Since September, '99 Tempest has been our department's primary, general-purpose Beowulf system. Usage has grown dramatically during the last few months as more department members have begun using Tempest to run research jobs previously run on other high-performance computer centers. What started out as a curiosity has now turned into the most heavily used computing resource in the department. Tempest routinely handles approximately 30 parallel and serial jobs per day in areas ranging from astrophysics and general relativity, to condensed matter and fluid dynamics. This system has already assisted two Ph.D students (with a third to soon follow) to complete their discertation work ahead of schedule while giving them the opportunity to apply parallel programming skills to real research problems.
Tempest was assembled entirely within our department from individually chosen commodity hardware components. With the exception of the ethernet switch, all of the computer hardware within the cluster was built from the component level. This low-level approach allowed the developers to optimize the overall performance by providing us the freedom to choose the fastest and most reliable components to fit our budget. Likewise, all matters of OS and software package configuration were carried out in-house in an effort to achieve optimal settings for speed and stability.
