Statistical software testing is an increasingly popular method in the software development cycle. An exact modeling of the usage profiles of a software system is an indispensable prerequisite for statistical testing. Recently, new techniques for obtaining optimal usage profiles even the presence of rarely used critical functions have been introduced. Although these techniques deliver unbiased dependability estimates with a single model (instead of using multiple models as it is the current practice) their applicability is hampered by their prohibitive computational complexity. In this paper we demonstrate that these techniques can be effectively employed even for large software systems by exploiting high performance computing. We discuss the parallelism potential within these techniques and present a parallel implementation based on a high level parallel language--High Performance Fortran. Our techniques are general enough to allow an efficient optimization of medium to large scale test plans on high-end parallel machines as well as on more cost effective PC-clusters. A real world software system is studied in detail to prove our claims.

Top