This diploma thesis is devoted to a modern attitude of desktop computer usage, that uses operating system and applications virtualization. The paper describes theoretical principles of virtualization techniques and selection of the proper solution for the organization. Diploma thesis also describes the project implementation.

Virtualization is a key technology for cloud computing as it allows several operating system instances to run on the same machine, enhances resource manageability and enables flexible definition of billing units. Virtualization works by adding a software layer, a hypervisor, on top of the hardware platform. Virtual Machines, \emph{VMs}, are run on top of the hypervisor, which provisions hardwares resources to the VM guests. In addition to enabling higher utilization of hardware resources, the ability to move VMs from one host to another is an important feature. Live migration is the concept of migrating a VM while it is running and responding to requests. Since VMs can be re-located while running, live migration allows for better hardware utilization. This is because placement of services can be performed dynamically and not only when the are started. Live migration is also a useful tool for administrative purposes. If a server needs to be taken off-line for maintenance reasons, it can be cleared of services by live migrating these to other hosts. This thesis investigates the principles behind live migration. The common live migration approaches in use today are evaluated and common objectives are presented as well as challenges that have to be overcome in order to implement an ideal live migration algorithm. The performance of common live migration approaches is also evaluated and it is found that even though live migration is supported by most hypervisors, it has drawbacks which makes the technique hard to use in certain situations. Migrating CPU and/or memory intensive VMs or migrating VMs over low-bandwidth links is a problem regardless of which approach that is used. To tackle this problem, two improvements to live migration are proposed and evaluated, delta compression and dynamic page transfer reordering. Both improvements demonstrate better performance than the standard algorithm when migrating CPU and/or memory intensive VMs and migrating over low bandwidth links. Finally, recommendations are made on which live migration approach to use depending on the scenario and also what improvements to the standard live migration algorithms should be used and when.

The network architecture of the current Internet cannot accommodate the deployment of novel network-layer protocols. To address this fundamental problem, network virtualization has been proposed, where a single physical infrastructure is shared among different virtual network slices. A key operational problem in network virtualization is the need to allocate physical node and link resources to virtual network requests. While several different virtual network mapping algorithms have been proposed in literature, it is difficult to compare their performance due to differences in the evaluation methods used. In this thesis work, we proposed VNMBench, a virtual network mapping benchmark that provides a set of standardized inputs and evaluation metrics. Using this benchmark, different algorithms can be evaluated and compared objectively. The benchmark model separate into two parts: static model and dynamic model, which operated in fixed and changed mapping process. We present such an evaluation using three existing virtual network mapping algorithms. We compare the evaluation results of our synthetic benchmark with those of actual Emulab requests to show that VNMBench is sufficiently realistic. We believe this work provides an important foundation to quantitatively evaluating the performance of a critical component in the operation of virtual networks.

Data centers consume a significant amount of energy. This problem is aggravated by the fact that most servers and desktops are underutilized when powered on, and still consume a majority of the energy of a fully utilized computer even when idle This problem would be much worse were it not for the growing use of virtual machines. Virtual machines allow system administrators to more fully utilize hardware capabilities by putting more than one virtual system on the same physical server. Many times, virtual machines are placed onto physical servers inefficiently. To address this inefficiency, I developed a new family of packing algorithms. This family of algorithms is meant to solve the problem of packing virtual machines onto a cluster of physical servers. This problem is different than the conventional bin packing problem in two ways. First, each server has multiple resources that can be consumed. Second, loads on virtual machines are probabilistic and not completely known to the packing algorithm. We first compare our developed algorithm with other bin packing algorithms and show that it performs better than state-of-the-art genetic algorithms in literature. We then show the general feasibility of our algorithm in packing real virtual machines on physical servers.

