- 1. Evolution of Data Centers and Software Deployment
- 2. Defining Virtualization
- 2.1. Advantages of Virtualization
- 2.2. Types of Virtualization
- 2.3. Players on the Virtualization Market
- 2.4. The Anatomy of a Virtual Machine
- 3. VMware Infrastructure
- 4. The Downside of Virtualization
- 5. Conclusion
2. Defining Virtualization
VMware Inc., “the global leader in virtualization solutions” (Reuters, 2008 [was at: http://www.reuters.com/article/idUS96187+15-Sep-2008+BW20080915, site now defunct, July 2019]), defines virtualization as “the separation of a resource or request for a service from the underlying physical delivery of that service” (VMware, 2006). That is, virtualization provides a layer of abstraction between the physical resources such as CPUs, network or storage devices and the operating system. The separation of hardware functionality from the actual physical hardware allows the creation of an idealized virtual hardware environment in which an operating system can run without being aware that it is being virtualized. In fact, the virtualization layer, also called hypervisor, makes it possible to run multiple operating systems simultaneously on the same underlying hardware.
Each of these virtual machines (VM) stands for a complete system and has its own set of freely configurable virtual hardware. It is just like a physical computer – with network card, hard drive, memory and processor. And as with every computer, it is possible to install almost any kind of operating system on it. The underlying virtualization software, i.e. the virtualization layer is responsible for the emulation and mapping of the host’s physical resources to the virtual machine and its guest operating system. That approach makes virtual machines almost hardware-independent, so that it can be moved from one host to another without having to argue with compatibility issues.
2.1. Advantages of Virtualization
Using virtualization technologies can bring enormous cost savings to companies which are managing their own data centers. In times where services have to be up 24/7 and every second of downtime leads to declining profits, a stable IT environment is an absolute necessity.
Without a virtualized data center, companies often waste a lot of resources by running only one application per server. In order to minimize the risk of attacks or total system failures, they try to make sure that an application can in no way affect another service in terms of security and workload. A consequence of this “one-application per box” approach is that many servers have a very little degree of capacity utilization and expensive machines almost lie idle CITE. The whole IT infrastructure is often designed for a worst case scenario instead of being optimized for optimal hardware utilization. Virtualization can create an IT environment in which memory and CPU resources are balanced and ideally loaded with up to 60-80 percent of the host’s resources (Wolf et al., 2005).
Furthermore, the considerably high amount of required physical machines in a non-virtualized environment leads to big and cost-intensive data centers with high cooling and operational costs. Not only environmental organizations and activists, but also many IT managers increasingly ask for small and “green” data centers (iX, 2008).
Another common challenge in traditional IT data centers is the scheduling and minimization of planned downtimes in which hardware can be repaired or software can be modified. Since virtual machines can be copied by simply cloning them, one can safely perform security updates or any kind of system changes in the exact same environment as the production machine is running. This allows to maximize server availability and uptime significantly. VMware for instance claims to have customers that are running their VMs for over three years without a second of downtime (VMware, 2007 [was at: http://www.vmware.com/files/pdf/VMware_paravirtualization.pdf, site now defunct, July 2019]).
2.2. Types of Virtualization
In contrast to desktop virtualization in which an operating system is the basis for the abstraction layer (hosted), most server virtualization solutions do not need a full underlying host OS (bare-metal). The bare-metal or hypervisor architecture “is the first layer of software installed on a clean [..] system” (VMware, 2006). There are generally three well known approaches being used by the top virtualization projects, that is full virtualization, paravirtualization and hardware assisted virtualization.
They all have in common that they make use of the typical privilege levels of the processor architecture in order to operate directly on the physical hardware – in the supervisor mode. The major virtualization solutions support different processor architectures, but nearly all of them work with the x86 architecture. The x86 has a hierarchical ring architecture with four rings. Ring zero for privileged instructions of the operating system (kernel), ring one and two for device drivers and ring three for application instructions.
Full virtualization creates a fully simulated machine by placing the abstraction layer in the lowest ring and moving the guest operating system up to ring one. Every privileged instruction of the guest system is being trapped and rewritten (binary translation) by the virtualization software and then passed to the actual hardware. User applications can run their non privileged code directly and with native speed on the processor. The hypervisor simply passes the unmodified code to the CPU so that there is almost no virtualization overhead. Neither the guest OS nor the physical hardware are aware of the virtualization process which is why this approach supports the widest range of operating systems and hardware.
