Burkhard Zink

Distinguished Senior Postdoctoral Fellow
Center for Computation & Technology

Johnston Hall
Louisiana State University
Baton Rouge, LA 70803

Tel: +1 (225) 578 0046
Fax: +1 (225) 578 5362
Email: bzink@cct.lsu.edu

Curriculum Vitae
GPU Computing and High-Performance Computing |--| General Relativistic Astrophysics |--| Ray-Tracing and Visualization |--| Publications

Research Interests

GPU Computing and High-Performance Computing

High performance computing is the field of studying scientific and engineering problems by means of large computational systems, usually orders of magnitude more powerful than a regular desktop computer. The computing performance is usually achieved by massive parallelism, i.e. the execution of a program on many processors at once, which brings unique requirements for algorithmic expression and implementation.

Desktop computers, on the other hand, have traditionally been restricted to single CPUs for economic reasons, and have therefore been addressed with serial programming models (although modern CPUs support a certain level of internal parallelism). However, there is another component in some recent desktop computers which uses massive parallelism for high performance computation: the graphics processing unit (GPU).

GPUs are enormously powerful: While a regular dual-core CPU has a peak performance of roughly 20 GFlop/s, a modern GPU delivers about 350-500 GFlop/s. In addition, GPU performance has increased much more rapidly than CPU performance during the last years. Modern programming tools like NVIDIA CUDA now admit to use GPUs for general parallel processing, pushing GPUs directly into the field of high-performance computing.

I have recently become interested in using GPU accelerators for scientific computation and visualization, and I am driving activities at CCT to enable the efficient use of GPUs in hybrid many-core/GPU distributed clusters within the Cactus middleware. We have entered into a close collaboration with NVIDIA Corporation and NCSA to develop scalable parallel models, algorithms and tools, and will be organizing joint workshops to explore these matters.

An experimental implementation of one of our scientific codes shows exciting results: The massively parallel architecture of an NVIDIA Quadro FX 5600 admits us to accelerate our code by up to a factor of 26.5 over a CPU compute core! The details of this result have been published in a Technical Report, see below.

General Relativistic Astrophysics

Investigating the properties of black holes, neutron stars and other extreme astrophysical objects requires a description of fluids in terms of general relativity (relativistic CFD). However, to actually obtain quantitative results, large-scale simulations are a necessity. These kind of projects are among the most resource-intensive computations done on modern high-performance clusters, and nowadays require as much attention to the physical motivation as to issues in parallel algorithms, software engineering, data handling and visualization.

Research in this exciting and complex field was the main part of my PhD thesis in Garching, where I dealt with the question how black holes can form from gravitational instabilities in supermassive stars. At CCT, I am leading an effort to describe the fluid flow in terms of a so-called multi-block infrastructure, which distributes the computational domain into several distinct coordinate patches. This approach has advantages in terms of the numerical description, since it is better adapted to approximate symmetries of the physical system without introducting coordinate singularities.

However, multi-block approaches also admit to formulate parallel algorithms which are conceptually easier to scale to 100,000s of computational cores than adaptive mesh refinement, another popular approach. Processor counts like this are the order of magnitude expected to be present in future petascale machines, i.e. those with a peak performance exceeding 1 PFlop/s.

Ray-Tracing and Visualization

Visualization is usually concerned with the representation of large datasets in scientific and engineering problems, the enhancement of particular features, and the interactive control of display parameters. In these representations, realism is secondary to ease of use and expressive content.

However, I have been working on a different aspect of visualization, the approximately realistic imaging of astrophysical objects by means of volume ray-tracing. Ray-tracing is a well-known technique in computer graphics to construct synthetic images: virtual rays are emitted from a projection point into an analog of the screen plane to produce individual pixel colors, by determining the intersection of the ray with scene objects. Various algorithms exist to increase the realism of the produced images, increase computational efficiency, and reduce sampling artifacts.

In the context of relativistic astrophysics, rays appear as curved lines and need to be integrated numerically. The interaction of rays with objects like accretion disks around black holes is then governed by a general relativistic rendering equation which transports spectral data under consideration of relativistic effects like Doppler shift and beaming.

Currently, I am also collaborating with Shalini Venkatamaran from the CCT Graphics Group on on-demand visualization techniques in high-performance computing using GPUs.

Publications