Presenter

Mr. Hasan FAYYAD-KAZAN - ETRO dept. [Email]

Abstract

Over the last 10 years, virtualization has become much more widely popular as a result of fast and cheap processors. It is indeed one of the hottest trends in information technology today. It is a mechanism allowing multiple environments, called guest Virtual Machines (VMs), where each VM is an aggregation of an operating system and an application set, to be executed on the same physical machine simultaneously. It is provided by a software layer called Virtual Machine Manager (VMM) or hypervisor. The VMM virtualizes (abstracts) the platform making the underlying physical resources (such as CPU, storage, memory and I/O devices) shareable.

Virtualization has gained great acceptance and is successfully used with many business application classes where cloud computing is the most visual one. Recently, it started to be used for soft-real time media-based applications. Adding the hypervisor layer between the hardware and the VMs obviously incurs extra overhead compared to non-virtualized system. This overhead may vary from one hypervisor to another, making it difficult for users to choose the hypervisor that fits their requirements.
There have been a number of publications comparing hypervisors. These comparison analysis were, unsurprisingly, initiated by the hypervisor vendors themselves. For commercial reasons, they were mostly comparing hypervisors’ supported features where each vendor claims to offer the best services or features. More recently, a number of unbiased scientific papers originated focusing on the I/O performance, network throughput and CPU utilization using public benchmark tools which are not really adapted to the comparison job. As a consequence, most publications are not at all quantifying their statements others give only limited scientific proof. Therefore, there is no quantitative benchmark comparing the performance of the available hypervisors. Also, there is no focus on the hypervisor architecture (including the components and policies) and its effect on the VMs performance. Our contribution started from this point. We want to go beyond theoretical statements and quantify the performance differences in virtualization solutions. We provide a benchmark by using test metrics and scenarios to measure the VMs performance and behavior on top of three popular bare-metal hypervisors’ families: Microsoft (MS) Hyper-V Server, VMware vSphere ESXi and Xen. The achievements of the benchmark are the enumeration of the factors that influence the VM overhead such as the hypervisor scheduling algorithms, the number of simultaneously running VMs, their different type of workload and others. Moreover, the benchmark shows that the VM latencies are not only software- but also hardware-related. Finally, the benchmark results can help users make informed decisions about their choice between these hypervisors.

Apart from enterprise virtualization, recent years have seen increasing demand for virtualizing embedded systems due to the introduction of multi-core in such systems. It is deployed in embedded systems domain such as avionics systems and industrial automation where a strong emphasis on real-time performance is required. Currently there are several commercial real-time hypervisors and few open source ones where RT-Xen is the only mature and ready-to-use product.
We extended our research on two real-time hypervisors, RTS (commercial) and RT-Xen (open source). We end by giving advice about the right use of virtualisation technology for some embedded applications.

Short CV

Master of Applied Sciences and Engineering: Applied Computer Science, Lebanon