Content area
Abstract
A virtual container is a computing unit that cuts out unnecessary libraries from classical virtual machines. The benefits of a containerized system include more efficient processing, and better resource utilization of distributed infrastructure. Large-scale adoption of virtual containers has stimulated concerns by practitioners and academics about the viability and reliability of data acquisition due to the decreasing window to gather relevant data points. A hallmark of a virtual container system is one in which individual containers are built, fulfill their functionality, and are then put back into a large resource pool. This means a container exists for only a few seconds in many cases.
This dissertation provides a solution for the viability and reliability of acquiring data from a container system in two ways. First, the process to construct a data stream is defined to gather data from points throughout a distributed computing environment through the use of introspection tools, which are able to acquire data from a system as it is running. The variety of data from the environment can be transformed into a more homogeneous form and visualized to provide both a health monitoring and forensic analysis system in a containerized environment. This process is the focus the second contribution, as well as novel visual comparisons between containerized workloads.
Three separate experiments were performed to show the proposed process for gathering viable data needed for health monitoring and forensics analysis is repeatable and scalable. First, an initial study focusing on how to gather data from a container engine using an introspection tool was performed. The second experiment focused on collecting information from a cloud stack, which incorporated a swarm of containers managed by different resource managers. These resource managers oversee containerized versions of popular software, and allocate memory, storage, and compute resources as they are needed within a computational cluster. The containers, using different resource managers, executed various types of jobs; the goal is to demonstrate that meaningful statistics can be captured that provide insight about the system. Finally, the third experiment demonstrates the usefulness of using visualizations to do a post-analysis to determine how different jobs and stacks of systems lead to different resource utilization; issues with default statistics collected by a popular introspection tool is also highlighted.
The contribution from the dissertation itself is two-fold. First, the process to construct the data pipeline utilized to construct the images combines metrics from various levels of a large-scale containerized system. Second, the images themselves are built on a series of transformations that quantify different levels within a container execution environment.
