Cloud Database Performance Prediction via Workload Fingerprinting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing performance prediction techniques for cloud-based databases are not designed to perform predictions across different hardware platforms and require extensive re-training for each new hardware configuration, failing to account for the execution environment and being non-portable.
Innovation Solution
A method for building a cloud database performance model using machine learning and collaborative filtering, which maps test server performance to cloud performance by iteratively executing workloads under varying resource constraints and using a fingerprint to characterize workloads across different hardware configurations, allowing for predictions across multiple platforms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing performance prediction techniques are used, then predictions can be made for single hardware platforms, but the techniques require extensive re-training for each new hardware configuration and are not portable across platforms
Solution Approach 1:
The patent creates a universal performance prediction model that can predict database workload performance across multiple different hardware platforms and cloud providers without requiring platform-specific retraining. The system uses a standardized set of performance metrics and workload characteristics that can be applied universally across AWS, Rackspace, and other cloud infrastructure providers, making the prediction tool multi-functional and platform-agnostic
Solution Approach 2:
The system performs preliminary characterization of workloads by collecting performance metrics across various hardware configurations during an initial training phase. This preliminary action creates a comprehensive model that captures workload behavior patterns, which can then be applied to predict performance on new platforms without requiring extensive retraining when hardware configurations change
2Measurement precision
If performance prediction models are trained extensively for each hardware configuration, then prediction accuracy for that specific configuration improves, but the complexity and cost of deployment increases significantly
Solution Approach 1:
The patent develops a single universal prediction model that maintains high accuracy across multiple hardware platforms without requiring separate models for each configuration. The system achieves this by training on diverse hardware configurations during the initial phase and using standardized performance metrics that capture workload behavior universally, thereby reducing deployment complexity while maintaining measurement precision
Solution Approach 2:
The system uses parameter-based workload characterization that captures essential performance characteristics without being tied to specific hardware configurations. By focusing on measurable performance parameters rather than hardware-specific details, the model can adapt to different hardware configurations through parameter variations rather than requiring complete retraining, thus maintaining accuracy while simplifying deployment
3Measurement precision
If platform-specific prediction models are used, then predictions are accurate for that platform, but the system cannot generalize performance across different cloud providers and hardware configurations
Solution Approach 1:
The patent creates a universal prediction framework that can accurately predict performance across multiple cloud providers (AWS, Rackspace, etc.) and different hardware configurations. The system uses standardized workload characterization and performance metrics that are platform-agnostic, allowing the same model to generalize across diverse infrastructure while maintaining prediction accuracy through comprehensive initial training on varied hardware
Data Source
AI summary
Systems and methods are disclosed to perform performance prediction for cloud-based databases by building on a computer a cloud database performance model using a set of training workloads; and using a learned model on the computer to predict database performance in the cloud for a new workload, wherein for each reference workload r and hardware configuration h, system throughput tr,h, average throughput of αr and standard deviation σr, comprising normalizing each throughput as:tr,h_=tr,h-arσr.