Cloud computing is changing the way we provision hardware and software for on-demand capacity fulfillment. Lately I have been thinking about the ways on-demand servers, storage, and CDNs are changing the way we develop web applications and make business decisions. Gone are the days of idle CPUs, empty memory, or unused drive space. The cloud charges us for what we use as we use it (assuming capacity is available). In this post I will provide an overview of the cloud hosting landscape with a particular focus on cloud utilization by web companies. I will walk through a managed infrastructure stack and examine a few major business targets.

  1. The hardware
  2. The platforms
  3. The managed cloud stack
    1. High availability
    2. Security
    3. Stable, efficient OS
    4. Programming Language Business Logic
  4. The client layer
    1. Attached storage
    2. Database
    3. Cache
  5. Cloud consumers
    1. Web application developers
    2. Back office tasks
    3. Disaster recovery
  6. Summary

The hardware

In 1943 Thomas J. Watson of IBM famously proclaimed “there is a world market for maybe five computers.” Today we look back and laugh at such a proclamation but the statement really did hold up for approximately 10 years. Into the 1950s IBM designed computers for a possible market of 20 companies, of which 5 were expected to purchase such a machine. In 1953 IBM was pleasantly surprised to find 18 of 20 companies purchased the IBM 701, provind the business of back office processing and a new division for the tabulating giant.

Last week Rick Rashid of Microsoft was quoted as saying around 20 percent of the world’s servers are sold to a handful of companies: Microsoft, Google, Yahoo!, and Amazon. Three of those four companies are cloud resellers, renting small slices of their compute farms to businesses all over the world. 198 megawatt datacenters may be the new mainframe, with consumption units charged in minutes and bytes much like the time sharing relationships of the 1970s.

IBM again caught my interest last year with its Kittyhawk project from Jonathan Appavoo, Volkmar Uhlig, and Amos Waterland in New York. IBM is currently researching ways to repurpose the massively parallel Blue Gene supercomputers for the datacenters of the Web. It’s possible your future web application will run on a computer originally designed for gene sequencing and nuclear weapons testing.

Hardware and data operations are again consolidating towards major players. These specialist providers are building at a scale and specialization most web businesses can’t match. On-demand infrastructure of the cloud makes it cheaper and more efficient to outsource needed operational function to teams of experts already keeping some of the largest web companies in the world running every day.

The platforms

Microsoft and Google are the newest entrants into the cloud computing arena, focusing their efforts their respective programming languages of expertise. Microsoft‘s Windows Azure services platform will likely be the best platform for C# and ASP.Net development as it is tuned by the creators of .Net, IIS, and SQL Server. Google has similarly applied its expertise in the Python language and distributed web nodes to its Google App Engine product. The App Engine cloud is tuned by top contributors to the Python language including its BDFL, Guido van Rossum. App Engine utilizes custom Google software, Google Front End and Megastore, for web serving and storage. Cloud developers on either platform are using a similar set of hardware and software as the proven web-scale platforms of and Google. I expect Google App Engine will add support for Java in the near-future, their second major language offering and the most popular language among Google’s own services.

Language specialists are building managed stacks on top of generic cloud platforms such as Amazon Web ServicesEC2. Engine Yard sells a custom, managed AMI optimized for the Ruby language and its Rails framework. Rackspace’s Mosso subsidiary and others optimize for the latest versions of PHP + MySQL, attracting performance-minded applications in search of a tuned cloud instance. I am not aware of any major language contributors of Ruby or PHP employed at either company but the platforms do attempt to find their own niche among a broad offering of scalable hosting providers.

Amazon’s EC2 is the most well-known cloud computing provider and, as previously mentioned, the baseline service for other companies building value-added solutions. The AMI, a machine image formatted deployment in the Amazon cloud, is the basic building block of EC2 virtualization and the primary interaction point of Amazon’s customers. Amazon resells premium operating system and application packages on behalf of companies such as Microsoft, IBM, and Oracle but it’s possible such specializations will instead be absorbed by the software publishers themselves as they roll out their own hosted clouds (such as Azure or IBM Blue Cloud).

The cloud computing software stack is trending towards an integrated, managed experience maintained by some of the top contributors to each programming language and related components. More generic cloud platforms will need to stay up-to-date with managed technologies on their platform and/or establish a strong reseller relationship to more specialized cloud managers.

The managed cloud stack

Cloud computing stack

Managed cloud providers handle an entire stack of infrastructure needed to deliver web applications at scale. A solid cloud computing environment abstracts the basics of a computing environment away from the implementors and lets them focus on adding value with each new application. Managed cloud hosting providers need to offer the following basic layers to stay relevant in a web developer’s world.

