Data Intensive Volume Visualization on the Tera Mta and Cray T E

Raycasting is a DVR Direct Volume Rendering tech nique DVR is a class of techniques for graphically visualizing D data The data can be from any source as long is it is representable by discrete sample values at each point in a three dimensional grid Common sources of such data include CT Computed Tomog raphy MRI Magnetic Resonance Imaging and ul trasound scans The data can come from simulation or direct measurement of physical phenomena Direct volume rendering is by nature highly memory inten sive For example a high resolution representation of a human female The Visible Female dataset is gigabytes A raycasting algorithm must typically visit each byte of this data to produce an image A further challenge is to optimize DVR algorithms for data locality Desktop workstations do not currently have the memory or compute power to visualize such large datasets quickly Yet quick response to user requests for view changes are important to allow in teractive exploration We describe a software system the MPIRE Massively Parallel Interactive Rendering Environment developed at SDSC San Diego Super computer Center which brings the power of a super computer to the desktop via the Web and allows in teractive explorations of massive datasets Currently MPIRE runs on two very di erent supercomputers the Cray T E and the MTA Tera Multithreaded Ar chitecture We compare and contrast the architec tures of these machines and characterize the porting e ort and performance of MPIRE on each Introduction Complex physical phenomena are often best investi gated by visual inspection There is no substitute for seeing for ones self the contours of the problem under scrutiny whether it be the ocean oor the human brain or the reactive surface of a molecule Today computers aid in the work of science by turn ing masses of numbers into pictures on the screen If view changes can be completed quickly the scientist can interact with the data in real time virtually skim ming the ocean oor slicing through a diseased or gan or inspecting the shape of a protein docking site However when the dataset to be visualized is large the practical challenge of turning data bytes into im age pixels can be immense Speci cally changing the view point and forming a new image of a multi gigabyte dataset in under seconds is beyond the current capabilities of most desktop system Clearly powerful supercomputers such as the Cray T E and the MTA Tera Multithread Architecture provide an alternative But most researchers do not have a su percomputer sitting next to them What they do have in many cases is a PC or workstation run ning Netscape with Java Using SDSC s San Diego Supercomputer Center s MPIRE Massively Parallel Interactive Rendering Environment software this is enough to deliver the power of the supercomputer to the desktop The MTA and the T E embody very distinct ap proaches to supercomputing The machines are dif ferent architecturally and this leads to di erent pro gramming styles The T E re ects current main stream thinking on how to build a supercomputer from commodity components The MTA is an exper imental machine which is constructed from custom components At SDSC we are tasked with evaluating the Tera MTA approach While we have studied the MTA s performance on a number of benchmarks and applications Carter et al this paper presents our rst evaluation of a full user application on the