Читать книгу Mastering Microsoft Azure Infrastructure Services - Savill John - Страница 11

When I talk to people about Azure or even the public cloud in general, where possible I start the conversation by playing Titanfall (www.titanfall.com), a game published by Electronic Arts. The game is primarily a first-person shooter, but in addition to running around as a normal person, you get to pilot these massive robots, known as Titans, that are great fun to fight in. Unlike many other games, it is exclusively online and requires a large infrastructure to support the many players. There are many reasons I try to play Titanfall when starting my cloud conversations:

• I need the practice, as my teenage son will attest.

• I can write off the console and game because I use it in a business scenario.

• I can present a perfect example of a use case for the public cloud.

Why is Titanfall a perfect example of a use case for the public cloud? That is something that will become clear later in this chapter, but in the meantime, I definitely recommend supporting the public cloud and specifically Azure by playing lots of Titanfall.

Introducing the Cloud

Every organization has some kind of IT infrastructure. It could be a server sitting under someone’s desk, geographically distributed datacenters the size of multiple football fields, or something in between. Within that infrastructure are a number of key fabric (physical infrastructure) elements:

Compute Capacity Compute capacity can be thought of in terms of the various servers in the datacenter, which consist of processors, memory, and other hardware (such as the motherboard, power supply, and so on). I will use the term compute throughout this book when referring to server capacity.

Storage A persistent method of storage for data – from the operating system (OS) and applications to pure data such as files and databases – must be provided. Storage can exist within a server or in external devices, such as a storage area network (SAN). SANs provide enterprise-level performance and capabilities, although newer storage architectures that leverage local storage, which in turn replicate data, are becoming more prevalent in datacenters.

Network These components connect the various elements of the datacenter and enable client devices to communicate with hosted services. Connectivity to other datacenters may also be part of the network design. Options such as dedicated fibre connections, Multiprotocol Label Switching (MPLS), and Internet connectivity via a DMZ are typical.

Datacenter Infrastructure An often overlooked but critical component of datacenters is the supporting infrastructure. Items such as uninterruptable power supplies (UPSs), air conditioning, the physical building, and even generators all have to be considered. Each consumes energy and impacts the efficiency of the datacenter as well as its power usage effectiveness (PUE), which provides a measure of how much energy a datacenter uses for computer equipment compared to the other aspects. The lower the PUE, the more efficient the datacenter – or at least the more power going to the actual computing.

Once you have the physical infrastructure in place, you then add the actual software elements (the OS, applications, and services), and finally the management infrastructure, which enables deployment, patching, backup, automation, and monitoring. The IT team for an organization is responsible for all of these datacenter elements. The rise in the size and complexity of IT infrastructure is a huge challenge for nearly every organization. Despite the fact that most IT departments see budget cuts year after year, they are expected to deliver more and more as IT becomes increasingly critical.

Not only is the amount of IT infrastructure increasing, but that infrastructure needs to be resilient. This typically means implementing disaster recovery (DR) solutions to provide protection from a complete site failure, such as one caused by a large-scale natural disaster. If you ignore the public cloud, your organization will need to lease space from a co-location facility or set up a new datacenter. When I talk to CIOs, one of the things at the top of the don't-want-to-do list is write out more checks for datacenters – in fact, write out any checks for datacenters is on that list.

In the face of increased cost pressure and the desire to be more energy responsible (green), datacenter design becomes ever more complex, especially in a world with virtualization. If the three critical axes of a datacenter (shown in Figure 1.1) are not properly thought out, your organization’s datacenters will never be efficient. You must consider the square footage of the actual datacenter, the kilowatts that can be consumed per square foot, and the amount of heat that can be dissipated expressed in BTU per hour.

Figure 1.1 The three axes of datacenter planning

If you get any of these calculations wrong, you end up with a datacenter you cannot fully utilize because you can’t get enough power to it, can’t keep it cool enough, or simply can’t fit enough equipment in it. As the compute resources become denser and consume more power, it’s critical that datacenters supply enough power and have enough cooling to keep servers operating within their environmental limits. I know of a number of datacenters that are only 50 percent full because they cannot provide enough power to fully utilize available space.

