Jonathan Bergh conducted the following benchmarks against our facilities. Tests were run on the HEX 64core AMD 6376 machine to test the
scaling of a standardised CFD problem over an increasing no. of cores.
The test mesh was a turbomachinery blade passage and exhaust solving for
a low speed, incompressible, turbulent flow. A medium size mesh of
~2.2M cells was used and flow was solved using a pressure-based, coupled
solver where all flow variables are solved in a blocked-coupled fashion
and the additional turbulence variables were solved for in a segregated
fashion. The coupled approach requires x2.5-3.5 times the amount of RAM
than a traditional segregated approach. Total mesh RAM requirements
were ~8GB which is not large but emphasis was on total solution time as
the optimisation algorithm which which will be coupled to the CFD
project to need (conservatively) 150-200 solutions to be calculated.
The tests scaled to jobs over 16 / 32 / 48 and 64 cores on the
ICTS Dell (Dell C6145 rack server). Total solution time was then
compared to two existing nodes (32) core based on the AMD 6276 and 6376
CPU's (HP Proliant DL385p rack server). All machines use DDR3-1600MHz
RAM and Intel MPI. The test was not disk read-write intensive as the
simulations performed were steady-state and so only minimal amounts of
data were written out to disk during each iteratio.
