Scalingin- Numerical- Linear- Algebra课件.pptVIP

ScaLAPACK Team Susan Blackford, UT Jaeyoung Choi, Soongsil U Andy Cleary, LLNL Ed DAzevedo, ORNL Jim Demmel, UCB Inder Dhillon, UCB /scalapack Jack Dongarra, UT/ORNL Sven Hammarling, NAG Greg Henry, Intel Osni Marques, NERSC Antoine Petitet, UT Ken Stanley, UCB David Walker, Cardiff U Clint Whaley, UT scalapack@ Possible Data Layouts ScaLAPACK supports all layouts 2D block cyclic recommended, for load balance and BLAS3 1D blocked 1D cyclic 1D block cyclic 2D block cyclic ScaLAPACK Structure ScaLAPACK BLAS LAPACK BLACS MPI/PVM/… PBLAS Global Local Parallelism in ScaLAPACK Level 3 BLAS block operations All the reduction routines Pipelining QR Iteration, Triangular Solvers, classic factorizations Redundant computations Condition estimators Static work assignment Bisection Task parallelism Sign function eigenvalue computations Divide and Conquer Tridiagonal and band solvers, symmetric eigenvalue problem and Sign function Cyclic reduction Reduced system in the band solver ScaLAPACK Performance Models (1)ScaLAPACK Operation Counts ScaLAPACK Performance Models (2)Compare Predictions and Measurements (LU) (Cholesky) Making the nonsymmetric eigenproblem scalable Axi = li xi , Schur form A = QTQT Parallel HQR Henry, Watkins, Dongarra, Van de Geijn Now in ScaLAPACK Not as scalable as LU: N times as many messages Block-Hankel data layout better in theory, but not in ScaLAPACK Sign Function Beavers, Denman, Lin, Zmijewski, Bai, Demmel, Gu, Godunov, Bulgakov, Malyshev Ai+1 = (Ai + Ai-1)/2 ? shifted projector onto Re l 0 Repeat on transformed A to divide-and-conquer spectrum Only uses inversion, so scalable Inverse free version exists (uses QRD) Very high flop count compared to HQR, less stable The “Holy Grail” (Parlett, Dhillon, Marques) Perfect Output complexity (O(n * #vectors)), Embarrassingly parallel, Accurate To be propagated throughout LAPACK and ScaLAPACK Making the symmetric eigenproblem and SVD scalable ScaLAPACK Summa