Abstract:

[Context] Fast Radio Burst (FRB) search is one of the important scientific goals of the 500-meter Aperture Spherical Radio Telescope (FAST), which has high computational complexity and a large amount of raw data. The GPU utilization rate is low by the current FRB search algorithms, and data processing demands more manual intervention. [Object] The research studied the process-level GPU parallel optimization method to achieve the effects that realizing multi-process parallelism of multiple, improving GPU utilization, simplifying the operation and scheduling of algorithms, and supporting the use of automated scripts on the premise of not modifying the algorithm implementation. [Methods] The containerized-encapsulation FRB search algorithm with GPU aggregation technology is utilized to support GPU idle time multiplexing, and shield technical details, like GPU calls and dependency library management. [Results] The experimental results show that the parallel optimization program can achieve a speedup of 5.3 for the FRB search algorithm, and a superior parallel efficiency of 0.88, running on a single GPU binding to 6 computing processes, without modifying the original algorithm or increasing GPU resources. [Conclusions] It can improve GPU utilization, and computing efficiency, and support automated processing effectively by the way of parallel optimization based on containerized encapsulation and process-level GPU aggregation. Also, the program has good generality and can be applied to parallel optimization of similar applications.

Key words: fast radio burst, containerization, process-level GPU parallel optimization, GPU aggregation