Network Configuration and Management via Two-Phase Online Optimization

Automated configuration and management of highly dynamic networks is a challenging problem for network practitioners. Such online optimization of systems can be performed in two ways: (i) using a separate model of the system for experimenting new configurations, (ii) using the system itself for experimentation without a separate system model. The former approach fails for dynamic networks with high failure rates or variable demand profile. In this paper, we take the latter approach and perform in- situ trials in a network to find better configurations of IGP link weights giving result to higher network throughput. Our approach follows a two-phase model where the online optimization process periodically goes into a "search" phase followed by network operation with the parameters found in the latest search phase. We use a black-box optimization algorithm, called Probabilistic Trans- Algorithmic Search (PTAS), to search for better IGP link weights and evaluate our approach in terms of key parameters such as search phase frequency and length.