Method for predicting the cost of live migration of virtual machines in a customer environment
Publication Date: 2010-Oct-19
The IP.com Prior Art Database
Please enter 2-3 line Abstract
Page 01 of 4
In dynamic virtualized environments, it is often necessary to migrate virtual machines between physical hosts (for load balancing, maintenance, etc.). When multiple alternative new placements are possible, specially designed placement algorithms determine the most optimal placement. Optimization typically considers the 'benefit' of the candidate placement, and also the 'cost of change', that is the penalty of moving running virtual machines from their current location to another location. Live migration is commonly used to minimize the impact of such a move on the applications running on the virtual machines. The cost of live migration may comprise several factors, such as the time needed to complete the migration, and the measurable impact on performance and/or availability of applications during the migration.
The problem addressed by this invention is how to predict (i.e., estimate in advance) the cost of migrating a given virtual machine in an operational customer environment..
One approach to solving this problem is to build an analytical model, based on characteristics of the application (e.g., throughput and latency of requests), the virtual machine (e.g., memory size, CPU utilization), the physical host (e.g. various migration capabilities), and the environment (e.g., network bandwidth). Such a model can be built and validated in a laboratory environment, and then applied in the customer environment. However, due to the variety of characteristics affecting the migration process, the behavior of live migration in the customer environment is likely to be different than observed in the laboratory (e.g., due to dynamic changes in the environment, or characteristics not included in the model). Thus, the predictions given by the model would be inaccurate, and might lead to sub-optimal placement decisions.
An enhancement to the analytical model is to apply an adaptive learning
algorithm, and to tune the model based on actual migrations performed in the customer environment.
Although this approach would make the model more
accurate, and is likely to provide better predictions, it may require a lengthy period of time until the model properly adjusts itself to the specifics of the customer environment.
The core of this invention is a framework that enables efficient adaptive learning of live migration costs in the customer environment.
The framework includes a 'virtual appliance' that can be deployed and configured in the customer environment as a virtual machine, migrated between physical hosts, and used to gain information on live migration performance in the customer environment. The virtual appliance will have an application stack identical to (or comparable with) the customer's virtual machines. To facilitate active learning, the appliance provides a management interface, to control configuration and performance characteristics of the virtual machine and application, as well as to monitor various metrics of the virtual machine...