Parallel sparse and banded matrix – multiple vectors multiplication

Download
2014
Cincioğlu, Meftun
In this thesis, performance of two important primitives, namely sparse and banded matrix – multiple vectors multiplication are studied. Sparse matrix – multiple vectors multiplication (SpMM) is one of the basic and most time consuming operations in many problems in science and engineering. Hence, any improvement in the performance of SpMM operations has a great impact on the wide spectrum of problems. One of the objectives of this thesis is to improve the performance of parallel SpMM operation by reducing indirect memory access, improving communication pattern, and load balancing. For this purpose, partitioning tools and permutation algorithms are used. Banded matrix – multiple vectors multiplication is used as a primitive operation in iterative solution of banded linear systems or in other applications. An improved method is presented that has an advantage especially for banded matrices having small bandwidth and multiplied by large number of vectors. All these numerical experiments are performed in two different computing platforms.
Citation Formats
M. Cincioğlu, “Parallel sparse and banded matrix – multiple vectors multiplication,” M.S. - Master of Science, Middle East Technical University, 2014.