GOSPEL |
Grid Optimisation Software for Problems of Elastohydrodynamic Lubrication |
Running in a Problem Solving Environment inside NAG's IRIS Explorer package the optimiser investigates a multi-dimensional parameter space for the fluid's characteristics. There is a measured friction calculation for each of several different ambient temperatures, loadings and slide to roll ratios against which the parameters are minimised.
Shown below is the latest version of one such problem, optimising against ten different parameters. Here there are 36 different EHL calculations for each function call from the optimiser. In total that means that there have been 985x36=35460 separate cases run.
In the screenshot two output graphs are shown. The left hand one shows the relative change from the initial guess as the simplex finds increasingly better solutions - the nearly parallel lines towards the end show that the solution has converged. The right hand graph shows parallel coordinates with a separate line for each set of parametes.
The e-Science aspects of this project can be broken down into three main sections: (i) getting the optimiser running in parallel to speed up the real time of obtaining these computational results; (ii) using a parallel solver on each individual EHL case to further increase the speed of the solution process and allow much larger cases to be tackled than have previously been feasible; (iii) getting all this functionality on a Grid resource - possibly split over several different machines - still interactive from the PSE.
The parallelism of the optimiser has been achieved using MPI. Communication is restricted to syncronisation of the calculated residuals (the differences between the numerical and experimental friction). This means that all 36 cases could theoretically be run in parallel. However this is an inefficient way of doing this since the results from one case lead to a good guess for a similar case with a higher slide-to-roll ratio. We therefore have six continuation runs with a different temperature and loading pair in each. The speed-up of these cases is therefore limited by the time taken for the slowest run to converge. Timings are given in the All Hands paper, linked below.
The parallelisation of the 2-d circular contact solver has further developed the solver developed in a previous EPSRC project. This code has remarkable speed-ups for individual cases and hence the inclusion of the solver has enabled 2-d cases to be optimised for the first time. This has not been previously done due to the prohibitive computational costs involved. MPI is again used within the solver although each of the continuation runs now uses a separate MPI communication group. Another group for the head processes of each continuation run is used for synchronising the residuals, as before. This low bandwidth communication is what enables the use of MPICH-G2 for MPI communication between separate Grid resources.
The Grid aspects concerned with the communication between the PSE and the optimisation process has made extensive use of work in the gViz project, another UK e-Science project. By using and extending the gViz library with both the application and the PSE all the Globus authentication, Globus RSL script writing, communication with the remote Grid resource job schedulder, the job is launched from a single click on the PSE interface. Communication of the visualisation streams is done through gViz setting up extra threads with socket connections back to the PSE and, whenever new data is available, the datasets are sent to all connected participants. Steering data is handled in a similar manner.
The work was also featured as a demo on the UK e-Science stand and as a poster on the University of Utah SCI Institute stand at Supercomputing 2003 in Phoenix, Arizona. It is was also both a poster and a presentation at the National e-Science Centre, Edinburgh, coordination meeting in April 2004.
|
| Return to homepage | E-mail comments |