All the research work I do gets implemented in software and gets
delivered to different applications. The software I write are delivered free
of charge, with popular open-source licenses. Lately, all my work has been
delivered through the popular Trilinos framework. I lead several efforts in
Trilinos, especially in the solvers and kernels areas.
I lead two linear solver related packages in Trilinos - ShyLU and Amesos2. I also lead a node-level kernels package called Kokkos Kernels.
ShyLU (Scalable Hybrid LU, pronounced Shy-Loo) was
originally developed as a hybrid proconditioner/solver. ShyLU uses the hybrid
(direct + iterative) approach to be robust and MPI + Threads to be a scalable
solver. ShyLU is used in Xyce circuit
simulation package at Sandia. ShyLU also includes a number of other features
for threaded factorizations such as threaded LU factorization, threaded
incomplete Cholesky factorization, and iterative ILU factorizations.
See ShyLU home
page for more details.
Amesos2 is an interface to different direct
solvers in Trilinos. Amesos2 is designed to work with the next generation
Trilinos stack, with interfaces to multiple direct solvers. See Amesos2 home page for more details.
Kokkos Kernels is package to deliver performance portable
kernels on top of the Kokkos library. It supports the multicore and manycore
architectures that are expected to be the architectures for future
supercomputers. Kokkos Kernels supports sparse/dense linear algebra kernels
and it supports graph kernels on next-generation architectures.
I have also contributed to other Trilinos packages such as IFPACK, IFPACK2 and Tpetra at different times.
Zoltan and Zoltan2, are toolkits for parallel combinatorial
algorithms like load balancing and dynamic partitioning, ordering and
coloring. My work related to combinatorial algorithms for scientific
computing and graph algorithms for data analytics get implemented in
See Zoltan home page and Zoltan2 home page for more details.
CHOLMOD/CCOLAMD generates constraint preserving ordering that reduces fill-in and requires fewer floating point operations in sparse Cholesky and sparse LU factorizations. It is based on COLAMD. CCOLAMD is part of CHOLMOD.
PIRO_BAND reduces symmetric/unsymmetric band matrices to tridiagonal/bidiagonal form. PIRO_BAND supports double and single precision arithmetic for real and complex matrices in architectures with 32-bit and 64-bit integers. PIRO_BAND can also compute the SVD for a band matrix.
PIRO_SKY finds the singular value decomposition of sparse matrices. PIRO_SKY uses a sparse QR factorization and sparse tridigonalization of the R from the QR factorization to compute the SVD of the sparse matrix.
(505) 844-7181 (Phone)
Mailing address (USPS)
Sandia National Laboratories
P.O. Box 5800, MS 1320
Albuquerque, NM 87185-1320
Sandia National Laboratories
1515 Eubank SE
Albuquerque, NM 87123