Performance Evaluation at the Software Architecture Level

When tackling the construction of a software system, at the software architecture design level there are two main issues related to the system performance. First, the designer may need to choose among several alternative software architectures for the system, with the choice being driven especially by performance considerations. Second, for a specific software architecture of the system, the designer may want to understand whether its performance can be improved and, if so, it would be desirable for the designer to have some diagnostic information that guide the modification of the software architecture itself. In this paper we show how these two issues can be addressed in practice by employing a methodology relying on the combined use of AEmilia --- an architectural description language based on stochastic process algebra --- and queueing networks --- structured performance models equipped with fast solution algorithms --- which allows for a quick prediction, improvement, and comparison of the performance of different software architectures for a given system. The methodology is illustrated through a case study in which a sequential architecture, a pipeline architecture, and a concurrent architecture for a compiler system are compared on the basis of typical average performance indices.