Objective:

Approach:

1. A parallel volume rendering API for time varying 3-D scalar fields in structured grids.
2. An X-window based GUI with multiple control panels for interactive control of visualization parameters and viewing positions.
3. A network interface connecting the renderer to the GUI and supporting image compression and various output devices.

Accomplishments:

• Added functional pipeline and out-of-core rendering capability.  The ParVox functions were separated into three functional modules, namely, the input module, the render module and the output module.    The three modules can be placed on different partitions of a single parallel machine or three different machines.  They are interconnected using MPI communication calls. (see Figure 1)  The advantage of functional pipeline is three fold: 1) increase the parallelism and the performance by pipelining the input, rendering, and compression operations, 2) easy to plug in additional input or render modules for new data formats or new rendering algorithms, and 3) allow distributed rendering on heterogeneous environment.   Out-of-core rendering capability was added into ParVox after the functional pipeline was in place.   For time-sequence animation, a specified subset of data can be prefetched into the memory at initialization.  The resident data will be swapped out and new data will be read in when the renderer is ready to render a dataset not currently resident in memory.   This out-of-core rendering capability allows ParVox to render 4D datasets of total size beyond the physical memory of the machine.  One limit applies to the current implementation:  each time step has to be loaded into memory as a whole.  If one time step does not fit into the physical memory of the parallel machine,  ParVox will not be able to handle it.

Figure 1.  The ParVox Functional Pipeline

• Ported a cell-projection renderer for unstructured grid dataset.   A new parallel volume rendering algorithm for 4D unstructured grid datasets was added into the ParVox visualization system.  The algorithm was based on Ma's cell-projection algorithm on IBM SP-2.  The algorithm was implemented using MPI1.2 communication API and was ported to the ParVox functional pipeline as an additional render module.  A new input module was also written for multiple variable unstructured grid datasets in NetCDF format.  The algorithm was modified and tuned to improve the performance in input, rendering, as well as output operations. Figure 2 is a rendered image of an unstructured grid finite element dataset provided by Tom Cwik.

Figure 2.  Unstructured Grid Data Rendering

• Accomplished HPCC Level 1 milestone "Demonstrate portable scalable distributed visualization of multi-terabyte 4D data sets on TeraFLOPS scalable systems".  The data used in this demo was 245 time steps of 1/6 degree global ocean circulation datasets generated by the COSIM (Climate, Ocean and Sea Ice Modeling) project at Los Alamos National Laboratory.    A 240 node HP Exemplar machine at Caltech was used to run the demo.   Neither the machine nor the dataset is  at Teraflop or terabyte level, but it is the biggest dataset and the largest accessible machine available at the time of the demo.  The data was stored at a remote data server and retrieved using Storage Resource Broker (SRB) in realtime .  The rendered images were sent back and displayed on SGI Oynx at JPL.  A full report on the milestone demonstration was prepared.  Three animation movies were generated during the demo: velocity(1.8Mbyte), temperature (800 Kbyte), and salinity (850 Kbyte).

Significance:

Status/Plans:

1. Port ParVox to SGI Origin and Beowulf.  MPI2.0 has not been widely accepted by the MPI community.  Currently not many vendors support MPI2.0 on their hardware.  Therefore, although ParVox is ported to MPI2.0, it can only run on the HP Exemplar system.  We plan to port ParVox to MPI1.2 point-to-point communication and test it on the SGI Origin and Beowulf PC clusters.  The JPL 128 node SGI Origin system and the 32 node Beowulf system owned by JPL High Performance Computing Group will be used as test machines.

1. Deliver ParVox 2.0 to the JPL's HPCC Software Repository.  The new release will include:

• MPI1.2 version of ParVox running on HP Exemplar, SGI Origin, and PC clusters
• Functional pipeline code supporting out-of-core rendering
• Bug fix and various functional enhancements from ParVox1.0
• Complete documentation

Point of Contact: