Workload Partitioning for Spatial Architecture Resource Assignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex computing workloads, such as deep learning, face challenges in efficiently organizing and assigning operations to available hardware resources due to constraints like spatial architecture and bandwidth limitations, making it difficult to achieve computationally efficient workload management.
Innovation Solution
A computerized method that partitions workload operations into subgraphs to minimize their quantity and maximize resource utilization, then maps these subgraphs to logical processing engine units and assigns them to physical units to minimize network traffic, using techniques like Integer Linear Programming and Einstein Summation notation for efficient resource allocation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the workload graph is partitioned into more subgraphs, then resource utilization per subgraph is maximized, but the quantity of subgraphs increases
Solution Approach 1:
The workload graph is segmented into subgraphs that can be independently mapped to physical devices. The partitioning algorithm divides the graph into contiguous subgraphs with minimal cross-subgraph edges, allowing each subgraph to be optimized for resource utilization while maintaining manageable complexity through systematic segmentation strategies.
2Manufacturing precision
If operations are assigned to physical resources considering spatial architecture constraints, then resource assignment accuracy is improved, but computational efficiency of assignment decreases
Solution Approach 1:
The system performs preliminary actions by pre-computing valid mappings between logical PE units and physical devices based on spatial architecture constraints. Bandwidth thresholds and connectivity requirements are pre-established, allowing the assignment algorithm to quickly match subgraphs to physical resources without repeatedly validating complex constraints during the optimization process.
3Loss of energy
If the workload is mapped to physical devices considering bandwidth limitations, then network traffic is minimized, but the complexity of mapping increases
Solution Approach 1:
The mapping problem is segmented into multiple stages: first partitioning the workload graph into subgraphs, then mapping each subgraph to physical devices independently. This segmentation allows bandwidth constraints to be enforced at the subgraph level rather than across the entire workload, reducing the complexity of constraint satisfaction while minimizing network traffic through optimal subgraph placement.
Data Source
AI summary
Operations of a workload are assigned to physical resources of a physical device array. The workload includes a graph of operations to be performed on a physical device array. The graph of operations is partitioned into subgraphs. Partitioning includes at least minimizing the quantity of subgraphs and maximizing resource utilization per subgraph. A logical mapping of the subgraph to logical processing engine (PE) units is generated using features of the subgraph and tiling factors of the logical PE units. The logical mapping is assigned to physical PE units of the physical device array at least by minimizing network traffic across the physical PE units. The operations of the subgraph are performed using the physical PE units to which the logical mapping is assigned. This process enhances the computational efficiency of the array when executing the workload.