A Graybox Approach to Rehoming Service Chains

We now present one of our key contributions—analysis and modeling of the rehoming cost for different rehoming actions. We specifically consider the rebuild and cold migrate rehoming actions, and model two different costs:(i) time to perform the rehoming action, as logged by OpenStack, referred to as action time, and (ii) time until the service chain’s end-to-end connectivity is restored, referred to as connectivity downtime; we measure this by calculating the delay between successful pings from client to server in the service chain. In general, the rehoming cost can be expressed as a utility function in terms of various individual costs; we focus on these two costs in this paper as illustrative examples.We consider a simple iPerf [8] service chain deployed in CloudLab [9], consisting of an iPerf client VM, a router VM, a firewall VM, and an iPerf server VM, to collect empirical data for modeling (see Section VA for details of our experimental setup). We focus on modeling the rehoming costs for the firewall VM; we expect similar results for other VMs as well, though the post-bootup process may be different, depending on the VM’s functionality. It is important to note that the results in this section are for a single generic VM being rehomed; we show in Section VD how these results can be applied to predict the optimal simultaneous rehoming actions for multiple VMs in the service chain.