A systematic approach for performance assessment using process mining

Abstract

Software performance engineering is a mature field that offers methods to assess system performance. Process mining is a promising research field applied to gain insight on system processes. The interplay of these two fields opens promising applications in the industry. In this work, we report our experience applying a methodology, based on process mining techniques, for the performance assessment of a commercial data-intensive software application. The methodology has successfully assessed the scalability of future versions of this system. Moreover, it has identified bottlenecks components and replication needs for fulfilling business rules. The system, an integrated port operations management system, has been developed by Prodevelop, a medium-sized software enterprise with high expertise in geospatial technologies. The performance assessment has been carried out by a team composed by practitioners and researchers. Finally, the paper offers a deep discussion on the lessons learned during the experience, that will be useful for practitioners to adopt the methodology and for researcher to find new routes.

Appendix: Generalized Stochastic Petri Nets

In this appendix, we introduce the modeling formalism of Generalized Stochastic Petri Nets (GSPN).

A Generalized Stochastic Petri net (GSPN) (Ajmone-Marsan et al. 1994) is a bipartite graph, formally defined as a 8-tuple $\mathcal {N} = (P,T,I,O,H,{\Phi },W,M_{0})$ where:

• P is the set of places,

• T = T_i ∪ T_t is the set of transitions, divided into immediate (T_i) and timed (T_t) transitions,

• I, O, H : P × T → are, respectively, the input, output, and inhibitor arc multiplicity functions,

• Φ : T → assigns a priority to each transitions: timed transitions have zero priority, while immediate transitions have priority greater than zero,

• W : T → assigns to each immediate transition a weight, and to each timed transition a firing time delay. The firing time delay is the mean value of the negative exponential distribution,

• M₀ : P → assigns the initial number of tokens to each place.

Figures 6 and 7 show the graphical representation of two GSPN models, where places are depicted as circles, immediate transition as thin black bars and timed ones as thick white bars.

Transitions of a GSPN model represent actions or events in the modeled system, whereas places represent pre- and post-conditions for the actions/events occurrence. The dynamic of a GSPN model is governed by the concession, enabling and firing rules of transitions in a marking M : P →, which is reachable from the initial marking M₀ due to the firing of a sequence of transitions.

A transition has concession in a marking M when its input places contain at least as many tokens as the corresponding arc multiplicities, and its inhibitor places contain less tokens than the corresponding arc multiplicities.

A transition is enabled in a marking M iff it has concession in M, and its priority is greater or equal to the one of the transitions t^′ having also concession in M.

Consequently, only transitions of the same priority level can be enabled in a marking. A transition t, enabled in marking M, may fire then leading to a new marking M^′, according to the equation:

$$M^{\prime}(p) = M(p) + O(p,t) - I(p,t), p\in P. $$