Time-Synchronized Storage Delivery for Multi-AI Training Bandwidth
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI model training processes face challenges in efficiently managing large data sets and synchronizing data delivery across multiple AI processing units, leading to inefficiencies in processing and training due to lack of effective bandwidth management and timing synchronization.
Innovation Solution
The proposed solution involves a method for time-synchronized storage delivery, where input/output commands from AI processing units are assigned timestamps based on their respective bandwidths, allowing for prioritized and synchronized data chunk transmission between AI model processing units and storage devices using global flow identifiers, ensuring that data is transmitted at the appropriate times to meet bandwidth and latency requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If data is transmitted to multiple AI processing units in parallel without timestamp synchronization, then processing speed is improved, but data delivery timing accuracy deteriorates
Solution Approach 1:
The system assigns timestamps to IO commands before data transmission occurs. This preliminary timing assignment ensures that when data is transmitted in parallel to multiple AI processing units, each unit receives data at the precisely scheduled time, maintaining both high processing speed and accurate timing synchronization.
Solution Approach 2:
Timestamps act as an intermediary mechanism between the storage device and multiple AI processing units. The timestamps provide a common reference framework that coordinates data delivery across parallel operations, enabling synchronized processing without compromising transmission speed.
2Ease of operation
If bandwidth is allocated equally to all AI model training operations, then system simplicity is maintained, but data delivery efficiency deteriorates
Solution Approach 1:
The system assigns different bandwidth allocations to different AI model training operations based on their specific requirements. Each IO command receives a timestamp reflecting its prioritized bandwidth allocation, allowing critical operations to receive data faster while less critical operations use standard bandwidth, optimizing overall system efficiency.
3Device complexity
If data transmission timing is not synchronized with bandwidth allocation, then system complexity is reduced, but processing latency increases
Solution Approach 1:
Timestamps are assigned to IO commands in advance, before actual data transmission occurs. This preliminary timing setup creates a synchronized framework that coordinates bandwidth allocation with data transmission timing, ensuring minimal latency without requiring complex real-time coordination mechanisms.
Data Source
AI summary
A method includes receiving, at a first computing device, a first input/output (IO) command from a first artificial intelligence processing unit (AI PU), the first IO command associated with a first AI model training operation. The method further includes receiving, at the first computing device, a second IO command from a second AI PU, the second IO command associated with a second AI model training operation. The method further includes assigning a first timestamp to the first IO command based on a first bandwidth assigned to the first AI model training operation. The method further includes assigning a second timestamp to the second IO command based on a second bandwidth assigned to the second AI model training operation.