Closer to the Cloud - A Case for Emulating Cloud Dynamics by Controlling the Environment

Cloud computing offers a wide range of benefits including potential cost-savings by leasing resources from cloud service providers and improved user experience by Geodistributing applications [17]. Cloud service providers typically share the platform and cloud resources across multiple users for better utilization. They lease resources on-demand to users in three broad flavors, namely, Infrastructure as service (IaaS)[1], Platform as Service (PaaS)[9] and Software as Service (SaaS) [8]. These shared resources make the cloud environment highly dynamic and often induce substantial performance variation in the cloud services [21], [22], [14]. A critical step in developing cloud based applications is the ability to test these solutions on a controlled but cloud-like environment. Also, it is important for commercial product vendors to profile the application performance in the presence of such cloud dynamics before a production deployment. While many testbeds [13], [12], [15], [18] provide the abstraction of cloud services via virtualization, control over the cloud environment is seldom available for systems that require precise emulation of the cloud dynamics. In this work, we emphasize the importance and need for testbeds that can provide repeatable and precise control over the environment for developing such latency sensitive applications. We substantiate our claim by presenting the case study of a multitier enterprise application called DayTrader [2] deployed on a large scale experimental testbed GENI [7]. We show using various controlled experiments on DayTrader and GENI, the importance and utility of controlling the environment (e.g., network and link parameters) in the testbeds to emulate cloud dynamics. We show the importance of emulating cloud dynamics to enable research on adaptive systems that help applications deal with performance variation in cloud services. We present our experience with performing controlled experiments on GENI and briefly discuss possible enhancements in such testbeds that can open up new avenues of research in cloud computing.