A Simulator to Forecast Cost and Performance in the Cloud

A key question for a Chief Information Officer (CIO) would be the future run-time cost and performance of complex business application software, before deciding to migrate it to a cloud. It is difficult for CIOs to accurately estimate cloud cost and performance in a fast and inexpensive manner. In this article, we describe “Silverlining”, a simulator for estimating the cost and performance of a cloud service before migration, to help the CIO not only with go/no-go decisions but also with the budgeting for an appropriate cloud configuration. Silverlining: A Simulator to Forecast Cost and Performance in the Cloud step-by-step depiction of the system operation workload (resource usage). The Google cloud infrastructure, employed in the case study, uses the following two primary classes of compute resources (shown in Figure 3): The Google App Engine (GAE) and the CloudSQL Database Engine. The CloudSQL Database Engine was simulated using three cloud configurations (Low-cost/low-powerCPU server D1, medium-cost/medium-power-CPU server D16 and high-cost/high-power-CPU server D32). The low-cost/lowpower-CPU server D1 provided adequate throughput to satisfy the management’s goal. Silverlining Simulation Process In order to provide cloud forecasting capability, Silverlining our simulator must first be primed with appropriate base information (examples in Figures 4 and 5). For this purpose, cost and performance goals are obtained from stakeholder requirements for the system, oftentimes in terms of service level agreements (SLAs) (See Figure 1 Step1), and their interdependencies are analyzed by means of a notational convention, called Softgoal Interdependency Graph (SIG), which is intended for representing and reasoning about NFRs. Next the characteristics of the intended software application is estimated (Step 2), e.g., using varying workloads. The characteristics of the intended software application are then loaded as input into the simulator (Step 3/4), and the simulator would output (Step 5) the cost and performance estimates for executing the software system, with varying workloads and cloud configurations, on the cloud. With proper adjustments for differences from the standard (Step6), the data from the simulator can be used to estimate the cost and performance of the cloud, as well as choosing among the available cloud configurations according to the particular cost and performance business goals that a CIO may have. Now, an important question is if and how much Silverlining is reliable – i.e., the accuracy of the simulation results. For Silverlining, experiments for a typical application were run on Google cloud (called Google App Engine, or GAE), with varying workloads and cloud configurations (such as platforms and infrastructure characteristics), for a variety of benchmark data, and, using the same workloads and cloud configurations, the results from Silverlining were compared against the benchmark data. The comparison showed the two sets of data were very close for the typical application (or class of applications) that was used for the experiments. Of course, more experiments would be needed, in order for Silverlining to help a CIO and a cloud service provider assess and predict the cost and performance of a variety of (classes) of software applications, as well as choose among alternative cloud platforms and configurations, or, if needed, even adjust cost and performance business goals. In this paper, a case study of a Vehicle Management System (VMS) Display-Status is presented, which has been in operation for almost three decades and will continue to be for many decades to come. This particular VMS is supposed to manage close to 100,000 vehicles, while carrying out a variety of tasks, such as keeping track of their locations and status (e.g., in normal operation or maintenance, or in emergency repair, moving or stationed), scheduling their routes, allocating crews, dispatching them, compiling statistics, reporting on work progress, etc. Figure 1. Silverlining simulation modeling framework steps.