研究目的
To overcome the loss of parallel efficiency in parallelized lattice kinetic Monte Carlo simulations by introducing a dynamical space partitioning method for better load balancing.
研究成果
The dynamical space partitioning method improves parallel efficiency in lattice kMC simulations by enabling better load balancing, as confirmed in hypothetical and practical cases. However, additional optimizations are needed for full effectiveness in real-world applications.
研究不足
The method may not be effective when active sites are localized on planes parallel to the slice planes, requiring slices thicker than the diffusion jump distance (about 5 bond lengths). In practical cases, data structure optimization and implementation of the 'TimeStop' approach are not yet complete, limiting efficiency improvements. The period of repartitionings and tstop parameters need further optimization.
1：Experimental Design and Method Selection:
The study involves designing a dynamical space partitioning method for parallelized lattice kMC simulations to improve load balancing. Two existing parallelization approaches ('TimeStop' and 'OneStep') are used and modified with the new partitioning method.
2：Sample Selection and Data Sources:
A hypothetical dopant diffusion simulation with a 512x64x63 simple cubic lattice and a practical epitaxial silicon growth simulation in a 54x90 nm3 cell for FinFETs are used.
3：List of Experimental Equipment and Materials:
An 18-core Intel Xeon Gold 6154 CPU is used for computations.
4：Experimental Procedures and Operational Workflow:
Simulations are run with equi-space, initial-only, and dynamical partitioning methods. For dynamical partitioning, repartitioning is performed at specific time intervals to equalize total rates. The numactl Linux command is used to bind OpenMP threads to cores.
5：Data Analysis Methods:
Computational times and the number of executed events are compared between serial and parallel simulations to evaluate parallelization efficiency.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
