|Search Duke CS||
Compared to the evaluation methods like prototyping and simulation, the advantage of analytical stochastic models lies in the lower cost of constructing and solving the model, meanwhile covering a large parameter/workload/configuration space. Nevertheless, tractability of such models requires simplifying assumptions that are sometimes seen as significantly impacting the fidelity of the model. In order to overcome such difficulty, hierarchical stochastic models can be used to capture important details about the system workload, fault load and various other hardware/software aspects to gain fidelity while retaining tractability.
This project proposal describes an approach for the analytical modeling of datacenter networks. Utilizing hierarchical modeling, this approach allows compact specification of large network architecture and easy comparison of different system structures and/or parameter profiles. To illustrate our approach, we’ll apply our modeling techniques to a datacenter network architecture commonly used in practice.