High availability

Any web application needs to be available to legitimate visitors from all over the world. A true cloud creates spans the entire globe, defeating the speed of light on behalf of its customers with a server point of presence in multiple simultaneous locations. The cloud provider needs to effectively receive and route incoming requests to the appropriate virtualized application instance on behalf of its customers.

Google and Microsoft replicate each application instance to multiple physical locations. AT&T Synaptic Hosting spans multiple locations for its enterprise customers.


Web applications should be protected from intrusion and abuse at the network layer. In a cloud computing world application security is a lot like click fraud in advertising: every bad action carries a marginal cost. Cloud providers need to guard customers against potential external abuse and intrusion.

Google, Microsoft, and Amazon have their eyes on many incoming requests each day. Google serves App Engine requests off the same hardware handling Google Front End, keeping bad requests away from search, ads, and your apps.

Stable, efficient OS

Web applications rely on a stable, efficient operating system to interface with hardware, manage filesystems, and allocate resources. The cloud server operating system is a stripped down version of standard installations without a need for direct hard drive interfaces or other peripherals.

Amazon EC2 AMI quick start

Amazon EC2 highlights the operating system behind every machine image. Older versions of Fedora and Windows Server are the default “quick start” options available to each new account. Google and Microsoft clouds run on custom operating systems tailored for web use. Windows Azure is a stripped-down version of the latest Windows Server. Google runs a Linux-based OS tuned by its infrastructure team.

Programming Language Business Logic

Every managed cloud platform includes a dynamic language virtual machine and an appropriate web services gateway. Language functions too closely associated with the parent operating system and its libraries are stripped away, leaving only a pure operating environment for a machine interpreter. External dependencies such as GNU tools and custom compilers will not function within the cloud language abstraction layer. Cloud services bundle a dynamic language runtime into an easily spawned instance for standard and efficient interpretation across many application instances.

Google App Engine supports most functions of the Python language with additional support for the Django framework, WebOb, and PyYAML. Developers may replace these built-in libraries with newer or customized versions at an additional performance and usage cost. App Engine passes web requests into the programming language environment through the Web Server Gateway Interface.

The cloud client layer

Attached storage

Cloud applications don’t operate in a vacuum. Dynamic applications persist their application state and logic through database and file storage. In the cloud world the database and the file server are cloud services unto themselves, operating in an isolated and specialized layer. This isolation makes the storage layer swappable from the rest of the cloud stack and presents new opportunities for competition.

Static files fall into two major categories based on their planned consumption. Files under 1 MB in size can be consumed by most clients in a single request, matching the expected simple request/response model of the platform. Files over 1 MB in size need to be broken into more manageable parts, or ranges, for a sequenced download. Static cloud storage can be broken up into differing solutions by file size or file type, providing the best possible solution for the storage and delivery task at hand.

Google App Engine offers static file storage separate from its dynamic runtime. App Engine supports up to 1,000 files and has a 10 MB HTTP response limit.

Amazon Web Services offers static file serving through its Simple Storage Service (S3) origin server and CloudFront CDN services. Amazon allows private and public file storage and can even charge individual users of third-party services for their use through DevPay.

Attached storage is by far the most diverse service offering for companies evaluating a specialized solution. I prefer storage providers with widely supported file management APIs, smart settings for MIME types and caching HTTP headers, and a primary functionality of serving files our to the worldwide web. I expect popular storage providers will bundle more CDN services in the future through an exclusive up-sell partnership. I also expect a new class of storage middleware optimized for minimizing files, cleaning up images, or transcoding video will set up new programmable front-ends backed by popular storage providers.

Database storage

Databases are the preferred way of persisting structured data powering web applications. Cloud service providers have tuned and rewritten database functionality for the cloud, opening up new opportunities for scalable data services across multiple dynamic application instances. Cloud databases are distributed, replicated, and largely transactional. Cloud databases can be separated from the rest of the cloud stack through RESTful APIs between different vendors but there is a definite latency advantage to coupling of data and its interpreter.

Microsoft offers SQL Server as a web service as part of the Azure services stack. Google App Engine offers Megastore, an abstraction layer on top of BigTable, as a service API within an App Engine instance or as a separate remote API. Amazon’s SimpleDB brings together EC2 processing with S3 data storage. Greenplum offers PostgreSQL as a stand-alone cloud offering.

Cloud databases are typically more limited in functionality than their local counterparts. App Engine returns up to 1000 results. SimpleDB times out within 5 seconds. Joining records from two tables in a single query breaks databases optimized for scale. App Engine offers specialized storage and query types such as geographical coordinates.

