Improving software middleboxes and datacenter task schedulers

• Autor(a) principal: Barreto, Hugo de Freitas Siqueira Sadok Menna
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da UFRJ
• Texto Completo: http://hdl.handle.net/11422/11669
• topic: Engenharia elétrica; Caixas intermediárias; Agendadores de tarefas; Equidade; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA

Over the last decades, shared systems have contributed to the popularity of many technologies. From Operating Systems to the Internet, they have all brought significant cost savings by allowing the underlying infrastructure to be shared. A common challenge in these systems is to ensure that resources are fairly divided without compromising utilization efficiency. In this thesis, we look at problems in two shared systems—software middleboxes and datacenter task schedulers—and propose ways of improving both efficiency and fairness. We begin by presenting Sprayer, a system that uses packet spraying to load balance packets to cores in software middleboxes. Sprayer eliminates the imbalance problems of per-flow solutions and addresses the new challenges of handling shared flow state that come with packet spraying. We show that Sprayer significantly improves fairness and seamlessly uses the entire capacity, even when there is a single flow in the system. After that, we present Stateful Dominant Resource Fairness (SDRF), a task scheduling policy for datacenters that looks at past allocations and enforces fairness in the long run. We prove that SDRF keeps the fundamental properties of DRF—the allocation policy it is built on—while benefiting users with lower usage. To efficiently implement SDRF, we also introduce live tree, a general-purpose data structure that keeps elements with predictable time-varying priorities sorted. Our trace-driven simulations indicate that SDRF reduces users’ waiting time on average. This improves fairness, by increasing the number of completed tasks for users with lower demands, with small impact on high-demand users.