Unlike full virtualization, the technique of paravirtualization requires the guest OS to be aware of the fact that it’s being virtualized. In order to communicate with the virtualization layer, the guest system has to be modified. Instead of trapping all kernel instructions, the paravirtualized guest only handles and translates the non-virtualizable code and therefore runs much more instructions with near-native performance. Moreover, it can use the guest’s device drivers and therefore support a wide range of hardware. The downside is that since the kernel has to be modified, only open source operating systems can be used as a paravirtualized guest. A well known project using this technique is Citrix’s server virtualization product XenServer.
A fairly new approach is the hardware assisted virtualization. It is based on a new CPU mode (Intel VT, AMD-V) level below ring zero in which the virtualization layer can execute privileged instructions. Each non-virtualizable instruction call “is set to automatically trap to the hypervisor, removing the need for either binary translation or paravirtualization”. Due to the hardware based virtualization support, the guest operating system does not need to be modified and therefore mostly provides a better performance. A drawback of this approach is that the high number of traps leads to a high CPU utilization and hence might affect scalability (VMware, 2007).
2.3. Players on the Virtualization Market
The ongoing trend towards complete server virtualization solutions and the raising demand has led to a very competitive market. Even though VMware Inc., a company owned by EMC, still dominates the market, other big companies such as Microsoft or Citrix are actively pushing their own products.
The key product of the market leader VMware is a suite called VMware Infrastructure which was launched June 2006 and consists in the basic form of the hypervisor ESX, a managing software and backup functionalities. Additional features such as high availability functions or the ability to live migrate virtual machines can be purchased as add-ons. The ESX supports various operating systems and is a mature and production proven system.
Microsoft recently launched its Hyper-V hypervisor technology as a part of Windows Server 2008. The successor of Virtual Server 2005 is a relatively new technology that supports many different Windows versions and the Red Hat Enterprise Linux. At this point, it is not able to live migrate virtual machines and generally has a limited functionality. Since Microsoft’s product is much cheaper than for example VMware’s solution, it remains to be seen what product will be established (InformationWeek, 2008).
Citrix Systems, a virtualization specialist mostly known for the remote desktop application Metaframe/XenApp, competes with its server virtualization product called XenServer. It uses the open-source Xen hypervisor which is based on a Linux kernel and provides almost the same functionality as VMware Infrastructure.
2.4. The Anatomy of a Virtual Machine
After getting some insight in why virtualization is so important for today’s data centers and getting to know the main techniques, it might be interesting to know what components virtualization software typically uses and how it actually works.
Since the fundamental ideas of all virtualization solutions are similar, the main components of each software are the same and fulfill the same tasks. This paper will briefly describe the essential parts of a typical virtualization software and will take a closer look on VMware’s ESX/ESXi hypervisor.
2.4.1. The Hypervisor / Virtual Machine Monitor
As mentioned earlier, a hypervisor is the core technology of each virtualization software and is often referred to as the abstraction or virtualization layer. It “manages most of the physical resources on the hardware, including memory, physical processors, storage, and networking controllers” (VMware, 2007). Moreover, it includes other base functionalities such as scheduling and allows the creation of completely isolated VMs which then can be run within its scope. Depending on the architecture and implementation of the hypervisor, the extended functionality greatly varies. Each virtual machine is supervised and controlled by the virtual machine monitor (VMM). It implements the hardware abstraction and is responsible for the execution of the guest systems and the partitioning of available resources. Hence, the VMM is a part of the hypervisor that protects and isolates virtual machines from each other and decides which VM can use how many resources (Lo, 2005). Since the scope of functions overlaps and depends on the implementation, both terms are used as a synonym in most literature.
VMware provides two different bare-metal hypervisors, the ESX and the ESXi. Both are light-weight operating systems optimized for virtualization that use different versions of VMware’s proprietary system core VMkernel. The kernel controls RAM, CPU, NICs and other resources (resource manager) and provides access to the physical hardware (hardware interface layer). It also includes a small set of device drivers of the certified and ESX/ESXi compatible hardware. VMware intentionally tries to keep the host’s memory and disk footprint small in order to increase stability and minimize system failures. A small memory footprint also reduces the interference with running VMs and applications and has a direct impact on their speed (Vallee, 2008, was at: http://www.computer.org/portal/web/csdl/doi/10.1109/PDP.2008.85, site now defunct, July 2019).
As for the ESX, it can be additionally controlled by the service console, a full Red Hat Enterprise Linux distribution. The console provides tools to customize the ESX host and allows the installation of custom agents and scripts as on any other Linux system. Important to mention is that the ESX is not based on a Linux system (Zimmer, 2008). The service console is just an interface to access and configure the underlying system and can be thought of as a virtual machine with more access.
