Abstract:

[Objective] Phase Field simulation is a powerful tool for studying microstructure evolution in alloys. The intrinsic spatiotemporal heterogeneity of the evolution process requires extreme scale simulation. In-situ data processing becomes necessary in these scenarios for the raw simulation data being produced in vast amounts at high speed. [Methods] Through the systematic design of novel feature extraction and 3D convolution algorithms and careful optimization for adapting the new generation Sunway supercomputer, we present here a novel feature extraction framework that enables us to extract characteristic features of each grain on the fly in a real-world alloy simulation with 33.6 trillion dofs. For the feature extraction, we propose a fully 3D convolutional network structure which is more conducive to performing optimization. In the aspect of pose estimation, which we are particularly interested in, it surpassed the performance of the current state-of-the-art methods. For the 3D convolution optimization, we propose a three-level blocking scheme with a novel scatter communication strategy to make full use of the on-chip network bandwidth. It allows 3D convolution to be accelerated with the SIMD vector unit of SW26010Pro without explicit matrix reconstruction. The operator achieves up to 91% of the theoretical peak performance on the new generation Sunway processor. [Results] Using over 39 million cores on the new generation Sunway supercomputer, sustained performance of 637 PFlops in mixed double and single precision is reached. Meanwhile, the evolution process of over four million grains is extracted and saved at the cost of less than 10% of the overall simulation time. [Conclusions] This data can be considered a detailed record of the evolution process and open the gates to various new approaches toward a better understanding of the process-structure-property paradigm for alloys. For example, data-driven modeling for grain precipitation and growth mechanisms. In a bigger picture, our practice here also shed light on the path that HPC becomes a powerful producer of simulation data and facilitates big data association with supercomputing.

Key words: alloy, Phase Field, In-situ feature extraction