next up previous
Next: About this document ...

H. Carter Edwards
Performance Concerns when Iterating Hybrid-Parallel Kernels

Sandia National Laboratories
P O Box 5800 / MS 0382
Albuquerque
NM 87185
hcedwar@sandia.gov

Iterative algorithms are typically implemented by a sequence of calls to simple computational kernels, such as the BLAS or their sparse equivalent. Hybrid-parallelization of these kernels on clusters of nodes with multicore CPUs or GPGPUs has demonstrated performance gains for individual kernels. An iterative algorithm can realize a similar performance gain only if the programming model for calling a sequence of these kernels does not introduce significant overhead. Such a programming model for hybrid-parallel kernels has been implemented in Trilinos' ThreadPool library. A simple CG iterative solver is implemented using the ThreadPool library and its hybrid-parallel performance is assessed.





root 2010-03-03