Michelle Mills Strout
mstrout@cs.ucsd.edu
http://www.cs.ucsd.edu/~strout
Tiling for Iterative Sparse Matrix Computations
In modern computer architecture the use of memory hierarchies
allows a program's data locality to directly affect performance.
Data locality occurs when data is still in a cache upon reuse.
We describe a technique for tiling iterative sparse matrix
computations in order to improve the data locality in scientific
applications, such as Finite Element Analysis, that spend most of
their computation time solving sets of simultaneous linear equations,
Au=f. Specifically, we present sparse tiling which tiles
the iterative solver Gauss-Seidel at runtime.