Abstract
본 연구에서는 빠르게 전자파 해석을 수행할 수 있는 병렬 유한차분 시간영역(Finite-Difference Time-Domain: FDTD) 알고리즘을 구현하기 위하여 CPU 클러스터를 구축하였다. 병렬 FDTD 알고리즘은 단일 프로세서를 이용한 FDTD 알고리즘에 비해 해석 시간을 크게 줄일 수 있으며, 전기적으로 매우 큰 구조물에 대한 전자파 해석도 가능하다. 본 연구팀에서는 CPU 클러스터 기반의 병렬 FDTD 알고리즘에서 요구되는 프로세스 간의 통신을 위해 MPI(Message Passing Interface) 라이브러리를 이용하였으며, 3차원 공간분할을 적용하여 프로세스 간의 통신 시간을 최소화하였다. 단일 프로세서를 이용한 FDTD 알고리즘 대비 CPU 클러스터 기반의 병렬 FDTD 알고리즘의 계산속도 향상도를 기본 모드와 하이퍼 모드에서 분석하였으며, 전기적으로 매우 큰 콘크리트 구조물의 전자파 해석을 하였다.
In this work, we construct a CPU cluster to implement a parallel finite-difference time domain (FDTD) algorithm for fast electromagnetic analyses. This parallel FDTD algorithm can reduce the computational time significantly and also analyze electrically larger structures, compared to a single FDTD counterpart. The parallel FDTD algorithm needs communication between neighboring processors, which is performed by the MPI (Message Passing Interface) library and a 3-D domain decomposition is employed to decrease the communication time between neighboring processors. Compared to a single-processor FDTD, the speed up factor of a-CPU-cluster-based parallel FDTD algorithm is investigated for the normal mode and the hypermode and finally analyze an electrically large concrete structure by the developed parallel algorithm.
Introduction
The Finite-Difference Time-Domain (FDTD) method is a cornerstone of electromagnetic (EM) simulation because it can model complex materials and provide wideband frequency data in a single run. However, it is notoriously memory-intensive. As the size of the structure being analyzed increases relative to the wavelength, the number of grid cells grows exponentially. For "electrically large" structures—such as full-sized concrete buildings—a single workstation quickly exceeds its RAM and processing capacity.
While parallel processing is a known solution, implementing it efficiently requires balancing the workload across multiple CPUs while minimizing "communication overhead"—the time CPUs spend exchanging data rather than calculating. This paper details the construction of a dedicated cluster and the optimization of a 3D parallel algorithm to overcome these scaling barriers.
System Architecture and Hardware
To address the computational demands of large-scale EM analysis, the research team constructed a cost-effective 15-node CPU cluster.
The hardware setup utilized 15 Dell OptiPlex 990SFF nodes connected via a 1,000 Mbps (Gigabit) Ethernet network. Each node featured an Intel Core i7-2600 quad-core processor and 16GB of RAM, resulting in a total of 60 physical cores (120 logical cores) and 240GB of system memory. This distributed architecture allows for the simulation of structures that would be impossible to process on a single machine.
3-D Domain Decomposition and MPI Implementation
A critical algorithmic optimization in this study is the transition from 1D to 3D domain decomposition. Instead of slicing the simulation space into 1D layers, the authors split the space into smaller sub-cubes.
This 3D approach is vital because it minimizes the surface area of the boundaries between processors, thereby reducing the volume of data that must be exchanged via the Message Passing Interface (MPI) library. To handle the edges of the simulation space where a processor might lack a neighbor, the team implemented "dummy tables" using MPI_PROC_NULL. This allowed for uniform communication code across all processors, avoiding slow conditional "if-else" statements for edge cases.
Results and Performance Analysis
The performance of the cluster was validated using a $200 \times 200 \times 200$ cell free-space benchmark and a large-scale $4m \times 4m \times 4m$ concrete structure.
- Significant Speedup: Using 60 physical cores, the system achieved a 51.76x speedup compared to a single processor. Scaling remained remarkably linear up to 60 cores.
- Hyper-threading vs. Physical Cores: Testing revealed that while "Hyper Mode" (using logical cores) increases the core count, it does not provide a linear speedup for FDTD. The bottleneck often lies in physical execution units or memory bandwidth rather than thread availability.
- Real-world Application: The cluster successfully simulated the EM characteristics of a large concrete structure, capturing complex phenomena like attenuation, scattering, and resonance that require high-resolution distributed memory.
Conclusion
This research demonstrates that a high-performance EM simulation environment can be built using standard workstation components and optimized parallel algorithms. While the current 1,000 Mbps Ethernet interconnect eventually limits the speedup curve for very high core counts, the 3D domain decomposition strategy significantly mitigates this bottleneck. Future work will apply this framework to even more complex scenarios, including human body EM exposure models and aircraft analysis.
Summary updated