Parallel version of SIP ?
hello all !
Have any of you come across a parallel version of Stone's Strongly Implicit Procedure. I have a vectorized version of the SIP, but I would like to make a transition from the Cray architecture to the Origin2000 architecture.
I would appreciate getting any information on a parallel version of the SIP (or a third party optimized version of the SIP for parallel architecture), and would greatly help if i could get the source code.
Re: Parallel version of SIP ?
I cannot point you directly at parallel code but would expect it to exist. However, I can make a suggestion based on parallelising implicit methods on the Origin 2000. Get the compiler to tell you where the data dependencies are and if they are not particularly important (does not matter too much if it is an old value) tell it lies. At the cost of an extra iteration or two you can sometimes get a very efficiently parallelised code. This advice is, of course, completely useless if you are not solving within an iterative loop.
|All times are GMT -4. The time now is 17:57.|