1. Introduction

BLAS¹[8,4,3] are a set of subroutines written in Fortran 77 which provide a standard API for simple linear algebra operations. BLAS are now widely used and base of library such as LAPACK, PETSC, SCALAPACK... A recent development in parallel computing was to use on the shelves components to build high performance computers, these components could include low and dual Intel processor systems. This choice leads to the use of two parallel programming models: message passing between nodes end shared memory inside nodes. The goal of the blasth library is to provide a transparent support of shared memory for BLAS call: the program call BLAS as usual but the call is processed on 2 processors. BLAS are split in 3 level:

• level 1: vector operations,
• level 2: matrix-vector operations,
• level 3: matrix-matrix operations and triangular solve.

In this paper we first present some discussion and experiments for synchronizing threads with Linux on SMP computer, we will then focus on daxpy and ddot level 1 BLAS, dgemm (from level 3). The sequential BLAS library used will be the one from the f77 implementation ASCI Red project² and ATLAS³[9] (for level 3). We also present results for a block LU factorization.