Section: Scientific Foundations

Performance evaluation

Participants : Thomas Begin, Paulo Gonçalves, Shubhabraya Roy.

Broadly speaking, performance evaluation aims at quantifying the behavior of a system. To do this, we frequently have to rely our analysis on a theoretical model rather than directly observing the behavior of the system. Several reasons may explain this choice: the system is not instrumented or not available for measurements, analytical results are often faster to obtain and brings more insight, the system may still be at prototype stage, etc.

Constructive modeling basically consists to mimic the internal operations of a system in a theoretical model. To hold the complexity of the model at a tractable level, it is common to represent complex internal mechanisms by random variables so that the resulting model is stochastic. The choice of distributions for random variables is often driven by the our expertise on the system. We attempt to devise models that can be represented as Markov chains or as queueing model as this will ease their subsequent resolution.

The resolution of the theoretical model provides numerical values for customary performance parameters such as the steady-state distribution of the number of requests (packets) in the system, the average throughput, the rate of losses, the mean utilization rate of a resources, etc. Regarding the nature of the model and its complexity, we set up entirely (sometimes approximate) analytical solutions, numerical solutions or discrete-event simulations to assess the values of the sought performance parameters.