Benchmarks by NCSA Resource Modi Mercury Cobalt Copper Platinum Titan Tungsten Benchmarks by BLAS Routines DAXPY DGEMV DGEMM

About BLAS

Many applications utilize Basic Linear Algebra Subprograms (BLAS) which are typically provided by the hardware or compiler vendor such as SGI's SCSL (SGI/Cray Scientific Library), IBM's ESSL (Engineering and Scientific Subroutine Library), Intel's MKL (Math Kernel Library) or NAG (Numerical Algorithms Group), implying a reasonable level of performance can be expected. For portability, many users often bundle the open source version of BLAS from Netlib, thus relying on the compiler to produce efficient routines. Other open source projects like ATLAS (Automatically Tuned Linear Algebra Software) provide BLAS as either a pre-compiled library or source code with a method of compiling/optimizing for a given architecture. Also in the 'without-charge' category are high performance BLAS libraries hand-coded by Kazushige Goto for a variety of architectures.

Using the BLASBench component of LLCBench written by Phil Mucci, we present the evaluation of three of the more commonly used BLAS routines: DAXPY, DGEMV and DGEMM on NCSA HPC resources. The most recent version of the libraries is used when possible. Compilers are the default compilers for that platform. The configuration file for each machine, which shows how to link to the libraries, is provided.