Streaming Dynamic Coarse-Grained CPU/GPU Workloads with Heterogeneous Pipelines in FastFlow

摘要

Software pipelines permit the decomposition of a repetitive sequential process into a succession of distinguishable sub-processes called stages, each of which can be concurrently executed on a distinct processing element. This paper presents a heterogeneous streaming pipeline implementation using the FastFlow skeletal library for a numerical linear algebra code. By introducing minimal memory management, we implement a large-scale streaming application which allocates the different pipeline stages to multi-core CPU and multi-GPU resources in a cluster environment, demonstrating the suitability of the algorithmic skeleton approach to efficiently coordinate the pipeline operation. Our implementation shows that long- running heterogeneous pipelines can be effectively implemented in FastFlow.