Information Technology Reference
In-Depth Information
Fig. 3.
Strong scalability the Xeon Phi for the flow solution on the kiln furnace
simulation. Ideal and measured scalabilities, parallel eciency and mean number of
elements per core are shown.
-
A straight comparison with the speed with the Sandy Bridge hosts leaves
IXP way behind: the same example with
no MPI
runs around 20 times
faster in the Sandy Bridge, which is surprising considering the differences in
processors clock (slightly 2x favouring the host). That is to say that, running
the same sequential case, one MPI task takes around 20 times longer in the
Xeon Phi. Further tracing of the code to analyze its performance is required.
Also, it is very likely that a specific IXP compiling option set will make them
run much faster. In this paper we have left this optimization aside purposedly.
-
The powerful hosts were not used in the runs here, only the IXP. Now that
we are sure that the code scales well in the accelerators and once a proper
compiling option set is obtained for the IXP, we can establish the true speed
ratio between both computing units. Taking into account these figures, we
can shue the MPI tasks among the host and the accelerator, giving METIS
a weight to do a balanced partition.
-
Software threading is still to be analyzed, specially for the hybrid MPI-
OpenMP mode.
References
1. Metis, family of multilevel partitioning algorithms,
http://glaros.dtc.umn.edu/
2. Casoni, E., Jerusalem, A., Samaniego, C., Eguzkitza, B., Lafortune, P., Tjahjanto,
D., Saez, X., Houzeaux, G., Vazquez, M.: Alya: computational solid mechanics for
supercomputers. Archivesof Computational Methods in Engineering (in press, 2014)
3. Houzeaux, G., Vazquez, M., Aubry, R., Cela, J.: A massively parallel fractional
step solver for incompressible flows. J. Comput. Phys. 228(17), 6316-6332 (2009)
4. Houzeaux, G., Vazquez, M., Saez, X., Cela, J.: Hybrid mpi-openmp performance
in massively parallel computational fluid dynamics. Presented at PARCFD 2008
Internacional Conference on Parallel Computacional Fluid Dynamics, Lyon, France
(2008)
