Dynamic migration of virtual machines on a cluster of physical machines is designed to maximize resource utilization by balancing loads across the cluster. When the utilization of a physical machine is beyond a fixed threshold, the machine is deemed overloaded. A virtual machine is then selected within the overloaded physical machine for migration to a lightly loaded physical machine. Key to such threshold-based VM migration is to determine when to move which VM to what physical machine, since wrong or inadequate decisions can cause unnecessary migrations that would adversely affect the overall performance. We present in this paper a learning framework that autonomously finds and adjusts thresholds at runtime for different computing requirements. Central to our approach is the previous history of migrations and their effects before and after each migration in terms of standard deviation of utilization. We set up an experimental environment that consists of extensive real world benchmarking problems and a cluster of 16 physical machines each of which has on average eight virtual machines. We demonstrate through experimental results that our approach autonomously finds thresholds close to the optimal ones for different computing scenarios and that such varying thresholds yield an optimal number of VM migrations for maximizing resource utilization.

SUMMARYDispatching a large number of dynamically changing requests directly to a small number of servers exposes the disparity between the requests and the machines. In this paper, we present a novel approach that dispatches requests to servers through virtual machines, called Dispatching Requests Indirectly through Virtual Environment (DRIVE). Client requests are first dispatched to virtual machines that are subsequently dispatched to actual physical machines. This buffering of requests helps to reduce the complexity involved in dispatching a large number of requests to a small number of machines. To demonstrate the effectiveness of the DRIVE framework, we set up an experimental environment consisting of a PC cluster and four benchmark suites. With the experimental results, we demonstrate that the use of virtual machines indeed abstracts away the client requests and hence helps to improve the overall performance of a dynamically changing computing environment.

No abstract

Dispatching a large number of dynamically changing requests directly to a small number of servers exposes the disparity between the requests and the machines. In this paper, we present a novel approach that dispatches requests to servers through virtual machines, called Dispatching Requests Indirectly through Virtual Environment (DRIVE). Client requests are first dispatched to virtual machines that are subsequently dispatched to actual physical machines. This buffering of requests helps to reduce the complexity involved in dispatching a large number of requests to a small number of machines. To demonstrate the effectiveness of the DRIVE framework, we set up an experimental environment consisting of a PC cluster and four benchmark suites. With the experimental results, we demonstrate that the use of virtual machines indeed abstracts away the client requests and hence helps to improve the overall performance of a dynamically changing computing environment. DISPATCHING REQUESTS INDIRECTLY THROUGH VIRTUAL ENVIRONMENT399 relatively small number of backend servers. By placing a dispatcher at the entry to the infrastructure, the incoming requests can be quickly distributed to the computing resources in such a way that each server handles requests reasonably equally at any point in time. However, this 'reasonably equal' distribution is rather difficult to achieve since it involves various parameters such as job arrival rate, job distribution time, server status, processing time, etc.[1].Popular web sites dynamically assign and distribute a large number of client requests to the proper servers to balance the loads. One of the techniques involves packet rewriting, which is eventually delivering to the target servers [2,3]. The load unbalancing approach [4] to balance the loads across a cluster is based on the observation that using the correlation between the arrival rate and the processing rate is not necessarily a good approach; hence, resulting in uncorrelating the two. A hierarchical controller framework in [5] uses three levels of controllers to effectively manage the power consumed by a cluster, where forecast and operating environment parameters are used to manage the interactions between multilevels. While the dispatcher-based approach relays client requests directly to physical machines, server virtualization [6] offers an indirect buffering mechanism that abstracts away the underlying physical machines for workload management.Virtualization enables multiple virtual machines to run on a physical machine or host operating system [6]. Commercial virtualization techniques and products include EMC's VMware [7], Microsoft's Virtual Server [8], etc. Open-source and academic virtualization techniques include Xen and its hypervisor [9], User Mode Linux (UML) [10], Bochs [11], QEMU [12], etc. Hardware support for virtualization comes from Intel's Virtualization Technology [13] and AMD's AMD-V [14]. These virtualization technologies provide various means to abstract the resources and hence enable the users to have an illusi...