On Jan 29, 6:05 pm, sonu <[email protected]> wrote: > But one MUST get same output (when i use "double precison" for > output ), irrespective of the number of nodes.
A reasonable expectation from your side. A few point to look at (I can only help you search for the answer) 1. compiler flags: have the biggest effect (one point where control is possible, meaning changing the compile time options with a bit of hit- and-trial. Are you doing any compiler optimisation? Drop all -Ox flags now and check. Such things should be last on the list. 2. The related issue is that of language. FORTRAN/C ? Search for IEEE 754 floating point calculations and get informed about it . May help. 3. Another point could be the compiler itself. Using GCC suite or intel? Not quite sure if compilers for x86 architectures will be as good as those for other systems (SUN/IBM). My ideas/information could be outdated. Surely, there is prevalence of x86 based machines running linux and at times one has to use what one gets. 4. A genuine programming error. A little bit about the programming model will help. Things like is the final value a sum/average of partial sums/answers from different nodes. At times one does make mistakes in calculations before distributing and after gathering. > I think mpi(message passing interface) has bugs ? can you tell me how > mpi works? Because in parallel run, libraries of mpi are used. Tried using the accompanying examples/test cases that come with MPI? The MPI book has examples too. Will be a good idea to check with some simpler things like calculation of pi (I think this example is there in the book). If they give sound results then there is nothing wrong with the mpi (your installation, that is). Internal workings of the library may not be of use (at least at this stage). Good luck Chetan -- l...@iitd - http://tinyurl.com/ycueutm
