Sun Microsystem's AutoClient and management of computer farms at BaBar

Abstract

Modern HEP experiments require immense amounts of computing power. In the BaBar experiment at SLAC, most of it is provided by Solaris SPARC systems. AutoClient, a product of Sun Microsystems, was designed to make setting up and managing large numbers of Solaris systems more straightforward. AutoClient machines keep all filesystems, except swap, on a server and employ CacheFS to cache them onto a local disk, which makes them Field Replaceable Units with performance of stand-alone systems. We began exploring the technology in Summer 1998, and currently operate online, reconstruction, analysis and console AutoClient farms with the total number of nodes exceeding 400. Although the technology has been available since 1995, it has not been widely used, and the available documentation does not adequately cover many important details of AutoClient installation and management. This paper discusses various aspects of our experience with AutoClient, including tips and tricks, performance and maintainability, scalability and server requirements, existing problems and possible future enhancements.