Year: 2023
Type: conference-paper
Abstract: The Dataflow for C++(DFCPP) designed and implemented in this paper is a parallel programming library for dataflow computing on a general control flow hardware platform. Compared with existing dataflow programming libraries, DFCPP has an easy-to-use user interface and richer expression ability and can express different kinds of dataflow tasks. DFCPP designs a memory manager and task scheduler for the NUMA (Non-Uniform Memory Access) architecture and has filled the gap of existing dataflow runtime libraries in this field. The memory manager can allocate memory space on specific NUMA nodes according to the needs of tasks. As for the task scheduler, on the one hand, it places jobs at the location of their input data based on the idea, "The data is where it is; Move a thread to the data it needs most." On the other hand, the task scheduler uses an improved task-stealing algorithm to steal the most suitable task. Combined with the memory manager and task scheduler, DFCPP can decrease the times of remote NUMA node access, thereby improving execution efficiency.
Authors: Qiuming Luo, Jinke Huang, Jinrong Li, Zheng Du
Institutions: Shenzhen University, Tsinghua University
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
Award: G1848575682
OpenAlex ID: W4386504997