Efficient Online Virtual Machines Migration for Alert-Based Disaster Resilience

Several recent weather-based disasters had very negative impacts on cloud networks, causing Data Center (DC) shutdown, consequent data-loss and intolerable downtime of cloud services. This has put the reactive disaster-resilient design of cloud networks on top the agenda of several cloud DC operators. DC operators are investigating approaches to avoid downtime of cloud services in case a DC is affected by a disaster. Thanks to virtualization most cloud services run on Virtual Machines (VMs) hosted by DCs, so it is possible to keep these services alive if the VMs are evacuated (namely, migrated) before the disaster from a DC affected by the disaster to a DC in a safe location, in an online technique. This technique is known as online “VM migration”, which results without or with a minimal service downtime. In this paper, we present an Integer Linear Programming (ILP) model for efficient online VMs migration in case of an alerted disaster (e.g., most weather-based disasters, as hurricanes) such as to avoid service downtime. The ILP performs scheduling and assigns route and bandwidth to the migration of VMs towards a safe DC within an alert time, with the objective of maximizing the number of VMs migrated and minimizing service downtime, network resource occupation and migration duration. We present a comparative analysis of offline and online migration strategies such as to quantify the trade-off between downtime, network resource utilization and migration duration. Moreover, we investigate the impact of the memory dirtying rate on the online migration process, i.e., the number of VMs evacuated and network resource occupation.