The Private Cloud and Virtualization

In the early 2000s as organizations looked to better use their available servers and enjoy other benefits, such as faster provisioning, virtualization became a key technology in every datacenter. When I look back to my early days as a consultant, I remember going through sizing exercises for a new Microsoft Exchange server deployment. When sizing the servers required that I consider the busiest possible time and also the expected increase in utilization of the lifetime of the server (for example, five years), the server was heavily over-provisioned, which meant it was also highly underutilized. Underutilization was a common situation for most servers in a datacenter, and it was typical to see servers running at 5 percent. It was also common to see provisioning times of up to six weeks for a new server, which made it hard for IT to react dynamically to changes in business requirements.

Virtualization enables a single physical server to be divided into one or more virtual machines through the use of a hypervisor. The virtual machines are completely abstracted from the physical hardware; each virtual machine is allocated resources such as memory and processor in addition to virtualized storage and networking. Each of the virtual machines then can have an operating system installed, which enables multiple operating systems to run on a single piece of hardware. The operating systems may be completely unaware of the virtual nature of the environment they are running on. However, most modern operating systems are enlightened; they are aware of the virtual environment and actually optimize operations based on the presence of a hypervisor. Figure 1.2 shows a Hyper-V example leveraging the VHDX virtual hard disk format.

Figure 1.2 A high-level view of a virtualization host and resources assigned to virtual machines

Virtualization has revolutionized the way datacenters operate and brought huge benefits, including the following:

High Utilization of Resources Complementary workloads are hosted on a single physical environment.

Mobility of OS Instances between Completely Different Hardware A single hypervisor allows abstraction of the physical hardware from the OS.

Potentially Faster Provisioning Faster provisioning is dependent on processes in place.

High Availability through the Virtualization Solution This ability is most useful when high availability is not natively available to the application.

Simplicity of Licensing for Some Products and OSs For some products and OSs, the physical hardware is allowed to be licensed based on the number of processor sockets, and then an unlimited number of virtual machines on that hardware can use the OS/application. Windows Server Datacenter is an example of this kind of product. There is also an opposite situation for some products that are based on physical core licensing, which do not equate well in most virtualized environments.

There are other benefits. At a high level, if it were to be summed up in five words, I think “more bang for the buck” would work.

The potential of the datacenter capabilities can be better realized. The huge benefits of virtualization on their own do not completely revolutionize the datacenter. Many organizations have adopted virtualization, but have then operated the datacenter as if each OS is still on dedicated hardware. New OS instances are provisioned with dedicated virtualization hosts and even dedicated storage for different projects, which has resulted in isolated islands of resources within the datacenter. Once again, resources were wasted and more complex to manage.

In this book, I’m going to talk a lot about “the cloud.” But, for on-premises environments, I would be remiss if I didn’t also talk about another big change – the private cloud. Some people will tell you that the private cloud was made up by hypervisor vendors to compete against and stay relevant in the face of the public cloud. Others say it’s a revolutionary concept. I think I fall somewhere in the middle. The important point is that a private cloud solution has key characteristics and, when those are implemented, benefits are gained.

A customer once told me, “Ask five people what the private cloud is, and you will get seven different answers.” While I think that is a very true statement, the US National Institute of Standards and Technology (NIST) lists what it considers to be the five critical characteristics that must be present to be a cloud. This applies to both private clouds and public clouds.

On-Demand Self-Service The ability to provision services, such as a virtual machine, as needed without human interaction must be provided. Some organizations may add approval workflow for certain conditions.

Broad Network Access Access to services over many types of networks, mobile phones, desktops, and so on must be provided.

Resource Pooling Resources are organized in a multitenant model with isolation provided via software. This removes the islands of resources that are common when each business group has its own resources. Resource islands lead to inefficiency in utilization.

Rapid Elasticity Rapid elasticity is the ability to scale rapidly outward and inward as demands on services change. The ability to achieve large-scale elasticity is tied to pooling all resources together to achieve a larger potential pool.

Measured Service Clouds provide resources based on defined quotas, but they also enable reporting based on usage and potentially even billing.

The full document can be found here:

http://csrc.nist.gov/publications/nistpubs/800-145/SP800-145.pdf

People often say there is no difference between virtualization and private cloud. That is not true. The difference is the management infrastructure for a private cloud enables the characteristics listed here. To implement a private cloud, you don’t need to change your hardware, storage, or networking. The private cloud is enabled through software, which in turn enables processes. You may decide that you don’t want to enable all capabilities initially. For example, many organizations are afraid of end-user self-service; they have visions of users running amok and creating thousands of virtual machines. Once they understand quotas and workflows, and approvals, they understand that they have far more control and accountability than manual provisioning provided.

