Graph partitioning algorithm for model transformation frameworks

Software modeling has become an everyday practice. Modeling extra-large models have enormous constraints: both memory and computational capacity of a single computer might be insufficient for handling model transformations. One solution to overcome this barrier is to extend the infrastructure. Cloud computing provides feasible realizations for these needs. However, existing algorithms have to be extended/modified to support cloud computing and use its advantages most efficiently. Generally, models can be easily mapped to graphs. This paper provides an algorithm for partitioning graphs representing models. Models can be mapped onto several computational instances and processed on these instances in a distributed fashion. Our algorithm is based on the heuristic Kernighan-Lin method, but we allow manually altering the number of partitions dynamically based on the actual needs. Moreover, we do not build on knowing the entire model when creating the partitions, since it would not fit into the memory of a single instance. Instead, model nodes are received and processed one by one. Our algorithm is fine tuned to these special conditions. The efficiency of the algorithm is illustrated by a case study.