Optimization of Deepvariant Software for Domestic Heterogeneous Platforms
Gaotian Meng, Lin Han, Yingying Li, Hao Wu · 2024
The demand for genomic data analysis has increased, and the development of domestic supercomputing platforms. The purpose of this transplantation and optimization work is to make full use of local high-performance computing resources, improve the efficiency and accuracy of genomic data analysis, and promote scientific research and application progress in related fields. This study focuses on the porting and optimization of DeepVariant software for domestic heterogeneous platforms, addressing issues including long runtime, low efficiency, inability to run on the platforms, and low degree of parallelization on traditional platforms. DeepVariant utilizes deep neural networks (DNNs) to call genetic variants and detect mutation sites. Originally tested on a workstation with GPU devices for efficiency, we analyzed the performance of DeepVariant and ported it to domestic DCU accelerators, accompanied by code refactoring. Optimization strategies including CPU parallelization, memory access optimization, and compilation option optimization were employed to reduce the runtime and cost of DeepVariant. Experimental results using different datasets demonstrate that the optimized DeepVariant achieved 1.2-2.2 times speedup compared to existing platforms. The optimization methods employed in this study significantly enhance the efficiency and performance of DeepVariant software.