Existing  partitioning algorithms  provide limited  support  for  load
balancing  simulations that  are performed on  heterogeneous  parallel
computing platforms. On  such architectures, effective  load balancing
can only be achieved if the  graph is  distributed so that it properly
takes  into  account  the  available  resources  (CPU  speed,  network
bandwidth). We  developed  a  graph partitioning algorithm   that  can
address the partitioning requirements  of the scientific computations,
and can correctly model the  architectural characteristics of emerging
hardware platforms.