Data Centric Computing for Nuclear Physics

The infrastructure developed by Google for its MapReduce algorithm has obviously had an impact upon the world of search. There has been an emergence of algorithmic implementations of the MapReduce framework to support computing in the bio-sciences, machine learning, and general high performance computing. The basic attractiveness of MapReduce and other data-oriented topologies is that they provide a framework for data-centric computation: parallel computations are mapped across fixed data stores thus reaping the benefits of locality for situations where the cost of data transfer to a compute node would far outweigh the cost of transferring computation to the data store. The ultimate benefit of this paradigm therefore is that researchers could define virtual machines that contain standard analysis images, migrate those to the data, and then run MapReduce on datasets for which this analysis makes sense. The ultimate goal of this project (phase I and II) is to develop libraries and adaptation modules that will allow virtualized Nuclear Physics software to run efficiently under MapReduce