Scalable Libraries

Jim Demmel
Berkeley

Demmel Introduction

* They are building Castle which seems to be Berkeley's private system?

- Unclear how relates to industry standards

* Unreasonable number of acceses to netlib -- why so large?

- >100 accesses per day for major packages such as CLAPACK (C version) and PVM
- Surely number of actual users of say PVM averages at about 1-5 per day (thousand users is one per day for 3 years)

* LAPACK is 800,000 lines of code

- C version is one million lines of code and got by automatic conversion
- no object oriented techniques

Demmel -- (SCA)LAPACK

* New Sparse LU is 16 times faster than Matlab version -- GP
* Duff's UMFPACK2 is similar performance on Sparse LU
* SCALAPACK uses 2D block cyclic as best general decomposition except for nonsymmetric matrices
* Have performance models for everything
* Divide and conquer methods for eigenvector/eigenvalue determination

- Ultimate methods! Problem is finished .....

* Holy Grail for eigenvector determination copes with close and degenerate eigenvalues

Who uses Numerical Software

* Engineers and Scientists
* HPCC community
* Education
* Each set of users has different tradeoffs
* Redoing Templates with greater attention to helping user choose correct method

- focused on eigenvalues/vectors
- PETsc is a nice language

* see http:/www.cs.berkeley.edu/~demmel

- nline courses in Parallel Computing and Numerical Linear Algebra

MultiPol-Kathy Yelick

* Examine set of irregular problems including

- Circuit Simulators
- Irregular CFD

* Abstract set of Irregular data structures

- Just released
- includes Fast Multipoles etc.

* which have been implemented on top of active messages