AI Infrastructure Reproducibility Through Dataset Transformation Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence (AI) systems are resource-intensive and lack additional features useful to data scientists, focusing solely on model execution without providing a rich set of features.
Innovation Solution
Implementing a storage system with non-volatile RAM (NVRAM) as a buffer for data storage, offloading device management tasks from storage drives, and using direct-mapped flash storage systems to manage data blocks directly, reducing unnecessary operations and enhancing reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning models and artificial intelligence applications are executed with large amounts of data, then processing capability and model performance are improved, but resource consumption increases significantly
Solution Approach 1:
The patent segments the AI infrastructure into multiple storage array controllers that can independently manage different portions of data storage and processing tasks. Each controller operates semi-autonomously, allowing parallel processing of multiple machine learning workloads simultaneously, thereby improving overall system productivity while distributing resource consumption across multiple units rather than concentrating it in a single resource-intensive operation
Solution Approach 2:
The storage array controllers perform preliminary data preparation, filtering, and organization before data is fed into machine learning models. By pre-processing data and maintaining optimized data structures in advance, the system reduces the computational burden during actual model execution, enabling efficient processing of large datasets without proportionally increasing real-time resource consumption
2Productivity
If AI systems focus solely on model execution, then model performance is maintained, but additional useful features for data scientists are lost
Solution Approach 1:
The storage array controllers are designed with multi-functionality, serving both as high-performance storage devices and as intelligent data processing units. They can execute storage operations, perform data analytics, prepare training datasets, and provide real-time data access - effectively replacing multiple separate components and enabling data scientists to access diverse features and capabilities through a single unified system
Solution Approach 2:
The patent merges storage functionality with data processing and analytics capabilities into integrated storage array controllers. This consolidation allows the system to maintain model execution performance while simultaneously providing data scientists with additional features such as data exploration tools, preprocessing capabilities, and optimized data access patterns that would otherwise require separate systems
3Speed
If data processing operations are performed frequently, then data access speed is improved, but write operations increase and reduce efficiency
Solution Approach 1:
The storage array controllers implement periodic data caching strategies where data is loaded into high-speed memory buffers at regular intervals and maintained there for multiple read operations. This periodic loading approach allows rapid data access during the cache validity period without requiring continuous write operations, thereby maintaining high data access speed while minimizing the frequency of write operations to preserve write efficiency
Solution Approach 2:
The system performs preliminary data loading and caching operations in advance of actual data access needs. By pre-positioning data in high-speed storage buffers and maintaining optimized data structures beforehand, the system enables rapid data retrieval without requiring frequent write operations, thus achieving high data access speed while protecting write operation efficiency from degradation
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
Ensuring reproducibility in an artificial intelligence infrastructure that includes one or more storage systems and one or more graphical processing unit ('GPU') servers, including: identifying, by a unified management plane, one or more transformations applied to a dataset by the artificial intelligence infrastructure, wherein applying the one or more transformations to the dataset causes the artificial intelligence infrastructure to generate a transformed dataset; storing, within the one or more storage systems, information describing the dataset, the one or more transformations applied to the dataset, and the transformed dataset; identifying, by the unified management plane, one or more machine learning models executed by the artificial intelligence infrastructure using the transformed dataset as input; and storing, within the one or more storage systems, information describing one or more machine learning models executed using the transformed dataset as input.