Adaptive Control Technique for Workload Management on a Shared Hosting Platform

When enterprise applications are consolidated on shared platforms, meeting application-level quality of service (QoS) goals becomes a challenge because complex interactions and dependencies exist among the components of multi-tier applications. Current consolidation tools just use system utilization metrics, and do not consider such dependencies when allocating resources to individual application components. We have developed an adaptive controller that can adjust dynamically resources for individual application tiers within a consolidated environment to meet the QoS goals of the applications. We have evaluated our controller design experimentally using multi-tier applications hosted on testbeds using Xen virtual machines. Results indicate that the adaptive controller is able to adjust multiple resources (e.g., CPU and network bandwidth) for multi-tier applications in order to maintain target QoS levels in scenarios with different platform and application configuration. Problem Statement We address the problem of dynamically allocating resources to complex, multi-tier enterprise applications in consolidated environments. Because application components typically run in multiple servers, these applications pose a number of unique challenges during consolidation. First, the application places different resource demands for different components, e.g., the web tier may consume CPU and network bandwidth, whereas the database tier mainly consumes I/O bandwidth. Second, the resource demands across tiers have workload and temporal dependencies, e.g., the database tier only serves connections established through the web tier. Finally, when multiple applications share resources, the QoS of one application can depend on the resource consumption of other applications if the platform does not provide complete performance isolation. These challenges make it extremely difficult to understand the sizing requirements of such applications to meet QoS goals, and there is substantial benefit in the ability to adjust dynamically resource allocations for such applications based on actual resource demand at run-time. Our Solution Consider a shared hosting platform as depicted in Figure 1(a), where a number of multi-tier applications share a common pool of server resources. Each application is composed of 3 tiers, and each tier is hosted inside a virtual container on a shared physical server. We further assume that tiers co-located on a physical server share resources, including CPU capacity, network bandwidth, and storage I/O capacity. Application performance (e.g., response time or throughput) is observed at the clients for QoS control. S A QoS Sensor 1 A S