Virtualization is the backbone of cloud infrastructures. Its core subsystems include hypervisors and virtual machine monitors (VMMs). They ensure the isolation and security of co-existent virtual machines (VMs) running on the same physical machine. Traditionally, driver domains -- isolated VMs in a hypervisor such as Xen that run device drivers -- use general-purpose full-featured OSs (e.g., Linux), which has a large attack surface, evident by the increasing number of their common vulnerabilities and exposures (CVEs). We argue for using the unikernel operating system (OS) model for driver domains. In this model, a single application is statically compiled together with the minimum necessary kernel code and libraries to produce a single address-space image, reducing code size by as much as one order of magnitude, which yields security benefits. We develop a driver domain OS, called Kite, using NetBSD OS's rumprun unikernel. Since rumprun is directly based on NetBSD's code, it allows us to leverage NetBSD's large collection of device drivers, including highly specialized ones such as Amazon ENA. Kite's design overcomes several significant challenges including Xen's limited para-virtualization (PV) I/O support in rumprun, lack of Xen backend drivers which prevents rumprun from being used as a driver domain OS, and NetBSD's lack of support for running driver domains in Xen. We instantiate Kite for the two most widely used I/O devices, storage and network, by designing and implementing the storage backend and network backend drivers. Our evaluations reveal that Kite achieves competitive performance to a Linux-based driver domain while using 10x fewer system calls, mitigates a set of CVEs, and retains all the benefits of unikernels including a reduced number of return-oriented programming (ROP) gadgets and advanced gadget-related metrics. General-purpose VMMs include a large number of components that may not be used in many VM configurations, resulting in a large attack surface. In addition, they lack intra-VMM isolation, which degrades security: vulnerabilities in one VMM component can be exploited to compromise other components or that of the host OS and other VMs (by privilege escalation). To mitigate these security challenges, we develop principles for VMM compartmentalization and specialization. We construct a prototype, called Redwood, embodying those principles. Redwood is built by extending Cloud Hypervisor and compartmentalizes thirteen critical components (i.e., virtual I/O devices) using Intel MPK, a hardware primitive available in Intel CPUs. Redwood has fifteen fine-grained modules, each representing a single feature, which increases its configurability and flexibility. Our evaluations reveal that Redwood is as performant as the baseline Cloud Hypervisor, has a 50% smaller VMM image size and 50% fewer ROP gadgets, and is resilient to an array of CVEs. I/O acceleration architectures, such as Data Plane Development Kit (DPDK) enhance VM performance by moving the data plane from the VMM to a separate userspace application. Since the VMM must share its VMs' sensitive information with accelerated applications, it can potentially degrade security. The dissertation's final contribution is the compartmentalization of a VM's sensitive data within an accelerated application using the Intel MPK hardware primitive. Our evaluations reveal that the technique does not cause any degradation in I/O performance and mitigates potential attacks and a class of CVEs. / Doctor of Philosophy / Instead of using software on a local device like a laptop or a mobile phone, consumers can access the same services from a remote high-end computer through high-speed Internet. This paradigm shift in computing is enabled by a remote computing infrastructure known as the "cloud,'' wherein networked server computers are deployed to execute third-party applications, often untrusted. Multiple applications are consolidated on the same server to save computer resources, but this can compromise security: a malicious application can steal co-existent applications' sensitive data. To enable resource consolidation and mitigate security attacks, applications are executed using a virtual machine (VM) -- an abstract machine that runs its own operating system (OS). Multiple VMs run on a single physical machine using two software systems: hypervisor and virtual machine monitor (VMM). They ensure that VMs are spatially isolated from each other, localizing security attacks. This dissertation focuses on enhancing the security of hypervisors and VMMs. The hypervisor and VMM have multiple responsibilities toward supporting the OS running on the physical computer and VMs. The OS runs software called device drivers, which communicate with input-output (I/O) hardware such as network and storage devices. Device drivers, usually written by third-party and I/O device manufacturers, are highly vulnerable to security attacks. To mitigate such attacks, device drivers are often run inside special VMs, called driver domains. State-of-the-art driver domains use a general-purpose full-featured OS such as Linux, which has a large code base (in the tens of millions of lines of code) and thus, a large attack surface. To address this security challenge, the dissertation proposes using lightweight, single-purpose VMs called unikernels, as driver domain OSs. Their code size is smaller than that of full-featured OSs by as much as one order of magnitude, which yields security benefits. We design and develop a unikernel-based driver domain, called Kite, for network and storage I/O devices. Kite uses NetBSD OS's rumprun unikernel for creating a driver domain OS. Using rumprun unikernel as a driver domain OS requires overcoming many technical challenges including a lack of support in a popular hypervisor such as Xen for performing I/O operations and communicating with rumprun, among others. Kite's design overcomes these challenges. Our empirical studies reveal that Kite is ten times less likely to be affected by future attacks and ten times faster to start than existing solutions for driver domains. At the same time, Kite domains match the performance of state-of-the-art driver domain OSs such as Linux. The hypervisor and VMM are responsible for creating VMs and providing resources such as memory, processing power, and hardware device access. Existing VMMs are designed to be versatile. Thus, they include a large number of components that may not be used in many VM configurations, resulting in a large attack surface. In addition, VMM components are not well spatially separated from each other. Thus, vulnerabilities in one component can be exploited to compromise other components. To address these security challenges, the dissertation proposes a set of principles for i) customizing a VMM for each VM's needs, instead of using one VMM for all VMs, and ii) strongly isolating VMM components from each other. We realize these principles in a prototype implementation called Redwood. Redwood is highly configurable and separates critical I/O components from each other using a hardware primitive. Our evaluations reveal that Redwood significantly reduces the VMM's size and VMM's vulnerabilities while maintaining performance. To enhance VM performance, I/O acceleration software is often used that eliminates communication overheads in the VMM. To do so, the VMM must share VMs' sensitive information with accelerated applications, which can potentially degrade security. The dissertation's final contribution is a technique that strongly isolates and limits access to sensitive information in the application using a hardware primitive. Our evaluations reveal that the technique improves security by localizing attacks without sacrificing performance.

