Interactive visualization of 3D mantle convection

Abstract

The current availability of thousands of processors at many high performance computing centers has made it feasible to carry out, in near real time, interactive visualization of 3D mantle convection temperature fields, using grid configurations having 10–100 million unknowns. We will describe the technical details involved in carrying out this endeavor, using the facilities available at the Laboratory of Computational Science and Engineering (LCSE) at the University of Minnesota. These technical details involve the modification of a parallel mantle convection program, ACuTEMan; the usage of client–server socket based programs to transfer upwards of a terabyte of time series scientific model data using a local network; a rendering system containing multiple nodes; a high resolution PowerWall display, and the interactive visualization software, DSCVR. We have found that working in an interactive visualizastion mode allows for fast and efficient analysis of mantle convection results.

Appendix A

It has been shown in previous studies that taking advantage of network speeds has been shown to decrease the overall computation time for scientific models (Zhou et al. 2006). A natural consequence of interactive visualization is the reduction in wall-clock time. In our case, it takes less wall clock time to visualize our ACuTEMan results immediately because we have been able to reduce the amount of data that is written to local disk on the compute nodes. We find it sufficient to store only partial data sets at specified intervals, as opposed to archiving the output of an entire simulation. Most important are the final conditions, or the state of the model at the last time step. We can restart the model using these conditions as the initial ones and rerun a partial case in the event that the data becomes of interest. Saving the final conditions of our model allows us to start a simulation from various times of interest when it becomes necessary to re-analyze the data and saves us from having to store large data sets at each time increment.

The decrease in computation time is a result of both minimizing the frequency of local data writes and also utilizing the local network for data output. In an interactive mode, data is transferred close to the theoretical maximum rate of the network and not at the rate of the local disk. The data is transferred from the local RAM on the nodes of the Blade system to the ram disk on the LCSE rendering system through the InfiniBand interface. Once the ram disk on the LCSE system is inundated with data, the speedup would no longer be observed. When the grid is increased to a large enough size, the capacity of the LCSEs RAM disk would be reached, thus filling the memory and slowing down the overall visualization and computation performance. We want to be careful to find a balance between doing real-time visualization and over-exerting the resources currently in place at the LCSE. The data throughput rate can be improved by adding more nodes to the LCSE system for accepting incoming transfers from multiple nodes on the Blade system.