We propose to develop and empirically evaluate the software tools,analysis techniques,and compiler optimizations necessary to generate automatically a customized Linux image,and to affect the necessary ap- plication transformations that use it.Moreover,we will quantify the success of our techniques using scienti .c applications and software prototypes executing on commodity cluster platforms.We will employ parallel and .oating-point intensive codes as well as scientic database programs.We will focus on execution platforms viable at the high-end .clusters of SMP,IA32 and IA64 systems .the latter allowing us to investigate relevance to NSF's current TeraGrid [65 ]effort.The result,we believe,will be a software system that auto- matically enables high-performance scientic computing on commodity systems through application-specific customization and dynamic adaptation of a low-cost,popular,and familiar Linux operating system.
