On the Elasticity of Parallel Components in a Cloud of High Performance Computing Services

Abstract

Cloud computing offers virtually unlimited set of resources and flexibility to allocate them through elasticity. But cloud limitations, such as the complexity of configuration and environment dynamicity, may jeopardizes the assurance of QoS requirements. HPC Shelf is a cloud of HPC services that employs a component-oriented architecture to describe hardware and software resources of parallel computing systems. We design a framework for HPC Shelf that employ cloud elasticity concepts for keeping the values of QoS metrics of parallel computing systems inside an acceptable range, enabling adaptations to fulfill the QoS contract restrictions. In our evaluation, using a linear algebra application, we show how HPC Shelf takes advantage of cloud elasticity to reinforce QoS requirements, rectifying assumptions from ill-defined QoS models.