AI Data Storage Padding to Reduce Write Amplification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The inconsistent data amounts between data units used by AI models and the capacity of physical units in rewritable non-volatile memory modules lead to frequent triggering of garbage collection programs, increasing write amplification and reducing performance.
Innovation Solution
A data storage method that generates second data with a greater data amount than the first data, including dummy data if necessary, to fill the capacity of physical units, thereby stabilizing operation by reducing write amplification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If data units from AI models are directly stored in physical units of rewritable non-volatile memory module, then storage operation is simple, but write amplification increases due to capacity mismatch
Solution Approach 1:
The patent segments the storage process into two distinct stages: first storing AI model data units in a buffer memory without immediate writing to physical units, and then performing batch consolidation later. This segmentation allows the system to decouple the storage operation from the write amplification problem, as the buffer can accommodate variable-sized data units while the physical units are filled in optimized batches that match their capacity, thereby reducing unnecessary write operations.
Solution Approach 2:
The patent applies preliminary action by pre-storing AI model data units in the buffer memory before committing them to physical units in the rewritable non-volatile memory module. This preliminary storage step allows the system to accumulate data units and perform capacity-matched consolidation later, avoiding the need to write partial or mismatched data amounts to physical units, thus reducing write amplification while keeping the initial storage operation simple.
2Speed
If data units from AI models are directly stored in physical units of rewritable non-volatile memory module, then storage process is fast, but garbage collection is frequently triggered
Solution Approach 1:
The patent segments the storage system into a buffer memory layer and a physical unit layer, allowing fast initial storage of AI model data units in the buffer without triggering garbage collection. The buffer acts as an intermediate layer that absorbs the speed requirements while the physical units are updated in controlled batches, preventing frequent garbage collection triggers and improving operation stability.
Solution Approach 2:
The patent performs preliminary storage of AI model data units in the buffer memory before writing to physical units. This preliminary action separates the fast storage operation from the garbage collection-triggering write operation, allowing the system to maintain high storage speed while avoiding frequent garbage collection by consolidating writes until physical unit capacity thresholds are met.
3Device complexity
If data units from AI models are directly stored in physical units of rewritable non-volatile memory module, then storage operation is straightforward, but performance deteriorates over time
Solution Approach 1:
The patent segments the storage operation into a simple buffer storage phase and an optimized consolidation phase. The buffer memory handles straightforward storage of AI model data units with minimal complexity, while the consolidation phase optimizes writes to physical units by matching capacity, thereby maintaining simple operations while improving long-term storage performance through reduced write amplification and fewer garbage collection cycles.
Solution Approach 2:
The patent performs preliminary storage in the buffer memory, which maintains straightforward storage operations while deferring the optimized consolidation to a later stage. This preliminary action preserves the simplicity of the storage operation while improving performance over time, as the consolidated writes to physical units are more efficient and reduce the frequency of performance-degrading garbage collection operations.
Data Source
AI summary
A data storage method, a host system, and a data storage system are disclosed. The method includes the following. An artificial intelligence (AI) model is executed. First data to be stored to a memory storage device is obtained. In response to the first data being generated by the AI model, second data is generated according to the first data, in which the second data includes the first data, and a data amount of the second data is greater than a data amount of the first data. A first write command is sent to the memory storage device according to the second data, so as to instruct the memory storage device to store the second data.