Investigating resource interference and scaling on multitenant PaaS clouds
Panagiotis Patros, Stephen A. MacKay, Kenneth B. Kent, Michael Dawson · 2016
Platform as a Service (PaaS) clouds are capable of both transparently allocating computing resources as well as providing part of the software stack and related services to tenant applications that execute on a subset of available cloud VMs. To deal with increased load, PaaS clouds enable applications to scale out, by creating extra instances, or scale up, by adding resources to the existing instances. However, good scalability is not necessarily attainable; in this paper we investigate the reasons for this. In particular, we propose a mathematical model that describes CPU allocation per tenant depending on interference from other tenants on the same VM, which we use to make predictions and confirm them in a variety of experimental situations. Furthermore, using a set of cloud tenants that target specific resources, we propose and evaluate a methodology for profiling the resource-intensiveness of cloud applications that uses slowdown in the presence of a resource-intensive cloud burner.