Enter the Public Cloud

The private cloud, through enhanced management processes and virtualization, brings a highly optimized on-premises solution. Ultimately, it still consists of resources that the organization owns and has to house year-round. As I mentioned earlier, CIOs don’t like writing checks for datacenters, no matter how optimal. All the optimization in the world cannot counter the fact that there are some scenarios where hosting on-premises is not efficient or even logical.

The public cloud represents services offered by an external party that can be accessed over the Internet. The services are not limited and can be purchased as you consume the service. This is a key difference from an on-premises infrastructure. With the public cloud, you only pay for the amount of service you consume when you use it. For example, I only pay for the amount of storage I am using at any moment in time; the charge does not include the potential amount of storage I may need in a few years’ time. I only pay for the virtual machines I need turned on right now; I can increase the number of virtual machines when I need them and only pay for those extra virtual machines while they are running.

Turn It Off!

In Azure, virtual machines are billed on a per-minute basis. If I run an 8-vCPU virtual machine for 12 hours each month, then I only pay the cost for 12 hours of runtime. Note that it does not matter how busy the VM is. You pay the same price whether the vCPUs in the VM are running at 100 percent or 1 percent processor utilization. It’s important to shut down and deprovision from the Azure fabric any virtual machines that are not required to avoid paying for resources you don’t need. (Deprovision just means the virtual machine no longer has resources reserved in the Azure fabric.) The virtual machine can be restarted when you need it again. At that point, resources are allocated in the fabric automatically; the VM will start as expected.

In addition to the essentially limitless capacity, this pay-as-you-go model is what sets the public cloud apart from on-premises solutions. Think back to organizations needing DR services. Using the public cloud ensures there are minimal costs for providing disaster recovery. During normal running, you only pay for the storage used for the replicated data and virtual environments. Only in the case of an actual disaster would you start the virtual machines in the public cloud. You stop paying for them when you can fail back to on-premises.

There are other types of charges associated with the public cloud. For example, Azure does not charge for ingress bandwidth (data sent into Azure – Microsoft is fully invested in letting you get as much data into Azure as possible), but there are charges for egress (outbound) data. There are different tiers of storage, some of which are geo-replicated, so your data in Azure is stored at two datacenters that may be hundreds of miles apart. I will cover the pricing in more detail later in the book, but the common theme is you pay only for what you use.

If most organizations’ IT requirements were analyzed, you would find many instances where resource requirements for a particular service are not flat. In fact, they vary greatly at different times of the day, week, month, or year. There are systems that perform end-of-month batch processing. These are idle all month, and then consume huge amounts of resources for one day at the end of the month. There are companies (think tax accountants) that are idle for most of the year but that are very busy for two months. There may be services that need huge amounts of resources for a few weeks every four years, like those that stream the Olympics. The list of possible examples is endless.

Super Bowl Sunday and the American Love of Pizza

I’ll be up front; I’m English and I don’t understand the American football game. I watched the 2006 Super Bowl. After five hours of two minutes of action, a five-minute advertising break, and a different set of players moving a couple of yards, it’ll be hard to get me to watch it again. Nonetheless, it’s popular in America. As Americans watch the Super Bowl, they like to eat pizza, and what’s interesting is the Super Bowl represents a perfect storm for pizza ordering peaks. During the Super Bowl halftime and quarter breaks, across the entire United States, with all four time zones in sync, people order pizza. These three spikes require 50 percent more compute power for ordering and processing than a typical Friday dinnertime, the normal high point for pizza ordering.

Most systems are built to handle the busiest time, so our pizza company would have to provision compute capacity of 50 percent more than would ever normally be needed just for Super Bowl Sunday. Remember that this is 50 percent more than the Friday dinnertime requirement, which itself is much higher than is needed any other time of the week. This would be a hugely expensive and wasteful exercise. Instead Azure is used.

During normal times, there could be 10 web instances and 10 application instances handling the website and processing. On Friday nights between 2 p.m. and midnight, this increases to 20 instances of each role. On Super Bowl Sunday between noon and 5 p.m., this increases to 30 instances of each role. Granted, I’m making up the numbers, but the key here is the additional instances only exist when needed, and therefore the customer is charged extra only when the additional resources are needed. This elasticity is key to public cloud services.

