Arktos: A Hyperscale Cloud Infrastructure for Building Distributed Cloud

Abstract

Scalability and management cost in cloud computing are few of the top challenges for the cloud providers and large enterprises. In this paper, we present Arktos, a cloud infrastructure platform for managing large-scale compute clusters and running millions of application instances as containers and/or virtual machines (VM). Arktos is envisioned as a stepping-stone from current “ single-region” focused cloud infrastructure towards next generation distributed infrastructure in the public and/or private cloud environments. We present details related to the Arktos system architecture and features, important design decisions, and the results and analysis of the performance benchmark testing. Arktos achieves high scalability by partitioning its architecture into two independent components, the resource partition (RP) and the tenant workload partition (TP), with each component scaling independently. Our performance testing using a benchmark tool demonstrates that Arktos with just two RPs and two TPs system setting can already manage a cluster of 50K compute nodes and is able to run 1.5 million workload containers with 5 times system throughput (QPS)1 compared with an existing container management system. Three key characteristics differentiate Arktos from other open source cloud platforms such as OpenStack and Kubernetes. Firstly, Arktos architecture is a truly scalable architecture that supports a very large cluster by scaling to more RPs and TPs in the system, Secondly, it unifies the runtime infrastructure to run and manage both VM and container applications natively, therefore eliminating the cost of managing separate technology stacks for VMs and containers. Lastly, Arktos has a unique “ virtual cluster” style multi-tenancy design that provides both strong tenancy isolation, including network isolation and transparent resource view.