FMM is an efficient algorithm for N-body problems. It consists of multiple phases with various computational characteristics and dependencies. Both of input dataset characteristics (distribution of bodies) and user-specified parameter affect load balance between the phases. It is a challenging topic to adopt such performance variation and exploit computing resources on a heterogeneous computer. In this poster we implement and evaluate a dynamic task scheduling FMM implementation based on kifmm3d and StarPU. The performance is compared to existing approaches and best cases by brute-force. We show that our approach can adaptively schedule computation but it still has performance challenges.
各種検索
サポート
ヘルプ