To be clear, I totally understand the eating pizza part!

The pizza scenario is a case of predictable bursting, where there is a known period of increased utilization. It is one of the scenarios that is perfect for cloud computing. Figure 1.3 shows the four main scenarios in which cloud computing is the clear right choice. Many other scenarios work great in the cloud, but these four are uniquely solved in an efficient way through the cloud. I know many companies that have moved or are moving many of their services to the public cloud. It’s cheaper than other solutions and offers great resiliency.

Figure 1.3 The key types of highly variable workloads that are a great fit for consumption-based pricing

In a fast-growing scenario, a particular service’s utilization is increasing rapidly. In this scenario, a traditional on-premises infrastructure may not be able to scale fast enough to keep up with demand. Leveraging the “infinite” scale of the public cloud removes the danger of not being able to keep up with demand.

Unpredictable bursting occurs when the exact timing of high usage cannot be planned. “On and Off” scenarios describe services that are needed at certain times but that are completely turned off at other times. This could be in the form of monthly batch processes where the processing runs for only 8 hours a month, or this could be a company such as a tax return accounting service that runs for 3 months out of the year.

Although these four scenarios are great for the public cloud, some are also a good fit for hybrid scenarios where the complete solution has a mix of on-premises and the public cloud. The baseline requirements could be handled on-premises, but the bursts expand out to use the public cloud capacity.

For startup organizations, there is a saying: “fail fast.” It’s not that the goal of the startup is to fail, but rather, if it is going to fail, then it’s better to fail fast. Less money is wasted when compared to a long, drawn-out failure. The public cloud is a great option for startups because it means very little up-front capital spent buying servers and datacenter space. Instead, the startup just has operating expenditures for services it actually uses. This is why startups like services such as Microsoft Office 365 for their messaging and collaboration. Not only do they not need infrastructure, they don’t need messaging administrators to maintain it. Public cloud IaaS is a great solution for virtual machines. Once again, no up-front infrastructure is required, and companies pay only for what they use. As the company grows and its utilization goes up, so does its operating expenditure, but the expenditure is proportional to the business. This type of pay-as-you-go solution is also attractive to potential financers, because there is less initial outlay and thus reduced risk.

At the start of this chapter, I said that everyone should play Titanfall, and this is where it fits in. Titanfall has a large number of artificial intelligence (AI) players, which would be burdensome if their computations had to be performed on a player’s console. So, Titanfall leverages Azure to provide the services and processing needed. When lots of players are online, an increased number of environments are started in Azure. When fewer players are online, a decreased number of environments are required. Only the environments needed run, thus optimizing costs. The amount of infrastructure that would be required to host something like Titanfall would be immense, and leveraging the public cloud presents a perfect-use case, especially when the demand for the service will diminish after a few months as new games are released.

I see the public cloud used in many different ways today, and that adoption will continue to grow as organizations become more comfortable with using the public cloud and, ultimately, trust it. Key use cases today include but are not limited to the following:

Test and Development Test and development is seen by many companies as “low-hanging fruit.” It is less risky than production workloads and typically has a high amount of churn, meaning environments are created and deleted frequently. This translates to a lot of work for the IT teams unless the private cloud has been implemented.

Disaster Recovery As discussed, for most companies a DR action should never be required. However, DR capability is required in that extremely rare event when it’s needed. By using the public cloud, the cost to implement DR is minimal, especially when compared to costs of a second datacenter.

International DMZ I have a number of companies that would like to offer services globally. This can be challenging – having datacenters in many countries is hugely expensive and can even be politically difficult. By using a public cloud that is geographically distributed, it’s easy to offer services around the world with minimal latencies for the end users.

Special Projects Imagine I have a campaign or special analytics project that requires large amounts of infrastructure for a short period of time. The public cloud is perfect for this, especially when certain types of licensing (for example, SQL Server licensing) can be purchased as consumed and other resources are paid for only as required.

A Desire to Get Out of the Datacenter Business I’m seeing more companies that just don’t want to maintain datacenters anymore. These organizations will move as much as possible to the public cloud and maintain minimal on-premises infrastructure needed for certain services, such as domain controllers and file and print servers.

