NDSolve acceleration by forcing use of LAPACK functions
- To: mathgroup at smc.vnet.net
- Subject: [mg60481] NDSolve acceleration by forcing use of LAPACK functions
- From: "James Irwin" <irwinj at gmail.com>
- Date: Sat, 17 Sep 2005 02:32:15 -0400 (EDT)
- Organization: http://groups.google.com
- Sender: owner-wri-mathgroup at wolfram.com
Hi, I'm looking at two Mathematica notebooks. The first is Michael Trott's "Mathematics of Tsunamis" (http://mathworld.wolfram.com/news/2005-01-14/tsunamis/) http://mathworld.wolfram.com/news/2005-01-14/tsunamis/Tsunami.nb and the second is the Mathematica 5.2 promoting "Modeling, Simulating and Analyzing a Tsunami" (http://library.wolfram.com/infocenter/Demos/5699/) by Robert Knapp et al (noted animations on Jeff Bryant's pages). The Trott version has an NDSolve[] method options set up that invokes LAPACK's dgetrf() function. The Knapp version does not. Why this is important will now become more clear: The LAPACK function dgetrf() performas an LU matrix factorisation which is dominated by a matrix-matrix multiplication (i.e. Dot[] in Mathematica, also called 'dgemm' in BLAS/LAPACK speak). I work for ClearSpeed in the UK and we have PCI-X co-processor cards that deliver 50GFLOPS of dgemm performance. We wish to show the capability improvement users will see running 'hard' computations. Each of the notebooks is easily scalable. In fact, the Knapp example was designed to demonstrate the impact of this scaling. By increasing the resolution of the computation, the wave propogation is smoother and more faithful. Until now, the cost of this more refined analysis has been presented in minutes turning into hours of compute. Using our card, I have measured a six-fold time-difference for the LAPACK-function (given a large enough problem size) versus a 3.4GHz Xeon. Obviously, it would be more interesting to use the Knapp version since it shows the wave propogation on a per-step basis and has a more interesting sea-bottom. I have thus far failed to comprehend the changes required to make the method options in the Knapp version invoke the LAPACK functions. I would appreciate some helpful pointers. Thanks in advance, James Irwin ClearSpeed Technology PLC