Production visualization for the ASCI One TeraFLOPS machine

Abstract

The delivery of the first one tera-operations/sec computer has significantly impacted production data visualization, affecting data transfer, post processing, and rendering. Terascale computing has motivated a need to consider the entire data visualization system; improving a single algorithm is not sufficient. This paper presents a systems approach to decrease by a factor of four the time required to prepare large data sets for visualization. For daily production use, all stages in the processing pipeline from physics simulation code to pixels on a screen, must be balanced to yield good overall performance. Performance of the initial visualization system is compared with recent improvements. "Lessons learned" from the coordinated deployment of improved algorithms also are discussed, including the need for 64 bit addressing and a fully parallel data visualization pipeline.