PERFORMANCE OF HYPERPLANE ORDERING ON VECTOR COMPUTERS

[摘要] We consider the hyperplane ordering vectorization with the least effect of memory bank conflict on vector computers. The hyperplane ordering is often used in the incomplete Cholesky conjugate gradient (ICCG) method in order to solve systems of linear equations which result from discretization of partial differential equations. We show that the performance of hyperplane ordering on currently used vector computers depends upon the magnitude of the access stride to memory and the total count of the access strides. We estimate precisely the magnitude of the access stride to memory and its total count which are determined by the numbers of gridpoints in the three directions of 3D domains. We give a recipe on how to choose the numbers of gridpoints in the x-, y-, z-directions for efficient solution of 3D problems by the ICCG method.

[发布日期] 1991-12-23

[效力级别]  Proceedings Paper

[关键词] HYPERPLANE ORDERING;INCOMPLETE CHOLESKY CONJUGATE GRADIENT METHOD;MEMORY BANK CONFLICT;VECTOR COMPUTERS