swKokkos: An Athread Backend for Enhanced Kokkos with the Sunway Heterogeneous Architecture
Junlin Wei, Jinrong Jiang, Wu Wang, Chen Li, Yehong Zhang, Yue Yu, Lian Zhao, Xiang Han, Zhenjia Li, Feng Zhang, Haoyuan Zhang, Yidi Bai
Abstract
Performance portability is crucial for enabling the efficient execution of high-performance applications across evolving heterogeneous architectures. The Kokkos library offers an abstraction layer that decouples algorithmic logic from hardware-specific details. However, Kokkos lacks support for the Sunway processor, which is an essential heterogeneous architecture. We present swKokkos, an Athread backend that enables Kokkos on the Sunway many-core architecture. It introduces a compile-time kernel registration and runtime callback mechanism that breaks prior barriers preventing the template-intensive C++ Kokkos library from executing on Sunway compute processing elements (CPEs). We further implement three targeted optimizations: (1) vectorized hash-based callback dispatch to cut runtime abstraction overhead, (2) tile-based parallel task mapping algorithm to improve flexibility and load distribution, and (3) a fast device scheduling library that reduces device-side scheduling latency. We evaluate swKokkos for performance across multiscale scenarios. Moreover, swKokkos achieves up to 34.0× speedup over the host backend. Compared to the native device backend solution, it ranges from 0.7× to 2.8×, indicating that the abstraction overhead is low and occasionally reversed into gains through improved scheduling. In particular, swKokkos scales efficiently and stably to over 100000 core simulations in a real-world application, demonstrating its ability to support extreme-scale simulations on the Sunway exascale supercomputer.
Ask about this paper
Ask your agent about it.
Lune has read the top-tier papers around this one, so every answer names the papers it rests on.
Your agent calls
Lunesearch_papers
Free to start. No credit card required.
Terminal
Install the CLIlune papers get 952e3518-5e52-4e0f-8706-bcd0f3977df1Related papers
- A performance-portable nonhydrostatic atmospheric dycore for the energy exascale earth system model running at cloud-resolving resolutionsLuca Bertagna, Oksana Guba, Mark A. Taylor, James G. Foucar et al.SC 2020 · 20 citations
- SYCL++: A Unified Programming Framework for Heterogeneous Supercomputers at ScaleZitao Shen, Yuyang Jin, Kinman Lei, Zixuan Ma et al.HPDC 2026
- Enabling and scaling the HPCG benchmark on the newest generation Sunway supercomputer with 42 million heterogeneous coresQianchao Zhu, Hao Luo, Chao Yang, Mingshuo Ding et al.SC 2021 · 44 citations
- TensorKMC: kinetic Monte Carlo simulation of 50 trillion atoms driven by deep learning on a new generation of Sunway supercomputerHonghui Shang, Xin Chen, Xingyu Gao, Rongfen Lin et al.SC 2021 · 16 citations
- 5 ExaFlop/s HPL-MxP Benchmark with Linear Scalability on the 40-Million-Core Sunway SupercomputerRongfen Lin, Xinhui Yuan, Wei Xue, Wanwang Yin et al.SC 2023 · 11 citations
