Eli M. Dow

A quantitative study of virtual machine live migration

Virtual machine (VM) live migration is a critical feature for managing virtualized environments, enabling dynamic load balancing, consolidation for power management, preparation for planned maintenance, and other management features. However, not all virtual machine live migration is created equal. Variants include memory migration, which relies on shared backend storage between the source and destination of the migration, and storage migration, which migrates storage state as well as memory state. We have developed an automated testing framework that measures important performance characteristics of live migration, including total migration time, the time a VM is unresponsive during migration, and the amount of data transferred over the network during migration. We apply this testing framework and present the results of studying live migration, both memory migration and storage migration, in various virtualization systems including KVM, XenServer, VMware, and Hyper-V. The results provide important data to guide the migration decisions of both system administrators and autonomic cloud management systems.

Leveraging virtualization to optimize high-availability system configurations

Leveraging redundant resources is a common means of addressing availability requirements, but it often implies redundant costs as well. At the same time, virtualization technologies promise cost reduction through resource consolidation. Virtualization and high-availability (HA) technologies can be combined to optimize availability while minimizing costs, but merging them properly introduces new challenges. This paper looks at how virtualization technologies and techniques can augment and amplify traditional HA approaches while avoiding potential pitfalls. Special attention is paid to applying HA configurations (such as active/active and active/passive) to virtualized environments, stretching virtual environments across physical machine boundaries, resource-sharing approaches, field experiences, and avoiding potential hazards.

The Emerging Role of Software Testing in Curricula

In order to produce quality products, companies require new engineering students that have good problem solving, debugging, and analysis skills. Many graduates enter the work force with exceptional development skills, but lack proficiency in test, debugging, and analysis skills. This is in part because academic curricula emphasize development at the expense of teaching software testing as a formal engineering discipline. The majority of curricula today emphasize the initial phases of a development life cycle, namely: requirements gathering, architecture design, and implementation. The skills which are retained in this area of test are often learned ad-hoc while working on solutions for an implementation-oriented course. The lack of formal test education among graduates forces industry to spend substantial resources to properly educate graduates in the art and science of software testing. The contribution of this paper to the literature includes an evaluation of software testing as an industry profession, a survey of current curricula guidelines, a survey of software testing education in practice today, and a discussion of ongoing efforts to advance the status of software testing in academic curricula through a novel, crowd-sourced, industry-expert, approach to software test education.

Virtual machine migration plan generation through A∗ search

Modern virtual machine (VM) management software includes components that enable consolidation of VMs for power savings or load-balancing for performance. While the existing literature provides various methods for computing a better load-balanced, or consolidated goal state, it fails to adequately suggest the best path from the systems current state to the desired goal allocation. This paper discusses an approach to efficient path finding in VM placement plan generation for cloud computing environments. Initial results indicate our prototype solution is viable for managing hundreds of VMs through the application of A* search.

Decomposed multi-objective bin-packing for virtual machine consolidation

In this paper, we describe a novel solution to the problem of virtual machine (VM) consolidation, otherwise known as VM-Packing, as applicable to Infrastructure-asa-Service cloud data centers. Our solution relies on the observation that virtual machines are not infinitely variable in resource consumption. Generally, cloud compute providers offer them in fixed resource allocations. Effectively this makes all VMs of that allocation type (or instance type) generally interchangeable for the purposes of consolidation from a cloud compute provider viewpoint. The main contribution of this work is to demonstrate the advantages to our approach of deconstructing the VM consolidation problem into a two-step process of multidimensional bin packing. The first step is to determine the optimal, but abstract, solution composed of finite groups of equivalent VMs that should reside on each host. The second step selects concrete VMs from the managed compute pool to satisfy the optimal abstract solution while enforcing anti-colocation and preferential colocation of the virtual machines through VM contracts. We demonstrate our high-performance, deterministic packing solution generation, with over 7,500 VMs packed in under 2 min. We demonstrating comparable runtimes to other VMmanagement solutions published in the literature allowing for favorable extrapolations of the prior work in the field in order to deal with larger VMmanagement problem sizes our solution scales to. Subjects Autonomous Systems, Operating Systems