The ESXi is a slim version of the ESX which is given away by VMware for free. Apart from many other enterprise features, it lacks of the service console and the support for high availability, live migration and consolidated backups. The disk footprint of the ESXi is only 32 MB which makes it possible to run it from a USB stick or to directly embed it in servers. Most OEMs such as Dell, IBM or HP offer servers with an embedded version of ESXi.
Since both the ESX and the ESXi are very similar and are based on the same code, it can also be referred to as the ESX/ESXi hypervisor.
2.4.2. Resource Management
The host computer on which the virtualization software is running provides resources such as memory or processors. As any other operating system, the hypervisor is responsible for their management and the provision of interfaces to access them. In addition to the normal OS functionalities, it has to partition and isolate available resources and provide them to the running virtual machines.
Modern hypervisors like the ESX/ESXi treat a physical host as a pool of resources that allows their dynamic allocation to its virtual machines. Each resource pool defines a set of lower and upper limits for RAM and processor time. Virtual machines in a pool are only provided with its resources and cannot see whether there are more available. In a resource pool with 8×2 GHz and 8 GB RAM for instance, one could easily run two VMs with each 4 GHz and 2 GB RAM and another eight with each 1 GHz and 500 MB RAM.
2.4.3. Virtual Processors
Each VM is featured with one or many freely configurable virtual CPUs. Each virtual CPU behaves like a real processor, including registers, buffers and control structures. The virtual machine monitor makes sure that a VM only uses the resources that have been assigned to them and ensures a complete isolation of the context. Most virtualization solutions allow to equip VMs with more virtual CPUs than there are actually available. Extensions make it possible to use a physical multi-core environment by dynamically assigning CPUs to virtual machines. The exact behavior of the physical processors depends on which technology the virtualization software uses.
The ESX/ESXi hypervisor is an implementation of the full virtualization approach with few elements of paravirtualization. That is, the VMM directly passes and executes user-level code to the real CPU with no need for emulation (direct execution). This approximates the speed of the host and is much faster than the execution of privileged code. The guest OS code and application API calls are executed in the much slower virtualization mode (Zimmer, 2008). The hypervisor only supports the x86 architecture and makes high demands on the hardware, but is very flexible in terms of CPU time partitioning. The frequency of all physically available processors is simply summed up and can be freely assigned to any virtual machine. By assigning minimum and maximum boundaries, one can make sure that a VM does not eat up more CPU than allowed. The VMM is in charge of controlling these limits.
Even though the guest operating system does not need to be aware of the virtualization and perfectly works without any additional code, VMware provides a package called VMware Tools for the guest OS. If one installs the package on the guest system, it automatically provides optimized device drivers, e.g. for network cards or shared folders between host and guest.
3.4.4. Virtual Memory
The memory management of virtual machines is very complex and also handled by the VMM. Every running VM is provided with virtual RAM and has no direct access to the physical memory chips. The hypervisor is responsible for reserving physical memory for the assigned virtual RAM of all guest systems. Just like a normal operating system creates and manages page tables and a virtual address space for each application, the virtualization layer presents a contiguous memory space to the VM. The guest OS is not aware of the fact that every performed “physical” memory access is trapped by the VMM and translated to a real physical address. Hence, the actual process of mapping the VM’s memory to the host’s physical RAM is completely transparent.
In the scope of CPU virtualization, the total amount of assigned processor time has to be less or equal to the actual available amount. That is, it’s impossible to assign 2 and 3 GHz to two virtual machines of one host if the total available amount is only 4 GHz. In contrast to that, some hypervisors allow to assign more memory to a VM than is physically available. Similar to normal operating systems, the hypervisor uses a swap to extend the RAM. That means, a VM sees the amount of memory that has been assigned to it, no matter how much is really physically available. One could for example assign 3 GB RAM even if the host system provides only 2 GB.
Today’s operating systems have various swapping strategies but are typically programmed to swap only if it is really required. Hence an operating system naturally uses most of the memory that the hypervisor provides. Since the virtual machine’s guest does not know that the data in its “physical” RAM is being swapped, it rather fills the “fast” memory (Wolf et al., 2005). While this brings performance advantages in normal environments where the RAM is not being partially swapped, it leads to a dramatic loss of performance in virtualized operating systems when the limit of physically available RAM exceeds.
The ESX/ESXi hypervisor implements various functionalities to completely control and partition the available memory. In addition to the described over-commitment of memory, it includes a number of technologies to control the usage of RAM within the virtual machines. It might for instance be necessary to take back physical memory resources from a VM in order to give it to another one. Maybe one has added a new VM and likes to reserve at least 500 MB memory for it. At that moment, the VMM has to get back the physical resources from running VMs.
