Proactive Management of Service Instance Pools for Meeting Service Level Agreements

Existing Grid schedulers focus on allocating resources to incoming jobs as per the resource requirements expressed directly by end users. This demands detailed knowledge of application behavior for different resource configurations on the part of end-users. Additionally, this model of job submission incurs significant delay in terms of the provisioning overhead for each request, for meeting the real requirement of an end-user – the invocation of an application service. In contrast, for interactive workloads, services are commonly pre-configured by an application server according to long-term steady-state requirements and the workload manager prioritizes execution requests coming from multiple end-users. In this paper, we propose a framework for bridging the gap between these two extremes, especially in the context of application services beyond simple interactive workloads, such as long running applications or a parallel numeric application. Such a service may be frequently requested and may need to be custom-configured to run over multiple nodes to meet enduser service level agreements (SLAs). In our approach, end users are shielded from lower-level resource configuration details and deal only with service metrics like desired throughput or average response time. These SLAs are then translated into concrete resource allocation decisions. Since multiple services share a common resource pool, and demand for a service fluctuates over time, static pre-configurations may not maximize utility of the resources. Our approach involves dynamic reprovisioning to achieve maximum utility, while accounting for overheads incurred during reprovisioning. We find that it is not always beneficial to re-provision resources according to perceived benefits and propose a model for calculating the optimal amount of re-provisioning for a particular scenario.