In Fall 2001 entered the production state a computing
system dedicated to the Direct Numerical Simulation (DNS) of wall turbulence,
in particolar turbulent plane channel flows and pipe flows. The system has
been designed on top of a related computer code, written in the programming language
CPL, developed by Prof. P.Luchini
The system (codename: dilbert) is a sigle computing
machine, designed for running in a very efficient way just our code, and
is (as to now) formed by 8 computing nodes. Each of them is a commodity Dual
Pentium III Personal Computer, with 2 733MHz CPUs (namely one dilbert and
seven dogberts :-)
The nodes are connected each other with two commodity
Fast Ethernet cards, without hubs or switch altoghether. On every node runs
the operating system Debian/GNU Linux
. Parallel computing does not require message-passing libraries like
MPI.
The nodes of the cluster are connected with a simple
ring topology, which reflects how data are exchanged between nodes for parallel
execution of simulations. In the diagram below the correspondence is shown
between the cluster nodes and the slices into which the channel is partitioned,
in the wall-normal direction. Color represent the interface between
slices, with the corresponding ethernet cable connecting the two neighbouring
nodes.
The computer code solves the incompressible Navier-Stokes
equations both in cartesian and cylindrical coordinates, with the highest
efficiency. Together with the dilbert system, the DNS of turbulent wall flows
at relatively high Reynolds numbers can be afforded easily, without the nedd
of resorting to external, cost-uneffective supercomputers.
A synthetic description of the comuting system, the
numerical method and a comparison of their performaces with those os some
amongs the bleeding-edge researc DNS codes in the world is reported in this paper.
|