The recent trend of migrating legacy computer systems to a virtualized, cloud-based environment has expanded to real-time systems. Unfortunately, modern hypervisors have no mechanism in place to guarantee the real-time performance of applications running on virtual machines. Past solutions to this problem rely on either spatial or temporal resource partitioning, both of which under-utilize the processing capacity of the host system. Paravirtualized solutions in which the guest communicates its real-time needs have been proposed, but they cannot support legacy operating systems. This thesis demonstrates the shortcomings of resource partitioning using temporally-isolated servers, presents an alternative solution to the scheduling problem called the KairosVM Flattening Scheduling Algorithm, and provides an implementation of the algorithm based on Linux and KVM. The algorithm is analyzed theoretically and an exact schedulability test for the algorithm is derived. Simulations show that the algorithm can schedule more than 90% of all randomly generated tasksets with a utilization less than 0.95. In comparison to the state-of-the-art server based approach, the KairosVM Flattening Scheduling Algorithm is able to schedule more than 20 times more tasksets with utilization of 0.95. Experimental results demonstrate that the Linux-based implementation is able to match the deadline satisfaction ratio of a state-of-the-art server-based approach when the taskset is schedulable using the state-of-the-art approach. When tasksets are unschedulable, the implementation is able to increase the deadline satisfaction ratio of Vanilla KVM by up to 400%. Furthermore, unlike paravirtualized solutions, the implementation supports legacy systems through the use of introspection. / Master of Science

In industry there has been a large focus on system integration and server consolidation, even for real-time systems, leading to an interest in virtualization. However, many modern hypervisors do not inherently support the strict timing guarantees of real-time applications. There are several challenges that arise when trying to virtualize a real-time application. One key challenge is to maintain the guest's real-time guarantees. In a typical virtualized environment there is a hierarchy of schedulers. Past solutions solve this issue by strict resource reservation models. These reservations are pessimistic as they accommodate the worst case execution time of each real-time task. We model real-time tasks using probabilistic execution times instead of worst case execution times which are difficult to calculate and are not representative of the actual execution times. In this thesis, we present a probabilistic hierarchical framework to schedule real-time virtual machines. Our framework reduces the number CPUs reserved for each guest by up to 45%, while only decreasing the deadline satisfaction by 2.7%. In addition, we introduce an introspection mechanism capable of gathering real-time characteristics from the guest systems and present them to the host scheduler. Evaluations show that our mechanism incurs up to 21x less overhead than that of bleeding edge introspection techniques when tracing real-time events. / Master of Science