A Host-Agnostic, Supervised Machine Learning Approach to Automated Overload Detection in Virtual Machine Workloads

This paper evaluates a mechanism for applying machine learning (ML) to identify over-constrained IaaS virtual machines (VMs). Herein, over-constrained VMs are defined as those who are not given sufficient system resources to meet their workload specific objective functions. To validate our approach, a variety of workload-specific benchmarks inspired by common Infrastructure-as-a-Service (IaaS) cloud workloads were used. Workloads were run while regularly sampling VM resource consumption features exposed by the hypervisor. Datasets were curated into nominal or over-constrained and used to train ML classifiers to determine VM over-constraint rules based on one-time workload analysis. Rules learned on one host are transferred with the VM to other host environments to determine portability. Key contributions of this work include: demonstrating which VM resource consumption metrics (features) prove most relevant to learned decision trees in this context, and a technique required to generalize this approach across hosts while limiting required up front training expenditure to a single VM and host. Other contributions include a rigorous explanation of the differences in learned rulesets as a function of feature sampling rates, and an analysis of the differences in learned rulesets as a function of workload variation. Feature correlation matrices and their corresponding generated rule sets demonstrate individual features comprising rule sets tend to show low cross-correlation (below 0.4) while no individual feature shows high direct correlation with classification. Our system achieves workload-specific error percentages below 2.4% with a mean error across workloads of 1.43% (and strong false negative bias) for a variety of synthetic, representative, cloud workloads tested.

A Host-Independent Supervised Machine Learning Approach to Automated Overload Detection in Virtual Machine Workloads

This paper evaluates a mechanism for applying machine learning (ML) to identify over-constrained IaaS virtual machines (VMs). Herein, over-constrained VMs are defined as those who are not given sufficient system resources to meet their workload specific objective functions. To validate our approach, a variety of workload-specific benchmarks inspired by common Infrastructure-as-a-Service (IaaS) cloud workloads were used. Workloads were run while regularly sampling VM resource consumption features exposed by the hypervisor. Datasets were curated into nominal or over-constrained and used to train ML classifiers to determine VM over-constraint rules based on one-time workload analysis. Rules learned on one host are transferred with the VM to other host environments to determine portability. Key contributions of this work include: demonstrating which VM resource consumption metrics (features) prove most relevant to learned decision trees in this context, and a technique required to generalize this approach across hosts while limiting required up front training expenditure to a single VM and host. Other contributions include a rigorous explanation of the differences in learned rulesets as a function of feature sampling rates, and an analysis of the differences in learned rulesets as a function of workload variation. Feature correlation matrices and their corresponding generated rule sets demonstrate individual features comprising rule sets tend to show low cross-correlation (below 0.4) while no individual feature shows high direct correlation with classification. Our system achieves workload-specific error percentages below 2.4% with a mean error across workloads of 1.43% (and strong false negative bias) for a variety of synthetic, representative, cloud workloads tested.

WAYFINDER: parallel virtual machine reallocation through A* search

Modern virtual machine (VM) management software enables consolidation of VMs for power savings or load-balancing for performance. While existing literature provides various methods for computing a better load-balanced, or consolidated goal state, it fails to adequately suggest the best path from the system’s current state to the desired goal allocation. This paper discusses an approach to efficient path finding in VM placement problems for cloud computing environments of moderate scale with results indicating the solution is reasonable for managing hundreds of VMs. We present an overview of known approaches to dynamic VM placement and discuss their shortcomings with respect to dynamic reallocation. We then describe a novel design and implementation of a heuristic search algorithm to determine optimal sequential migration plans to transition from a given VM-to-host allocation to an arbitrary desired allocation state. We then elaborate nuances of A* application to this domain along with our simulation-based validation approach. Finally, this work demonstrates a novel and highly effective technique for exploiting migration parallelism in order to rapidly achieving VM reallocation convergence suitable for continual workload rebalancing in cloud data centers.