Improving Application Performance in the Emerging Hyper-converged Infrastructure

University of Minnesota Ph.D. dissertation.April 2019. Major: Computer Science. Advisor: David Du. 1 computer file (PDF); viii, 118 pages.In today's world, the hyper-converged infrastructure is emerging as a new type of infrastructure. In the hyper-converged infrastructure, service providers deploy compute, network and storage services on inexpensive hardware rather than expensive proprietary hardware. It allows the service providers to customize the services they can provide by deploying applications in Virtual Machines (VMs) or containers. They can have controls on all resources including compute, network and storage. In this hyper-converged infrastructure, improving the application performance is an important issue. Throughout my Ph.D. research, I have been studying how to improve the performance of applications in the emerging hyper-converged infrastructure. I have been focusing on improving the performance of applications in VMs and in containers when accessing data, and how to improve the performance of applications in the networked storage environment. In the hyper-converged infrastructure, administrators can provide desktop services by deploying Virtual Desktop Infrastructure application (VDI) based on VMs. We first investigate how to identify storage requirements and determine how to meet such requirements with minimal storage resources for VDI application. We create a model to describe the behavior of VDI, and collect real VDI traces to populate this model. The model allows us to identify the storage requirements of VDI and determine the potential bottlenecks in storage. Based on this information, we can tell what capacity and minimum capability a storage system needs in order to support and satisfy a given VDI configuration. We show that our model can describe more fine-grained storage requirements of VDI compared with the rules of thumb which are currently used in industry. In the hyper-converged infrastructure, more and more applications are running in containers. We design and implement a system, called k8sES (k8s Enhanced Storage), that efficiently supports applications with various storage SLOs (Service Level Objectives) along with all other requirements deployed in the Kubernetes environment which is based on containers. Kubernetes (k8s) is a system for managing containerized applications across multiple hosts. The current storage support for containerized applications in k8s is limited. To satisfy users' SLOs, k8s administrators must manually configure storage in advance, and users must know the configurations and capabilities of different types of the provided storage. In k8sES, storage resources are dynamically allocated based on users' requirements. Given users' SLOs, k8sES will select the correct node and storage that can meet their requirements when scheduling applications. The storage allocation mechanism in k8sES also improves the storage utilization efficiency. In addition, we provide a tool to monitor the I/O activities of both applications and storage devices in Kubernetes. With the capabilities of controlling client, network and storage with hyper-convergence, we study how to coordinate different components along the I/O path to ensure latency SLOs for applications in the networked storage environment. We propose and implement JoiNS, a system trying to ensure latency SLOs for applications that access data on remote networked storage. JoiNS carefully considers all the components along the I/O path and controls them in a coordinated fashion. JoiNS has both global network and storage visibility with a logically centralized controller which keeps monitoring the status of each involved component. JoiNS coordinates these components and adjusts the priority of I/Os in each component based on the latency SLO, network and storage status, time estimation, and characteristics of each I/O request