A publication of the Advanced Simulation & Computing Division, NA-121.2, NNSA Defense Programs
March 2009
NA-ASC-500-09—Issue 10
Return to this issue’s stories
Application-Level Services Enable In Situ Analysis of Shock Physics Codes
The ability to perform meaningful, real-time analysis and visualization of scientific data, either simulated or real, is a significant advance that has the potential to fundamentally change the way computational scientists work. In situ analysis provides immediate feedback that allows the scientist to quickly detect bugs, identify areas of interest, or interactively modify the behavior of a running application. Such a capability will dramatically improve productivity of the application scientist and significantly reduce time-to-solution for time-critical ASC applications.
In a joint effort between the Scalable Data Analysis and Scalable I/O groups of the ASC Computational Systems and Software Environments (CSSE) project, researchers at Sandia National Laboratories are exploring the use of “Application-Level Services” for in situ analysis of the shock physics code CTH. An application-level service is a separate application that executes on either compute nodes or service nodes of an HPC system. In contrast to system-level services (e.g., file services), an application-level service is dedicated to a single application.
The application-level service for CTH detects and tracks material fragments generated by the simulation of high-speed impacts. Fragment detection and tracking provide valuable insight to the scientist, but present a number of computational challenges. First, fragment tracking is data intensive because it requires data from every time-step calculation. The I/O requirements (both capacity and write bandwidth) of performing fragment tracking as an off-line task make this option impractical. Second, fragment detection is computationally expensive. If integrated with the CTH code, fragment detection adds as much as 30% to each simulated time-step calculation, significantly impacting scalability of the application. In addition, integrating detection into the CTH code adds complexity and significant programming burden to the application developer.
The application-level service to detect and track fragments executes on a separate partition of compute nodes on the HPC system and leverages technologies developed by the Lightweight File Systems project and Paraview, two ASC-funded projects. Offloading fragment detection to a separate partition of compute nodes provides several advantages: it allows concurrent CTH computation and fragment detection; it significantly reduces the I/O to storage by outputting only the fragment information; and, perhaps most importantly, it does not require major modifications to the CTH code. The application-level service operates on requests generated by the spyplot I/O API (the API that is already used by CTH).
The application-level service approach represents a significant advance in the ability to perform in situ analysis of scientific data and will have a broad impact in the greater computational-science community. Our work with CTH demonstrates the first significant use of this technology and will serve as a model for future efforts with other codes.
The fragment detection service provides on-the-fly data analysis with no modifications to CTH.
DOE Privacy Disclaimer | Sandia Privacy Disclaimer |
Developed and maintained by Sandia National Laboratories for NA-121.2