DMA Bandwidth Allocation Using Adaptive MO Values for AI Tasks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional DMA operations in AI services fail to accurately reflect priorities and perform data transmission in real time, especially when multiple operations are performed simultaneously, leading to inefficiencies in data transfer for AI computation tasks.
Innovation Solution
A data processing method that derives Multiple Outstanding (MO) values based on waiting times for computation tasks, adjusting bandwidth allocation to prioritize urgent DMA operations by reflecting changed waiting times and predicted execution times.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple DMA operations are performed simultaneously using conventional bandwidth allocation methods, then data transfer capacity is increased, but priority accuracy and real-time responsiveness deteriorate
Solution Approach 1:
The patent applies dynamics by making the MO value adjustable and adaptive rather than fixed. The system dynamically modifies the MO value for each DMA operation based on real-time priority information and predicted execution times, allowing the bandwidth allocation to change flexibly according to current system conditions and task urgency levels
Solution Approach 2:
The patent changes the parameter of MO (Multiple Outstanding) values to control DMA operation priorities. By deriving different MO values based on priority levels and predicted execution times, the system adjusts bandwidth allocation parameters to reflect real-time requirements, enabling urgent tasks to receive higher priority treatment while maintaining overall system throughput
2Device complexity
If conventional bandwidth allocation is used for multiple DMA operations, then system simplicity is maintained, but real-time priority reflection and data transfer efficiency worsen
Solution Approach 1:
The patent applies preliminary action by predicting the execution time of computation tasks before they actually execute. The command processor calculates predicted execution times in advance and uses these predictions to derive appropriate MO values, allowing the system to proactively allocate bandwidth based on anticipated task requirements rather than reacting after tasks are submitted
Solution Approach 2:
The patent implements feedback by continuously monitoring DMA operation priorities and predicted execution times, then using this information to adjust MO values dynamically. The system creates a closed-loop control mechanism where bandwidth allocation is continuously optimized based on feedback from task priority levels and performance predictions, improving data transfer efficiency through adaptive resource management
Data Source
AI summary
A data processing method using a DMA, the method comprising: deriving a waiting time of a waiting time of a first DMA operation for a first computation task and a waiting time of a second DMA operation for a second computation task, deriving a MO value of the first DMA operation and a MO value of the second DMA operation based on the waiting time of the first DMA operation and the waiting time of the second DMA operation, and performing the first DMA operation based on the MO value of the first DMA operation, and performing the second DMA operation based on the MO value of the second DMA operation.