PC 클러스터를 이용한 수치최적의설계의 병렬처리

Abstract

3차원 천음속 날개의 공력형상 설계최적화를 위하여 저렴한 비용의 PC 클러스터를 제작하였다. 클러스터는 3개의 노드로 구성되었으며, 리눅스 0S를 설치하고 메시지 전달방식의 통신 라이브러리인 MPI(Message Passing Interface)를 사용하였다. 이 시스템을 이용하여 24개의 설계변수를 가진 3차원 천음속 날개의 항력최소화 문제를 해석하였다. 유동장 해석을 위해 오일러 코드를 사용하였고, 일반적인 영역분할방식의 유동장 병렬처리를 택하는 대신에 수치최적화 알고리즘 중 민감도를 계산하는 부분을 병렬처리 하였다. 설계결과는 단일 노드를 사용한 순차계산의 결과와 동일하였으나, 설계 소요시간은 큰 폭으로 감소하였다. 본 연구의 결과로 저렴한 비용으로 우수한 성능을 갖춘 3차원 항공기 날개의 형상최

적설계 시스템을 구성할 수 있었으며, 향 후 이를 이용하여 보다 복잡하고 정교한 전산원용설계시스템으로 발전할 수 있는 기초를 마련하였다.

A cheap FC-Cluster was assembled for the aerodynamic design optimization of three dimensional transonic wing. The cluster has 3 nodes in which Linux Operating System and MPI as message passing library are installed. The system was applied to solve a wave drag minimization problem for transonic wing with 24 design variables. The Euler method was used to solve the flow around the wing, The process of sensitivity-derivative computation was parallelized in the numerical optimization algorithm, instead of using domain decomposition technique as usual parallel processing in the computational fluid dynamics. The design result with the parallel processing is exactly same as the one with the serial computation, while the parallel computation time was considerably reduced. The cheap shape optimization system based on the PC-cluster developed here could be a starting of the computer-aided design system with more sophisticated geometry generating module and flow analysis methods in the future.

A cheap FC-Cluster was assembled for the aerodynamic design optimization of three dimensional transonic wing. The cluster has 3 nodes in which Linux Operating System and MPI as message passing library are installed. The system was applied to solve a wave drag minimization problem for transonic wing with 24 design variables. The Euler method was used to solve the flow around the wing, The process of sensitivity-derivative computation was parallelized in the numerical optimization algorithm, instead of using domain decomposition technique as usual parallel processing in the computational fluid dynamics. The design result with the parallel processing is exactly same as the one with the serial computation, while the parallel computation time was considerably reduced. The cheap shape optimization system based on the PC-cluster developed here could be a starting of the computer-aided design system with more sophisticated geometry generating module and flow analysis methods in the future.