Data Distribution Analysis of MCGLS Algorithm for Parallel Least Squares Problems.pdfVIP

Data Distribution Analysis of MCGLS Algorithm for Parallel Least Squares Problems Tianruo Yang Abstract In this pap er we mainly study dierent data distribution of MCGLS a mo died algorithm of CGLS which is a basic iterative metho d to organize the computation of conjugate gradient metho d applied to normal equations for solving sparse least squares problems on massively parallel distributed memory computers The p erformance of CGLS on this kind of architecture is always limited b ecause of the global communication required for the inner pro ducts MCGLS can achieve some improvements on parallel p erformance by creating situations where communication can b e overlapp ed with computation Here a theoretical data distribution analysis taking into account the time complexity and load balance is presented which allows us to decide b est data decomp osition for sparse and large least squares problems It is shown that the row blo ck decomp osition is much more ecient than other decomp osition mo dels Intro duction Many scientic applications such as tomography geo detic survey and structural analysis give rise to the least squares problems with a large and sparse co ecient matrix A Minimizing by solving the normal equations is a common and often ecient approach T b ecause A A is symmetric and p ositive denite and it can b e solved eciently by using conjugate gradient metho d The resulting metho d CGLS is used as the basic iterative metho d to solve the sparse and large least squares problems In theory it is a straightforward extension of the standard conjugate gradient metho d However numerical unstable variant