To tune granularity of tasks at runtime, we introduce an original algorithmic scheme. It is based on the coupling of two algorithms, one sequential, the other one parallel and fine grain. However, parallelism is generated only in case of idleness of some processor. Then, when executing the program on a limited number of resources, the overhead related to parallelism management is bounded with no restriction of potential parallelism. It is especially suited to applications for which parallelization drastically increases the number of operations or induces a loss of performance, despite a decreasing execution time. This scheme is applied to parallelisation of two applications: gzip (Gailly, 2003) that implements Lempel-Ziv compression, a P-complete problem; and PL (ProBayes, n.d.) a probabilistic inference engine.