To do so, the hypervisor must be able to dynamically force VMs to start the swapping process instantaneously and release data from its virtual RAM. VMware solves this problem with a function called Memory Ballooning. To function correctly, it requires the installation of VMware Tools inside the guest system.
Another function, the Memory Idle Tax, bars the VM from keeping unaltered data too long in the memory. By charging more for idle memory pages, the guest OS will tend to swap them earlier than it had done it without the tax (Zimmer, 2008 and VMware, 2008).
2.4.5. Virtual Networking
Virtualization would hardly be useful at all if it was not possible to somehow network VMs together and connect them to a LAN or the Internet. Virtual machines can be connected to a network just like any physical machine. For this purpose, they can be equipped with virtual network interface cards (vNIC) which behave just like normal network cards. Each vNIC has its own MAC address and therefore its own identity on the network. The operating system communicates with the network card through its standard device drivers and therefore requires no changes in the kernel.
A virtual network is a network of VMs running on the same host. Each VM has one or more virtual network cards and is connected to a virtual switch (vSwitch). The host system provides at least one virtual switch that logically connects the virtual machines with each other. A vSwitch routes sending and receiving data from or to virtual machines internally. That is, any traffic between VMs does not leave the host computer and is not send though any physical cables. In addition to the virtual ports where VMs can connect to, a virtual switch has one or more uplink ports to which Ethernet interfaces of the host computer can be connected. That allows to interconnect the virtual network on the host computer with the outside network beyond the host’s NIC. For connections between an outside net and a virtual machine, the vSwitch behaves just like a physical switch.
Even though the basic tasks are the same as for physical switch boxes, there are few differences where virtual switches extend or reduce the functional range. VMware enforces a single-tier architecture in their virtual network, i.e. there is only one layer of virtual switches. Due to the fact that this does not allow interconnecting vSwitches, “the Spanning Tree Protocol (STP) is not needed and not present” (VMware, 2007). STP is used by physical switches to prevent the configuration of network loops. If more than one Ethernet interface is connected to the uplink ports, VMware provides a technique to use them as a team. The feature called NIC teaming allows the creation of a fail-safe and load-balanced network environment. The team of NICs can be configured to spread the network traffic and share the load to avoid bottleneck situations of data packets. At the same time, it creates a failover configuration in which the malfunction of one NIC does not lead to a total breakdown of the whole system (VI Online Library, 2008).
Another important feature of VMware Infrastructure and the ESX/ESXi hypervisor are Port Groups. A port group can be defined as a group of virtual machines connected to a virtual switch that form their own Virtual Local Area Network (VLAN). That is, one chooses the VMs that are supposed to be in the same virtual network and assigns them to a specific port group. Hosts of the same port group or VLAN can communicate with each other as if they were connected to the same switch, even if they are located in different physical locations. Hence, migrating a VM from one ESX host to another does not change anything in its location within the virtual network (Marshall, 2006).
2.4.6. Virtual Hard Drive Disks
Just like providing memory, network cards or CPU resources, it is possible to assign many different virtual hard drives disks (HDD) to a VM. Unlike other resources, hard drive virtualization is much more flexible. The hypervisor does not necessarily map every access to a physical HDD but creates virtual disks encapsulated in normal files. That is, a virtual disk is nothing more than a large file stored on a physical disk. The operating system within the VM accesses virtual disks like normal IDE or SCSI drives by using its internal OS mechanisms. Hence, they can be formatted with every file system and behave like a normal physical disk.
The major advantage of virtual disks is that in addition to normal HDD functionalities, they can be moved and copied like any other file. Therefore it is possible to store them locally on the host’s disk as well as on drives inside the network. The latter one is more typical especially in enterprise data centers. In addition to a virtual machine network, companies usually create a second so called storage LAN to connect network devices with each other and the hosts.
The ESX/ESXi hypervisor supports storing virtual HDDs on direct attached SCSI drives or on shared SAN, NAS or iSCSI storage. VMware developed a special file system for the network devices, the Virtual Machine File System (VMFS). It allows “multiple physical hosts to read and write to the same storage simultaneously [and] provides on-disk locking to ensure that the same virtual machine is not powered on by multiple servers at the same time”. That means that a single storage device can be used to store many virtual disks for VMs on different hosts. If a VM is running, VMFS makes sure that the virtual disk is locked until the VM is being powered down or crashes. If it fails, it can be restarted on another ESX by using the same disk file (VMware, 2008).