Next: Bibliography
Up: BLASTH, a BLAS library
Previous: stack alignment and thread/processor
The use of low end Intel SMP computers inside Beowulf or c.o.w.
can help in getting better performances when applications does not consume
a lot of memory bandwidth: we have always to remember that a cluster of single
processors nodes has twice aggregate memory bandwidth of an equivalent cluster
of dual SMP with the same number of processors. In linear Algebra a lots of
high level computation kernel use block methods and the blasth library could
help in these cases as we see for LU factorization. The perspectives for
this work are important because we need to work on more LAPACK routines to
provide a useful library. There is also work to do on matrix matrix product
and LU factorization to improve acceleration in small cases and our choice of
simple parallelization for LU cannot be ideal for more processors.
Porting to other platforms such as alpha systems is an ongoing work and theses
systems can change scalability results for high bandwidth consuming
BLAS such as ddot and dgemv making the blasth library more interesting.
Thomas Guignon
2000-08-24