Machine-Learned Storage Performance Tuning for Reliable Throughput
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Non-volatile memory storage devices face challenges in providing performance that meets specific service requirements, such as reliability and throughput consistency, due to variations in management operations.
Innovation Solution
A storage system incorporating a controller with performance units and a machine learning model that maps performance indices to operating parameter values, allowing for dynamic adjustment of performance units to meet desired performance indices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If firmware solutions are used to maximize storage device performance, then productivity is improved, but reliability deteriorates due to inability to provide consistent performance for different service requirements
Solution Approach 1:
The patent implements dynamic performance adjustment by allowing the storage device to switch between different performance levels based on service requirements. The controller receives performance index values from the host and adjusts operating parameters in real-time, enabling the system to adapt between high-performance modes (for throughput-critical services) and reliable modes (for consistency-critical services), thus resolving the contradiction between maximizing productivity and ensuring reliability.
Solution Approach 2:
The patent changes operating parameters dynamically based on performance index values. The controller modifies parameters such as cache allocation, garbage collection frequency, and wear-leveling intensity according to the selected performance level. This parameter adjustment mechanism allows the storage device to provide different performance characteristics (high throughput vs. high reliability) as needed, resolving the contradiction between productivity and reliability.
2Reliability
If management operations are performed to manage non-volatile memory, then reliability is improved, but productivity deteriorates due to performance variations
Solution Approach 1:
The patent applies partial management operations based on the performance index. When high performance is required, the controller reduces the intensity of management operations (such as wear-leveling and garbage collection) to minimize performance impact. When high reliability is required, the controller increases management operation intensity. This selective application of management operations resolves the contradiction between reliability improvement and productivity maintenance.
Solution Approach 2:
The patent implements periodic management operations that can be adjusted in frequency based on performance requirements. The controller can schedule wear-leveling and garbage collection operations at different intervals depending on the current performance index, allowing the system to balance between performing necessary management tasks for reliability and minimizing their impact on productivity.
3Adaptability or versatility
If performance units are adjusted to meet specific service requirements, then adaptability is improved, but device complexity increases
Solution Approach 1:
The patent implements a self-service mechanism where the storage device automatically adjusts its performance units based on the performance index received from the host. The controller autonomously selects appropriate operating parameters and configures performance units without requiring complex external control logic or manual intervention. This self-adjusting capability provides high adaptability while keeping the control system relatively simple.
Solution Approach 2:
The patent uses feedback from the host device in the form of performance index values to automatically adjust storage device operations. The host provides feedback about its service requirements, and the controller uses this feedback to automatically configure performance units and operating parameters. This feedback mechanism enables adaptability without requiring complex bidirectional communication protocols or control algorithms.
Data Source
AI summary
A controller of a storage device includes a plurality of performance units for controlling performance of the storage device, and controls a non-volatile memory device. A host device receives a plurality of first operating parameter values of each performance unit from the controller, generates a plurality of combinations for the plurality of performance units based on the plurality of first operating parameter values, and inputs the plurality of combinations into a machine learning model to infer a plurality of performance indices respectively corresponding to the plurality of combinations.