The data centers are composed of resource-rich expensive server machines. A server, overloadeded with workloads, offloads some jobs to other servers; otherwise, its throughput becomes low. On the other hand, low-end embedded computers are low-power, and cheap OS-capable devices. We propose a system to use these embedded devices besides the servers and migrate some jobs from the server to the boards to increase the throughput when overloaded. The datacenters usually run workloads inside virtual machines (VM), but these embedded boards are not capable of running full-fledged VMs. In this thesis, we propose to use lightweight VMs, called unikernel, which can run on these low-end embedded devices. Another problem is that the most efficient versions of these boards have different instruction set architectures than the servers have. The ISA-difference between the servers and the embedded boards and the migration of the entire unikernel between them makes the migration a non-trivial problem. This thesis proposes a way to provide the unikernels with migration capabilities so that it becomes possible to offload workloads from the server to the embedded boards. This thesis describes a toolchain development process for building migratable unikernel for the native applications. This thesis also describes the alignment of the memory components between unikernels for different ISAs, so that the memory referencing remains valid and consistent after migration. Moreover, this thesis represents an efficient VM state transfer method so that the workloads experience higher execution time and minimum downtime due to the migration. / Master of Science / Cloud computing providers run data centers which are composed of thousands of server machines. Servers are robust, scalable, and thus capable of executing many jobs efficiently. At the same time, they are expensive to purchase and maintain. However, these servers may become overloaded by the jobs and take more time to finish their execution. In this situation, we propose a system which runs low-cost, low-power single-board computers in the data centers to help the servers, in considered scenarios, reduce execution time by transferring jobs from the server to the boards. Cloud providers run services inside virtual machines (VM) which provides isolation from other services. As these boards are not capable of running traditional VMs due to the low resources, we run lightweight VMs, called unikernel, in them. So if the servers are overloaded, some jobs running inside unikernels are offloaded to the boards. Later when the server gets some of its resources freed, these jobs are migrated back to the server. This back and forth migration system development for a unikernel is composed of several modules. This thesis discuss detail design and implementation of a few of these modules such as unikernel build environment implementation, and unikernel's execution state transfer during the migration.

Unlike infrastructure-based networks, mobile ad hoc networks consist of nodes independent of any infrastructure. Cooperation among these nodes is essential for the sustenance of multi hop communication. However, battery and bandwidth constraints may lead nodes in an ad hoc network to adopt energy- and bandwidth-conserving strategies. As routing and packet forwarding are end results of cooperation, network performance is affected when nodes in the network behave selfishly to conserve their resources. Our work involves characterizing selfish behavior by nodes in ad hoc networks and assessing the effectiveness of adopting tit for tat based strategies, which are meant to discourage selfish behavior in the network. We show that in an ad hoc network where other nodes act selfishly to conserve their resources, a node can benefit by adopting a generous tit for tat strategy. We also show that a node can gain benefit by avoiding selfish nodes in an ad hoc network, adopting a strategy that we call generous tit for tat with selfish avoidance (GTFT-SA) To analyze the effectiveness of cooperation strategies in selfish ad hoc networks we create an emulation environment based on virtualization. Such an emulation environment is more flexible to changes and is simpler to replicate than real life testbeds, while providing higher fidelity than simulations. / Master of Science

Containers are an alternative to virtual machines rapidly increasing in popularity due to their minimal overhead. To help facilitate their adoption, containers use management systems with central registries to store and distribute container images. However, these registries rely on other, preexisting services to provide load balancing and storage, which limits their scalability. This thesis introduces a new registry design for Docker, the most prevalent container management system. The new design coalesces all the services into a single, highly scalable, registry. By increasing the scalability of the registry, the new design greatly decreases the distribution time for container images. This work also describes a new Docker registry benchmarking tool, the trace player, that uses real Docker registry workload traces to test the performance of new registry designs and setups. / Master of Science / Cloud services allow many different web applications to run on shared machines. The applications can be owned by a variety of customers to provide many different types of services. Because these applications are owned by different customers, they need to be isolated to ensure the users’ privacy and security. Containers are one technology that can provide isolation to the applications on a single machine, and they are rapidly gaining popularity as they incur less overhead on the applications that use them. This means the applications will run faster with the same isolation guarantees as other isolation technologies. Containers also allow the cloud provider to run more applications on a single machine, letting them serve more customers. Docker is by far the most popular container management system on the market. It provides a registry service for containerized application storage and distribution. Users can store snapshots of their applications on the registry, and then use the snapshots to run multiple copies of the application on different machines. As more and more users use the registry service, the registry becomes slower, making it take longer for users to pull their applications from the registry. This will increase the start time of their application, making them harder to scale out their application to more machines to accommodate more customers of their services. This work creates a new registry design that will allow the registry to handle more users, and allow them to retrieve their applications even faster than what’s currently possible. This will allow them to more rapidly scale their applications out to more machines to handle more customers. The customers, in turn, will have a better experience.