1（College of Computer Science and Engineering,Shandong University of Science and Technology,Qingdao 266590,China）2（Shandong Computer Science Center（National Supercomputing Center in Jinan),Jinan 250101，China）
Abstract：Silicon-Crystal application uses molecular dynamics（MD）to simulate the thermal conductivity of crystals and uses Tersoff potential function to simulate silicon crystals growth.we successfully ported the Silicon-Crystal application on the Sunway TaihuLight using the Sunway Athread,and proposed five main optimizations for the problem caused by the memory constraints of SW26010 heterogeneous many-core processor：1）Prefetch the parameters required for the calculation to LDM；2）Transfer central particle data by DMA；3）Design the software cache reasonably and use the software cache to read the neighbor particle data；4）Customize the transcendental functions in CPE（Computing Processing Element）to avoid the discrete access when the CPE calls the transcendental functions；5）Use register level communication to realize step-to-step pipeline and double buffering between CPEs.After implementing these optimizations,the single-core group has achieved 12.89 times speed-up than the serial version in MPE（Management Processing Element),and 8.7 times speed-up than the serial version in the Intel Xeon E5-2620 v4 processor.This paper also conducted the scalability tests and analysis on the Silicon-Crystal application.Experimental results show that the Silicon-Crystal application has good scalability on the Sunway TaihuLight.
2021, Vol. 42 
