GPU Model Execution Using Storage-Transformed Datasets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing storage systems face inefficiencies in managing flash storage devices, particularly in the management of erase blocks and data operations, leading to unnecessary write operations and reduced reliability.
Innovation Solution
Implementing a direct-mapped flash storage system where the operating system directly addresses data blocks without translation by storage controllers, initiating and controlling processes such as data rewriting and erasure, and utilizing separate storage controllers for failover and collaboration to manage erase blocks and data allocation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If storage controllers manage data operations and erase blocks, then data management is automated, but write operations become redundant and reliability decreases
Solution Approach 1:
The patent extracts the data management functions from storage controllers and places them directly in the operating system. The OS now directly addresses data blocks on flash storage devices without translation by storage controllers, eliminating redundant write operations and improving reliability while maintaining automation at the OS level.
Solution Approach 2:
The patent introduces a direct mapping layer in the OS that acts as an intermediary between the file system and flash storage devices. This direct-mapped flash storage system eliminates the need for storage controllers to translate addresses, reducing redundant operations while maintaining automated management through OS-level control.
2Ease of operation
If storage controllers translate data blocks, then data access is simplified, but write operations increase and efficiency decreases
Solution Approach 1:
The patent removes the translation function from storage controllers and implements direct addressing in the OS. Data blocks are accessed directly without translation, eliminating redundant write operations and improving write efficiency while maintaining ease of operation through OS-level abstraction.
Solution Approach 2:
The patent replaces the mechanical translation process performed by storage controllers with a software-based direct addressing system in the OS. This substitution eliminates the need for physical translation operations, reducing write amplification and improving efficiency.
3Reliability
If multiple storage controllers are used for failover, then system reliability improves, but device complexity increases
Solution Approach 1:
The patent merges the data management functions previously distributed across multiple storage controllers into a single centralized OS management system. This consolidation maintains failover capabilities through OS-level redundancy while reducing overall system complexity by eliminating duplicate controller architectures.
Solution Approach 2:
The patent creates a universal OS-level data management system that handles multiple functions previously performed by separate storage controllers, including address translation, wear leveling, and failover management. This multi-functional approach reduces complexity while maintaining reliability.
Data Source
AI summary
Executing a machine learning model in an artificial intelligence infrastructure that includes one or more storage systems and one or more graphical processing unit (‘GPU’) servers, including: receiving, by a graphical processing unit (‘GPU’) server, a dataset transformed by a storage system that is external to the GPU server; and executing, by the GPU server, one or more machine learning algorithms using the transformed dataset as input.