Keita Teranishi, Research Associate, Penn State University
Keita Photo

Keita Teranishi, Ph.D.

teranish@cse.psu.edu
350 IST University Park, PA 16802 U.S.A

I moved to Cray Inc..

My research interest is high performance scientific computing, including parallel matrix and graph algorithms, unstracture or sparse schemes, and software and tools to enable large scale simulations in computational science and engineering.

Recently, I am working on software and tools for computational materials science. This project includes design of e-Science web portal, data compression/management of materials data in nano and microscopic scales, and computational schemes for microscopic scale simulations.

I obtained BS and MS degrees at Universiy of Tennessee, Knoxivlle and Ph.D. degree at the Pennsylvania State Univeristy.

DSCPACK-IC

A Parallel Direct-Iterative Hybrid Solver Through Flexible Incomplete Cholesky Preconditioning

Author: Keita Teranishi and Padma Raghavan

DSCPACK-IC is a general-purpose parallel drop-tolerance incomplete Cholesky preconditioner for the Conjugate Gradients iterative solution. This package is suitable for memory-efficient solution for systems where the coefficient matrix is symmetric, positive definite and sparse. The amount of nonzero entries in the preconditioner can be adjusted using the drop tolerance condition to model the range from nearly a pure iterative method to a pure direct method. Although the software supports this wide range of preconditioning, we expect it to perform best when the preconditioner allows a reasonable fraction (10% or more) of the fill-in for the true sparse factor. This preconditioner is written in C; it uses MPI for inter-processor communication and provides an interface for KSP (Krylov Subspace Solvers) solvers in PETSc package. The implementation of this package is based on the DSCPACK direct solver to compute parallel ordering and symbolic factorization to support the fully parallel construction of the preconditioner and its application.


Traditionally, preconditioning through incomplete factors (though widely accepted as the method of choice on serial computers) has been considered infeasible for multiprocessors and networks of workstations. This is primarily because the large latencies of interprocessor communication on such multiprocessors make applying the preconditioner using parallel substitution very inefficient. ...More

Research Interests

Sparse Matrix Computation

In many scientific and engineering applications, their computational time is dominated by solution of linear systems where the coefficient matrix is sparse. Therefore, efficient solution for such linear systems is very important.

Parallel/Distributed Computing

Because of the performance limitation of single computers, many scientific and engineering problems need to be solved with multiple computers. Their computational power can be effectively derived only if appropriate algorithm, software and tools are devised.

Computational Materials Science and Engineering

I have been working on software and tools for computational materials science and engineering. My project aims at enabling materials scientists to study materials in different scales through web-browsers to orchestrate simulations on remote computers.

Scientific Data Mining and Compression

Classic scientific computing disciplines have emphasized on computation intensive applications in order to simulate large scale problems. These applications also produce a huge amount of data, and thus the management of such data is very critical to understand the result of simulations.

Copyright © 2006 Keita Teranishi. All rights reserved.