Out-of-core computation of the QR factorization on multi-core processors

Abstract

We target the development of high-performance algorithms for dense matrix operations where data resides on disk and has to be explicitly moved in and out of the main memory. We provide strong evidence that, even for a complex operation like the QR factorization, the use of a run-time system creates a separation of concerns between the matrix computations and I/O operations with the result that no significant changes need to be introduced to existing in-core algorithms. The library developer can thus focus on the design of algorithms-by-blocks, addressing disk memory as just another level of the memory hierarchy. Experimental results for the out-of-core computation of the QR factorization on a multi-core processor reveal the potential of this approach. © 2009 Springer.