In this paper we describe an investigation of a part of testing folklore: that long tests are more likely to find bugs than short ones.

In the experiments we used formal traversal tools to generate tests that cover all transitions of a projection of an abstract FSM model of two processor units. We compared a suite of short tests, each of which covers a single new transition by a shortest path, to a suite of longer tests that combine several new transitions and do not necessarily use shortest paths. The results of the experiments show that long tests achieve better coverage on other projections of the state machine. Moreover, the test suites consisting of long tests, measured in total length in cycles or generation cost, are lower cost than the suites built of short tests.