|Speaker :||Michele Tortelli|
|Time:||2:00 pm - 4:00 pm|
|Location:||LINCS Meeting Room 40|
Large scale deployments of general cache networks, such as Content Delivery Networks or Information Centric Networking architectures, arise new challenges regarding their performance prediction and network planning. Analytical models and MonteCarlo approaches are already available to the scientific community. However, complex interactions between replacement, replication, and routing on arbitrary topologies make these approaches hardly configurable. Additionally, huge content catalogs and large networks sizes add non trivial scalability problems, making their solution computationally demanding. We propose a new technique for the performance evaluation of large scale caching systems that intelligently integrates elements of stochastic analysis within a MonteCarlo approach. Our method leverages the intuition that the behavior of realistic networks of caches, being them LRU or even more complex caches, can be well represented by means of much simpler Time-To-Live (TTL)- based caches. This TTL can be either set with the guidance of a simple yet accurate stochastic model (e.g., the characteristic time of the Che approximation), or can be provided as very rough guesses, that are iteratively corrected by a feedback loop to ensure convergence. Through a thorough validation campaign, we show that the synergy between modeling and MonteCarlo approaches has noticeable potentials both in accurately predicting steady state performance metrics within 2% accuracy, while significantly scaling down simulation time and memory requirements of large scale scenarios by to two orders of magnitude. Furthermore, we demonstrate the flexibility and efficiency of our hybrid approach in simplifying fine-grained analyses of dynamic scenarios.