CASE Cluster

Theory

The concept of using multiple independent networked computers to complete a computational task quickly has been applied to the Euler3d code of the STARS software package. This new feature allows STARS to use large arrays of inexpensive off the shelf computers to complete the aeroelastic predictions in a small faction of the time a cutting edge single processor would require.

The new Euler3d_dpp (dpp => Distributed Parallel Processing) automates the process of simulating the multiple free responses needed at varying dynamic pressure needed to complete an aeroelastic analysis. It also automates and accelarates the generation of training data for the system identification techniques pioneered here at the CASElab.

Euler3d_dpp takes advantage of the natural parallel independent processing needed to make an aeroelastic prediction of flutter. An aerostructure's response to change at all conditions in its flight envelope must examined. Those conditions are divided into discrete sets which are tested in parallel on independent processors. This is how Euler3d_dpp uses an array of computers to accelerate the completion of the analysis.

Hardware

CASECluster has eight (8) computing, or slave, nodes; one of which doubles as the control, or master, node. The wiring for the ethernet, power, video, keyboard and mouse is hidden behind the computers.

Node Specifics

Networking Specifics

The node were selected based on the best performance to price configuration available at the time the cluster was assemblied. Microsoft Windows 2000 was selected for two reasons 1) we are familiar with it and 2) Windows has shown a performance increase over Linux in our benchmarking studies. The MPICH software package is a free academic release that has performed excellently for us.

Results

As examples of the possible performance gains from cluster use, two test cases are presented here.

The first example is the AGARD445.6. This test case is a simulatio of a single wing in a wind tunnel test condition. The Mach number is 0.96; so the model is in the transonic region where many structures are most likely to flutter.

The Second exampel is the GHV, Generic Hypersonic Vehicle. This is a model a high speed aircraft in flight at Mach 2.2.

The following graphs show the relationships between the number of mode shapes and the time to complete a flutter prediction. The number of mode shapes is directly related to the complexity of the structural model. So these data can be applied to any test cases with the same Mach number and number of computational elements as the models presented here. The data in these graphs has been extrapolated from results for a few cases.

AGARD445.6

Notice that the system identification (SysId) values outperform the traditional free response method for the AGARD445.6.

GHV

Notice that the system identification (SysId) values perform poorly compared to the traditional free response method for the GHV.

Links