Speaker
Yusuke Osaki
(Hiroshima University, depertment of physical science)
Description
The modern graphic processing unit (GPU) has become one of the powerful computational
device. Applying GPU to the lattice QCD computation has been investigated especially for the
inversion solver of the discretized Dirac operators. Most of the previous works on lattice QCD
with GPU, however, have been done with single GPU card. Since the single GPU performance
and the memory size are limited at a few TFlops and of a few GBytes, the lattice sizes are
limited to rather small sizes. Thus, to simulate QCD on more realistic lattice we need a cluster
system with multiple GPU’s. The GPU cluster system requires a special care on the GPU-
GPU-intercommunication because the modern GPU lacks the direct GPU-GPU communication
feature. In order to make a good performance with a cluster with GPU, we investigated a domain
decomposition preconditioner for the O(a)-improved Wilson-Dirac fermion inverter, where the
lattice is decomposed into domains. In this talk we report some performance results on the
additive Schwarz preconditioner on GPU cluster computing.
| Please, insert your presentation type (talk, poster) | talk |
|---|
Primary author
Yusuke Osaki
(Hiroshima University, depertment of physical science)
Co-author
Dr
Ken-Ichi Ishikawa
(Hiroshima University, depertment of physical science)
