PI Dr. Paul A. Navrátil is a Research Associate and the Manager of the Scalable Visualization Technologies group at the Texas Advanced Computing Center (TACC). His research interests include efficient algorithms for large-scale parallel visualization and data analysis (VDA) and innovative design for large-scale VDA systems. He is an expert in high-performance visualization technologies and advanced rendering techniques, including algorithms for large-scale distributed-memory ray tracing. Dr. Navrátil holds BS, MS and Ph.D. degrees in Computer Science and a BA in Plan II Interdisciplinary Honors from the University of Texas at Austin.
PI Dr. Berk Geveci is the Senior Director of Scientific Computing at Kitware, Inc. where he provides technical leadership for Kitware’s scientific and information visualization teams. Dr. Geveci’s team of thirty scientific computing professionals work closely with an extended community of visualization experts, including members of National Labs, DOD research laboratories, and supercomputing centers throughout the world. He is the principal developer of many key VTK components, including the SMP infrastructure, and is a member of the VTK Architecture Review Board. Dr. Geveci completed his Ph.D. at Lehigh University in 1999.
High-performance computing (HPC) is increasingly important in scientific research, engineering design, and modern manufacturing. Unfortunately deploying HPC at scale is difficult due to the complexity of the programming model, and the lack of reference implementations with which to educate and accelerate the creation of new software solutions. Significant efforts are now under way to reduce these barriers, for example the recent introduction of the Intel Xeon Phi™ architecture shows great promise towards addressing the complexity challenge. However an important next step is to deploy example applications impacting a broad swath of scientific and commercial applications. In this proposal we advocate integrating Intel Xeon Phi™ into the VTK/ParaView platform, a general, open-source software system that is widely used across an extensive range of application areas ranging from engineering simulation to medical computing to climate modeling.
The ParaView large-data visualization application, which has been successfully demonstrated on large-scale HPC facilities (e.g., hundreds of thousands of processors at Argonne) is built on the Visualization Toolkit (VTK). In turn VTK provides fundamental data processing, analysis, and visualization capabilities that are HPC capable. VTK/ParaView has been in continuous development since 1993 using an agile open-source development model, with hundreds of millions of dollars of effort going into the system (based on COCOMO analysis). Because it is open-source, this platform is a superb demonstration vehicle for new computing technologies, since developers can study and modify code to suit their purposes, and the system is readily deployed and used around the world.
By integrating VTK/ParaView with Intel Xeon Phi™, significant impacts will be made across a wide variety of computing applications. Routinely ParaView is used at most prestigious HPC centers in the service of simulation (materials, fluids, nuclear physics, climate), and large-data analysis (e.g., cosmology, climate, high energy physics). It is utilized in medical research (brain shape analysis), hydrological modeling (Chesapeake Bay watershed analysis), and LiDAR capture (point-cloud library PCL visualization). VTK also serves as the foundation of a number of data analysis and visualization systems addressing science and engineering (e.g. CD-Adapco’s Star-CCM+, DOE’s VisIt, the Salome Platform, MayaVi) and health care (e.g. OsiriX, 3D Slicer). Thus deploying the Phi technology into VTK/ParaView will place it in a leadership role for the application of HPC to meaningful computational problems.
To realize the overarching aim of integrating Intel Xeon Phi™ into VTK/ParaView, we will optimize the system to support Phi’s wide vector processing capability, and take advantage of the large number of computing cores. In the work proposed here, we will focus on the vector processing capability, and leverage separate work (via a DOE grant) for many-core support. We will develop and demonstrate support for both coarse and fine-grained parallelism, and release open-source software demonstrating the use of these capabilities in a variety of scientific computing applications. Finally, we will also demonstrate that these optimizations will significantly increase the performance of VTK/ParaView on the Intel Xeon® processor family due to better leveraging of the multi-core and vector capabilities of these processors.