Sparsity-Aware Storage Format Selection

The simulations in many scientific applications re¬quire the solution to linear systems of equations. Libraries are often used for high-performance executions of direct solvers of linear systems. However, the performance of these libraries is often tied to the underlying hardware platform and the code is not portable. We propose to develop a framework called Axb, which will use information about the architecture and the application. Axb automatically selects a good storage format and implementations that generate efficient code for the sparse direct solver. The results demonstrate that the performance of direct solvers can be accelerated upto 11X compared to using a library approach for the direct solver on multicore architectures. The work is an extension to our previous publication [1].