Accelerating cryo-EM Reconstruction of RELION on the New Sunway Supercomputer.

Over the long journey of understanding the micro-world, cryo-EM has become an effective technique for biomolecular structure determinations. However, due to the complex algorithmic features and large amounts of computing data, sophisticated HPC solutions are in urgent demand. In this paper, we present our efforts of porting RELION to the new generation of Sunway supercomputer. Optimizations that fit well with the new hardware architecture have been proposed, including a multilevel parallel scheme that smartly maps and scales RELION onto the novel Sunway architecture, optimizations that address memory bottlenecks and improve memory efficiency, and a pipeline approach that obtains excellent computation and communication overlapping. Combining all proposed optimizations, the calculation time for one iteration is greatly reduced from 7,577 seconds to 2,017 seconds, with a speedup of 3.757x. The overall design is scaled to over 131,072 cores with a parallel efficiency of 95%.