Sarnoff Participation

Network-Based Visualization and Analysis

The purpose of Sarnoff's participation in the initial phase of the Matisse project is to implement network-based visualization and analysis components for use over Supernet. Specifically, we intend to implement a variety of software modules to support data access from remote, high-performance storage systems, data analysis and visualization with remote, high-performance computing clusters and high-quality, real- time visual feedback. The following sections describe, in turn, the data types which will be supported, the hardware required and the visualization and analysis functions which will be developed.

Data Types

The metrology tools which we will be working with during the initial phase of the project (Computer Microvision, Remote Microscope) produce, as their basic data type, imagery. Because of this, the fundamental unit of data which will be supported is a 2-D image. Various collections of 2-D imagery will also be defined, including:

• Computer Microvision 4D volumes, comprised of stacks of images in time (i.e. x,y,z,t)
• High-Speed Tomography 3D volumes, comprised of isotropic 3D images (i.e. x,y,z)
• High-Speed Radiography 3D volumes, comprised of sequences of 2D images (i.e. x,y,t)
• Video streams of generated imagery

Hardware

Numerous analysis and visualization software modules have been created at Sarnoff and employed for use with actual data over the course of the past several years. These software modules have the capacity to employ parallelism to handle the large quantities of data anticipated in this project. We intend to develop a mid-size, 8 PC cluster for the purpose of creating a computation facility for MEMS data. This PC cluster will leverage supercomputing technology developed by another group at Sarnoff, which has developed a large-scale, 256 processor cluster termed the Cyclone. All software modules developed for the mid-size cluster will also function on the Cyclone or any other similar-technology PC cluster.

Visualization and Analysis Functions

During the initial phase of the project, we intend to port the following existing visualization and analysis software modules:

• Multi-dimensional image processing and permutation
• Rigid-body motion estimation
• Voxel-based volume rendering

The essential computational aspects of these modules have already been designed, implemented and tested on Silicon Graphics workstations and single Linux PCs. They are currently in active use by researchers at both Sarnoff and MIT. The porting of the modules for this project will involve the modification of data access functions to interface with the high-speed cache at Lawrence Berkeley National Laboratories (LBNL), the necessary modifications to support Silicon Graphics functionality on PCs, the parallelization of compute-intensive inner loops across the PC cluster and the assembly of the individual cluster node outputs into a real-time video stream. These last two steps are non-trivial in the sense that the goal of our work is to maximize the usage of the network pipe which connects each component in the system. Thus, the parallelization is both across processing and data access. The end result of this work should support computational tasks which fill a Gigabit network and display functions which run at 200-300Megabits per second.