In this paper, we focus on the resource-constrained modulo scheduling problem, a general periodic scheduling problem, abstracted from the problem solved by compilers when optimizing inner loops at instruction level for VLIW parallel processors. Heuristic solving scheme have been proposed since many years to solve this problem, among which the decomposed software pipeling method. In this method, a cyclic scheduling problem ignoring resource constraints is first considered and a so-called legal retiming of the operations is issued. Second, a standard acyclic problem, taking this retiming as input, is solved through list scheduling techniques. In this paper, we propose an hybrid approach, which uses the decomposed software pipeling method to obtain a good retiming. Then the obtained retiming is used to build an Integer Linear Programming formulation of reduced size, which allows to solve it exactly. Experimental results show that a lot more problems are solved with this new approach. The gap to the optimal solution is really small (0 or 1%) on all the tested problem instances.

To cite this version:Abir Benabid, Claire Hanen. Worst case analysis of decomposed software pipelining for cyclic unitary RCPSP with precedence delays.

This paper studies the generalization of the first Garey-Johnson algorithm, which minimizes the maximum lateness on two parallel processors to the case of unitary typed tasks systems with constant delays. The performance of the extended algorithm is evaluated through worst-case analysis. If all the tasks have the same type and no delay is considered, then the upper bound obtained coincides with the upper bound for the Garey-Johnson algorithm on identical processors, which is one of the best known for the maximum lateness problem.

No abstract