Neural DAG Scheduling via One-Shot Priority Sampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning schedulers for artificial neural networks are computationally expensive due to their reliance on episodic reinforcement learning frameworks, requiring multiple rounds of neural network processing, which limits their scalability and effectiveness in scheduling operations on large graphs and timely domains like compilers.
Innovation Solution
A one-shot neural network encoder is combined with list scheduling to sample node priorities in parallel, generating a final schedule using a Gumbel Top-K function, reducing computational resources and improving performance by eliminating the need for multi-round processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning schedulers use episodic reinforcement learning frameworks with multiple rounds of neural network processing, then scheduling accuracy can be improved, but computational cost and processing time increase significantly
Solution Approach 1:
The patent applies preliminary action by performing a single round of neural network processing to generate node priorities before scheduling begins. This one-shot approach pre-computes the priority information that would otherwise require multiple reinforcement learning rounds, thereby reducing computational cost while maintaining scheduling accuracy.
Solution Approach 2:
The patent extracts only the essential priority information from the neural network processing in a single pass, rather than performing multiple rounds of processing. By taking out just the necessary priority values needed for scheduling decisions, the system achieves accurate scheduling with significantly reduced computational overhead.
2Measurement precision
If conventional schedulers perform multiple rounds of neural network processing, then scheduling quality improves, but processing speed decreases
Solution Approach 1:
The system performs preliminary action by completing all necessary neural network processing in one shot before scheduling execution. This single-round approach pre-determines node priorities that guide the entire scheduling process, eliminating the need for multiple processing rounds and thereby improving processing speed while maintaining scheduling quality.
Solution Approach 2:
The patent segments the scheduling process into two distinct phases: (1) a one-shot neural network processing phase that generates node priorities, and (2) a scheduling execution phase that uses these priorities. This segmentation allows the computationally intensive neural network processing to be performed only once, improving overall processing speed while preserving scheduling quality.
3Reliability
If conventional reinforcement learning schedulers use multi-round processing, then scheduling effectiveness improves, but scalability to large graphs deteriorates
Solution Approach 1:
The patent applies preliminary action by performing a single round of neural network processing to generate comprehensive node priorities that guide the entire scheduling process. This one-shot approach reduces the computational burden that would otherwise scale poorly with graph size, enabling the system to handle large graphs effectively while maintaining scheduling quality.
Solution Approach 2:
The patent changes the processing parameter from multiple reinforcement learning rounds to a single round of neural network processing. This parameter change reduces the computational complexity from O(n^2) or worse to O(n), where n is the number of nodes, thereby improving scalability to large graphs while maintaining scheduling effectiveness through the quality of the generated priorities.
Data Source
AI summary
A processor-implemented method includes sampling, according to a priority sampling policy, a set of node priorities from a computation graph. Each node priority of the set of node priorities may be associated with a respective node on the computation graph. Additionally, each node may represent an operation of a task performed by an artificial neural network. The method also includes converting, via a list scheduling function, the node priorities to a schedule that associates each node of the computation graph with a processor of a group of processors of a device associated with the artificial neural network, the schedule associated with a makespan. The method further includes performing the task in accordance with the schedule.