USENIX Association Proceedings of the 4 th Annual Linux Showcase

This paper presents a multi-threaded BLAS library for dual SMP Intel computer running Linux. We present simple techniques to obtain parallelism for BLAS call transparently from the client program. We discuss some synchronization methods available under Linux, show performances results for a representative set of BLAS and for a high level linear algebra kernel. We then explain some key points on cache management and how they can impact on performances of the blasth library. Next we'll draw some conclusions on the use of SMP computer for linear algebra and present evolution perspectives for the library.