Types of Service in the Cloud

Throughout this chapter, I have talked about making services available on-premises with a private cloud and off-premises in the public cloud, but what exactly are these services? There are three primary types of service: Infrastructure as a Service (IaaS), Platform as a Service (PaaS), and Software as a Service (SaaS). For each type, the responsibilities of the nine major layers of management vary between the vendor of the service and the client (you). Figure 1.4 shows the three types of service and also a complete on-premises solution. There are many other types of as-a-Service, but most of the other types of services use one of these three primary types. For example, Desktop-as-a-Service really has IaaS as a foundation.

Figure 1.4 The key types of highly variable workloads that are a great fit for consumption-based pricing

IaaS can be thought of as a virtual machine in the cloud. The provider has a virtual environment, and you purchase virtual machine instances. You then manage the operating system, the patching, the data, and the applications within. Examples of IaaS include Amazon’s Elastic Computing 2 (EC2) and Azure IaaS, which offer organizations the ability to run operating systems inside cloud-based virtual environments.

PaaS provides a framework where custom applications can be run. Organizations only need to focus on writing the very best application within the guidelines of the platform capabilities, and everything else is taken care of. There are no worries about patching operating systems, updating frameworks, backing up SQL databases, or configuring high availability. The organization just writes the application and pays for the resource used. Azure is the classic example of a PaaS.

SaaS is the ultimate in low maintenance. The complete solution is provided by the vendor. The organization has nothing to write or maintain other than configuring who should be allowed to use the software. Hotmail, a messaging service, is an example of commercial SaaS. Office 365, which provides cloud-hosted Exchange, SharePoint, and Lync services accessed over the Internet with no application or operating system management for the organization, is an enterprise example.

Ideally, for the lowest management overhead, SaaS should be used, and then PaaS where SaaS is not available. IaaS would be used only if PaaS is not an option. SaaS is gaining a great deal of traction with services such as Office 365. PaaS adoption, however, is fairly slow. The primary obstacle for PaaS is that applications have to be written within very specific guidelines in order to operate in PaaS environments. Many organizations have custom applications that cannot be modified. Others don’t have the budget to change their applications, which is why IaaS is so popular. With IaaS, an existing virtual machine on-premises can be moved to the IaaS solution fairly painlessly. In the long term, I think PaaS will become the standard for custom applications, but it will take a long time.

IaaS can help serve as the ramp to adopting PaaS. Consider a multitiered service that includes a web tier, an application tier, and a SQL database tier. Initially, all these tiers could run as IaaS virtual machines. The organization may then be able to convert the web tier from Internet Information Services (IIS) running in an IaaS VM and use the Azure web role, which is part of PaaS. Next, the organization may be able to move from SQL running in an IaaS VM to using SQL Azure. Finally, the organization could rewrite the application tier to directly leverage Azure PaaS. It’s a gradual process, but the reduced overhead and increased functionality and resiliency at the end state is worth it.

I saw an interesting analogy using the various types of service put in the context of pizza services. (Yes, it’s a second pizza example in one chapter; I like pizza.) Take a look at Figure 1.5. No matter where you plan to eat the pizza or how you plan to have it prepared, the actual pizza ingredients are the foundation. Other services and facilities, such as assembling the pizza, having an oven, cooking the pizza, having a table, and serving drinks, are also required. But as we move up the levels of service, we do less and less. At the highest level of service, pizza at a restaurant, we just eat and don’t even have to wash up.

Figure 1.5 Various types of Pizza-as-a-Service

The analogy is not perfect. Ideally, I would have had the oven and power as the core fabric. Then, with IaaS, the oven and power would be provided, and I would supply the ingredients, and assemble and cook the pizza (maybe in a pizza cooking class). For PaaS, the dough, sauce, and cheese are provided as a base, and I just add the toppings I want. For SaaS, I eat what I’m given, but only the poshest restaurants can get away with serving whatever they want. I doubt that a pizza restaurant would do well with that model, but you get the idea of the types of service. As you progress through the types of as-a-Service, you are responsible for fewer and fewer elements and can focus on what you care about: the end service/application.

There is another key area in which the pizza analogy is not perfect. In the pizza world, as you progress up the service levels, the service gets better but the total cost increases. When I make a pizza from scratch at home, it’s cheaper than eating out at a restaurant. In the IT service space, this is likely not the case. From a total cost of ownership (TCO) for the solution, if I can buy a service like Office 365 as SaaS, that solution is likely cheaper than operating my own Exchange, SharePoint, and Lync solution on-premises when you consider the server infrastructure, licenses, IT admin, and so on.