skip to: onlinetools | mainnavigation | content | footer
Siva Rajamanickam
Research
Publications
Software
Activities
Home

Research

My primary focus is on High Performance Computing. I focus on many inter-related subfields within high performance computing as described below. One common theme among all my work lately is to design architecture-aware algorithms for next-generation supercomputers.

Linear Solvers

Linear solvers are the foundational tools for many scientific simulations of national interest. While this subfield has been studied for many years, number of open problems still remain. My primary interest in linear solvers area is to develop algorithms for linear solvers that is targeted towards future supercomputers. I am interested both in the node-level and system-level solvers. At the node-level my interests lie in task-parallel/data-parallel factorization like preconditioners, smoothers and direct solvers. At the system-level, I focus on hybrid Schur complement methods and preconditioners for communication-avoiding (s-step) Krylov methods.

Related Publications


Combinatorial Algorithms for High Performance Scientific Computing

Combinatorial algorithms impact a number of areas in high performance scientific computing starting from partitioning the input for better load balance and increased parallelism, ordering techniques for fewer floating point operations and increased parallelism, matching techniques for better numerical stability of linear solvers etc. I have worked on all these aspects (partitioning, ordering, coloring, matching algorithms). My primary interest in algorithms for next-generation architectures in some cases (coloring, matching) and new algorithms for system-level efforts in other cases (partitioning, ordering). I am interested in graph, hypergraph and coordinate based algorithms for different type of problems.

Related Publications


Graph algorithms for analytics in HPC Platforms

As the amount of data from the web, social networks and other non-traditional data sources grows, the need to analyze these data using high performance computing systems grows as well. Data from these resources have very different properties when compared to data from traditional scientific computing problems metioned above. I am interested in specialized techniques for analyzing such data using HPC platforms. We have developed a number of algorithms from graph traversals, identifying connected components, community detection based partitioning algorithms.

Related Publications


Linear Algebra Kernels

Sparse and dense linear algebra kernels are foundational kernels for scientific computing and in some cases even data analysis. My primary focus is on developing performance-portable algorithms for sparse and dense linear algebra kernels on architectures like GPUs and Intel Knights Landing processors.

Related Publications


Applications

All the above algorithmic work is focused on delivering capabilities to important applications from a number of domains such as circuit, thermal fluids, solid mechanics, structural dynamics and climate simulations. My work impacts these applications both directly and indirectly.

Related Publications



Image placeholder Contact
E-mail: srajama@sandia.gov
(505) 844-7181 (Phone)

Mailing address (USPS)
Sandia National Laboratories
P.O. Box 5800, MS 1320
Albuquerque, NM 87185-1320

FedEX/UPS/DHL
Sandia National Laboratories
1515 Eubank SE
MS 1320
Albuquerque, NM 87123