Elastipipe: On providing cloud elasticity for pipeline-structured applications

Abstract

Although offering clear benefits for Web and business-critical demands, the use of cloud elasticity still imposes challenges when trying to reap its benefits over complex applications such as those modeled as pipelines. This often happens because replication, the standard technique for resource reorganization, by default, doesn’t address both function-level parallelism and communication between working VMs. Taking into account this background, here we are proposing a model named Elastipipe to provide automatic elasticity over pipelines-structured applications. Our main goal is to reduce total execution time for a set of tasks in a way that is effortless to cloud users, eliminating the need for any preconfiguration. Elastipipe’s contribution consists in a framework designed to provide a new concept named flexible superscalar pipelines, in which scaling operations and load balancing take place over different elasticity units. An elasticity unit refers to a set of one or more stages of a pipeline that will be grouped together under the same elasticity rules. Using a real workload from an IT Brazilian company, the Elastipipe prototype presented performance gains of up to 60% when confronted with a non-elastic approach. In addition, we demonstrated that the functional decomposition of pipeline stages (CPU-bound, memory-bound, and so on) in corresponding elasticity units was responsible for the best results in terms of performance and cost.