LARGE SCALE NETWORK PROTOCOL EMULATION ON COMMODITY CLOUD

Abstract

Network emulation allows us to evaluate network protocol implementations, typically in higher fidelity than simulations. Most emulators allow execution of code in the platform or environment similar to the target platform on which the protocol will be deployed. This advantage comes at a cost. Emulation often requires much larger IO or computational resources than simulations. As a result, it is common to see some research projects doing simulations with up to hundred thousand nodes while emulations typically scale up to a few hundred nodes. In this thesis, we present CloudNet, a network protocol emulation platform that leverages the commodity cloud computing service to scale emulations to thousands of nodes. CloudNet uses a lightweight virtualization technique called LXC containers to emulate a single node. The network protocol code and the protocol state for each node is maintained in its respective container. CloudNet then uses properties of the network topology to determine where to place these containers among many physical machines, researchers might rent on the cloud service. CloudNet's careful mapping of nodes to the containers makes network performance more predictable and suitable for emulation even on a shared commodity cloud, which was previously thought to be unsuitable for serious network emulation. Through extensive experiments, we establish that CloudNet is scalable to thousand node networks while providing accurate emulation results.