Publications

Large-Scale Data Analytics (LSDA) problems require finding meaningful patterns in data sets that are so large as to require leading-edge processing and storage capability. LSDA problems are increasingly important for government mission work, industrial application, and scientific discovery. Effective solution of some important LSDA problems requires a computational workload that is substantially different from that associated with traditional High Performance Computing (HPC) simulations intended to help understand physical phenomena or to conduct engineering. While traditional HPC application codes exploit structural regularity and data locality to improve performance, many analytics problems lead more naturally to very fine-grained communication between unpredictable sets of processors, resulting in less regular communication patterns that do not map efficiently on to typical HPC systems. In both simulation and analytics domains, however, data movement increasingly dominates the performance, energy usage, and price of computing systems. It is therefore plausible that we could find a more synergistic technology path forward. Even though future machines may continue to be configured differently for the two domains, a more common technological roadmap between them in the form of a degree of convergence in the underlying componentry and design principles to address these common technical challenges could have substantial technical and economic benefits.

Abstract not provided.

Abstract not provided.

Abstract not provided.

Abstract not provided.

Abstract not provided.

Abstract not provided.

Abstract not provided.

Abstract not provided.

Abstract not provided.

Abstract not provided.

Abstract not provided.

Abstract not provided.

Abstract not provided.

Abstract not provided.

A study was undertaken to validate the 'capability' computing needs of DOE's Office of Science. More than seventy members of the community provided information about algorithmic scaling laws, so that the impact of having access to Petascale capability computers could be assessed. We have concluded that the Office of Science community has described credible needs for Petascale capability computing.

Abstract not provided.

This paper analyzes what additional costs would be incurred in supporting dual-mode, i.e. both classified and unclassified use of the Institutional Computing (IC) hardware. The following five options are considered: periods processing in which a fraction of the system alternates in time between classified and unclassified modes, static split in which the system is constructed as a set of smaller clusters which remain in one mode or the other, re-configurable split in which the system is constructed in a split fashion but a mechanism is provided to reconfigure it very infrequently, red/black switching in which a mechanism is provided to switch sections of the system between modes frequently, and complementary operation in which parts of the system are operated entirely in one mode at one geographical site and entirely in the other mode at the other geographical site and other systems are repartitioned to balance work load. These options are evaluated against eleven criteria such as disk storage costs, distance computing costs, reductions in capability and capacity as a result of various factors etc. The evaluation is both qualitative and quantitative, and is captured in various summary tables.

Abstract not provided.

Finite elements are routinely used for analysis of real world problems in a wide range of engineering disciplines. The types of problems for which these are used include, but are not limited to, structural engineering, materials science, heat transfer, optics and electromagnetics. While linearity is a good assumption to start with in many problems, reasonable solutions to real-life problems require them to be treated as non-linear. It is, therefore, necessary that the users of finite element codes be aware of the capabilities and limitations of their analysis tools.

The authors developed and implemented a highly parallel computational algorithm for solution of the inverse scattering problem generated when an integrated circuit is illuminated by laser. The method was used as part of a system to measure diffraction grating line widths on specially fabricated test wafers and the results of the computational analysis were compared with more traditional line-width measurement techniques. The authors found they were able to measure the line width of singly periodic and doubly periodic diffraction gratings (i.e. 2D and 3D gratings respectively) with accuracy comparable to the best available experimental techniques. They demonstrated that their parallel code is highly scalable, achieving a scaled parallel efficiency of 90% or more on typical problems running on 1024 processors. They also made substantial improvements to the algorithmics and their original implementation of Rigorous Coupled Waveform Analysis, the underlying computational technique. These resulted in computational speed-ups of two orders of magnitude in some test problems. By combining these algorithmic improvements with parallelism the authors achieve speedups of between a few thousand and hundreds of thousands over the original engineering code. This made the laser diffraction measurement technique practical.

We describe an algorithm for the static load balancing of scientific computations that generalizes and improves upon spectral bisection. Through a novel use of multiple eigenvectors, our new spectral algorithm can divide a computation into 4 or 8 pieces at once. This leads to balanced partitions that have lower communication overhead and are less expensive to compute than those of spectral bisection. In addition, our approach automatically works to minimize message contention on a hypercube or mesh architecture.

We describe an algorithm for the static load balancing of scientific computations that generalizes and improves upon spectral bisection. Through a novel use of multiple eigenvectors, our new spectral algorithm can divide a computation into 4 or 8 pieces at once. These multidimensional spectral partitioning algorithms generate balanced partitions that have lower communication overhead and are less expensive to compute than those produced by spectral bisection. In addition, they automatically work to minimize message contention on a hypercube or mesh architecture. These spectral partitions are further improved by a multidimensional generalization of the Kernighan-Lin graph partitioning algorithm. Results on several computational grids are given and compared with other popular methods.

The point of this work is to create a tool for thorough analysis of the eigenspectrum of the linearized finite element equations for buoyancy driven flow.