All these applications motivate the construction of a large polyhedral benchmark which enables us to analyze implementations experimentally. Our benchmark, named POLYBENCH, provides an easy and straightforward way to compare methods, with a great quantity of operations encountered in program analysis and transformation.
Early experiments were carried out, focusing on the satisfiability test with only three implementations: JANUS [SOG,SOG96], Simplex and Fourier-Motzkin [PIPa,CAI00]. They revealed important differences among these algorithms. Then, we extended our framework and added the projection, minimization, dual conversion and convex hull operators. The following section describes this benchmarking environment.