Declaration

A normal Finite Element code typically takes about 30-40 % of the total time to cal-culate and assemble the global element matrices. The focus of this work is on efficient implementation of Finite Element subroutines which perform this task. An implementa-tion which takes into account the underlying hardware architecture’s features results in achieving good performance. Without these considerations it is not possible to achieve a good portion of the peak performance on any machine. The major factors influencing performance are hardware architecture, programming language and the compiler. To get a good performance on cache based machines, it is necessary to understand the memory hierarchy of these machines. The methods to gain performance on scalar and vector architectures are studied. The various features of C and Fortran programming languages which effect performance are looked into. The problems faced by the C compilers to optimize the code due to aliasing are examined in detail. Due to large differences in architectural designs, a code which runs efficiently on one machine need not necessarily perform well on the other