Parallelisation issues

The most straightforward way of parallellazing this kind of application is the exploit the parallelism the is present in the idea of multiple populations. Another approach would have been to parallelize some operations within the evolutionary loop but this would make it very hard to insure compatibility with the serial version of lil-gp.

One could very well have thought about a parallel evaluation function for instance. But since this is problem dependent, the user would be forced to invest him or herself in parallel programming techniques that are not immediately relevant for their research.

To change the kernel of lil-gp into a parallel kernel the following problems had to be solved:

  1. Distribution of the different subpopulation onto different CPU's
  2. Modification of the exchanges of individuals in order to accommodate for inter CPU exchanges
  3. Adapt the random number generator slightly
  4. Synchronous versus asynchronous exchanges

The information provided on this page is copyrighted © by the Free University of Brussels and provided as is.
From more information contact Johan Parent