Composing Pipeline Parallelism using Control Taskflow Graph
Cheng-Hsiang Chiu, Tsung‐Wei Huang · 2022
Graph-based propagation (GBP) is a common parallel pattern in many graph computing applications. Many GBP applications compose pipeline parallelism for each linear segment in the graph, where each task encapsulates a sequence of linearly dependent functions. This type of task-parallel pipeline parallelism is hard to express using mainstream programming frameworks (e.g., oneTBB) that count on data-parallel models to perform pipeline scheduling. In this paper, we introduce a new task-parallel method to compose pipeline parallelism in a GBP workload by leveraging the state-of-the-art control taskflow graph model. We demonstrate the promising performance of our method on a real circuit simulation workload.