Cloud Elasticity: going beyond demand as user load

Abstract

Cloud computing systems have become not only popular, but extensively used. They are supported and exploited by both industry and academia. Cloud providers have diversified and so did the software offered by their systems. Infrastructure as a Service (IaaS) clouds are now available from single virtual machine use cases, such as a personal server, to specialized high performance or machine learning engines. This popularity has been brought by the low-cost and risk-free feature of renting computing resources instead of buying them, in a large, one-time investment. Furthermore, clouds permit their clients the use of elasticity. Elasticity is the most relevant feature of cloud computing. It refers to the clients' ability to easily change the number of rented resources in a live environment. This permits the entire system to handle differences in load. Most cloud clients serve web applications or services to third parties. In these cases, load differences can be correlated to the number of users for the service and elasticity is used to handle differences in what is called user load. Most of the scientific literature approaches elasticity looking solely at user load. To give a clearer understanding, the majority of cloud frameworks in use today work as follows: they start a number of worker nodes, and tasks are assigned to them for execution. Only when the user load changes, the number of workers is adjusted, if any. In this paper, we propose an alternative approach, where the number of workers depends on the actual requirements coming from the different execution steps of an application. We show such an idea can be achieved for several workflows from different fields and that it can bring significant benefits to execution time and cost.