SciClone Cluster Project Computational Science Cluster
Home
Introduction
Sponsors
Research
Hardware
Software
User Info
Documentation

Name: ATLAS

Version: 3.4.1

Description:

ATLAS (Automatically Tuned Linear Algebra Software) is a high performance implementation of the complete BLAS library plus a few selected LAPACK routines, with interfaces for both C and Fortran. On SciClone, we have installed multiple copies of the ATLAS libraries, individually tuned for each type of node in the system, in both 32-bit (v8) and 64-bit (v9) versions. Threaded versions of the libraries are also available for multiprocessor node types (everything except c1 and c3 nodes).

Installation Path:

/usr/local/ATLAS/include
/usr/local/[v8plusa,v9a]/c3/lib
/usr/local/[v8plusb,v9b]/[c5,db1,n2,s3]/lib

Documentation:

Usage Notes:

Use of an ATLAS library on a node type other than the one for which it was built may result in poor performance (if it even runs at all). For example, a program linked to the c5 version of the library most likely will fail if it is executed on a c3 node. Similarly, a 32-bit (v8) library cannot be linked against a 64-bit (v9) application, and vice versa.

Sun's Performance Library contains an alternative BLAS implementation which has also been tuned for UltraSPARC architectures.

SciClone's ATLAS libraries were built using Sun's Forte Developer 7 compilers rather than the recommended GCC 3.0.4 compilers. Based on the results reported by the ATLAS installation procedure, the Sun compilers provided better performance overall.