizability and reliability.
Solution Approach 2:
The patent transforms the prediction problem by changing the parameter representation from metric-specific values to a unified mathematical framework using objective functions and model parameters. By representing different metrics through a common parameter space and optimization approach, the system achieves generalizability while maintaining prediction reliability through consistent mathematical principles.
2Measurement precision
If multiple prediction models are maintained for different metrics, then each metric can be predicted accurately, but the system complexity increases
Solution Approach 1:
The patent employs a single universal prediction model that can accurately predict multiple different metrics through a common mathematical framework. Instead of maintaining separate models for each metric, the system uses one versatile model with an objective function that can be optimized for any metric of interest, thereby reducing system complexity while preserving prediction accuracy.
Solution Approach 2:
The patent reduces complexity by changing from multiple metric-specific models to a single model with adjustable parameters. The objective function allows the same model structure to adapt to different metrics by optimizing different parameters, eliminating the need for multiple complex models while maintaining accuracy through parameter optimization.
3Adaptability or versatility
If dynamic workload changes occur, then system performance must adapt quickly, but ad-hoc methods cannot respond effectively to changing conditions
Solution Approach 1:
The patent implements a dynamic prediction system that continuously adapts to changing workloads through real-time optimization of the objective function. The system can dynamically adjust model parameters and re-optimize predictions as workload conditions change, enabling effective response to dynamic conditions while maintaining performance optimization through principled mathematical methods.
Solution Approach 2:
The patent incorporates feedback mechanisms where prediction accuracy is continuously evaluated and used to refine the objective function and model parameters. This feedback loop enables the system to adapt to dynamic workload changes by learning from actual performance data and adjusting predictions accordingly, improving both adaptability and productivity through iterative optimization.
Data Source
AI summary
An objective function is dynamically chosen from a pool of available objective functions, and a prediction model is dynamically chosen from a pool of available prediction models. Points of inflection are determined for the chosen objective function, based on past values of a metric, to obtain a set of equations that can be solved to obtain model parameters associated with the chosen prediction model. The equations are solved to obtain the model parameters, and a future value of the metric is predicted based on (i) at least some of the past values of the metric and (ii) the chosen prediction model, with the obtained associated model parameters.