UL HPC users'session: Mastering big data

We illustrate in this presentation an optimized HPC implementation for outranking digraphs of huge orders, up to several millions of decision alternatives. The proposed outranking digraph model is based on a quantiles equivalence class decomposition of the underlying multicriteria performance tableau. When locally ranking each of these ordered components, we may readily obtain an overall linear ranking of big sets of decision alternatives. The proposed optimization strategies tackles algorithmic refinements of the ranking algorithm, reducing the size of python data objects, typing the data for efficient cython and C compilation, efficient sharing of static data via global python variables, using a multiprocessing task queue, and, last but not least, use the efficient UL HPC equipements.