The database layer of a cloud instance can be abstracted as a separate best-of-breed layer within a cloud stack but developers are most likely to use the local solution for both its speed and simplicity.


Our web applications receive multiple requests for the exact same resource. We should be able to place a pre-assembled version of our web pages, images, and XHR data into a local memory cache for fast serving on multiple requests. On our own servers we frequently use memcached, Varnish, Squid, etc. The cloud stack should include a storage cache as its first layer of request processing.

Google App Engine includes a memcache API written by Brad Fitzpatrick, creator of memcached. Windows Azure will supposedly support Velocity caching in the near future.

Cloud consumers

Building corporate clouds

The target market of a cloud computing platform will affect its stack completeness, feature sets, and future support. Cloud terminology seems to be thrown around as a magical buzzword but there are major usage cases emerging.

Web application developers

New web applications start small and may sometimes experience exponential growth on a worldwide basis. Web developers evaluating the cloud stack are likely starting from scratch without the concerns of switching from a legacy system or alternate implementation.

Cloud computing abstracts tiered architecture, operations planning, and other nuances from companies specializing in bring new ideas to market quickly. Web developers prefer a cloud stack tuned for fast web performance. Geographically distributed dynamic instances are important at least as an upgrade option to protect a new business from a rewrite at varying levels of scale.

I believe cloud providers offering a complete managed stack will attract web development specialists to their platform. Google App Engine, Mosso, and Windows Azure compete in this space.

Back office tasks

Enterprise applications are moving out of the local server closet and into the cloud. Medium- to large-sized companies are replacing in-house maintenance of machines and applications with software and infrastructure as a service. Project management, employee tracking, payroll, and many other common functions have made their way into the software-as-a-service realm. More customized applications will migrate to cloud hosting and take their place alongside the anchor tenants of the groupware and collaboration suites.

Windows Azure, Salesforce‘s, and Google App Engine show strong promise as integrated back office add-ons. Microsoft and Google already have a solid footing in enterprise groupware services through Exchange Online and Google Apps respectively. can be closely tied to the popular Salesforce CRM application for sales and marketing teams.

More generic back office functions can operate on any cloud hosting provider with a properly maintained disk image. A new class of hosting provider operates as an abstraction layer between multiple clouds by maintaining the appropriate images and deployment scripts for any given task. Companies such as Aptana, CohesiveFT, RightScale, and many others span multiple cloud hosting providers with a single management interface. Cloud management companies can monitor multiple providers and create spot pricing market for computing resources.

Back office solutions represent the largest possible growth area for cloud hosting providers. Platforms with strong existing anchor tenants can add on new services combining software-as-a-service and infrastructure-as-a-service. Generic cloud hosting providers will likely be tapped for general tasks directly or though a cloud management layer.

Microsoft is promoting its cloud hosting solutions through its partner channels. Microsoft partners receive a 12% commission on the first year of revenues and 6% commission on all future revenues. Google offers a 20% discount to Google Apps Authorized Resellers over the life of the account.

Excess capacity

Hosting solutions need to scale up to meet peak demand. Peak demand could occur for an hour each day, one day a year (Black Monday in the retail sector), or one month out of twelve (college basketball playoffs). Cloud computing lets businesses pay only for what they use when they use it. Servers are not sitting around in your datacenter depreciating in value and consuming resources while you wait for peak load to occur.

Excess capacity needs may be predictable and cyclical, allowing a business to integrate cloud computing into their computing workflow with ease. Generic cloud computing platforms offer the best migration costs as businesses clone their own local machine images for execution in a cloud computing environment.

Disaster recovery

Business operations need to stay online when catastrophe strikes. An earthquake in California, a hurricane in Florida or Texas, or a power outage anywhere in the world could knock your business offline instantly. A hot backup in the cloud spins up when your primary site is down. An on-demand backup facility is a lot cheaper than physical investments as companies invest in contingency planning.

Amazon Web Services recently introduced reserved machine instances for companies who must be absolutely sure they will be able to operate in an environment of strained cloud capacity. Reserved instances receive priority allocation of cloud resources in exchange for an upgrade fee and lower monthly usage charges. Reserved instances are the VIP treatments of the cloud hosting world.

Demand response programs are common in utility sectors such as electricity. Businesses can opt to be the last ones kicked off the grid in a low-capacity environment in exchange for higher consumption costs.


Cloud computing is picking up steam and there are a few early winners. The most promising solutions from large vendors are still in a technology preview stage but should be open for general use by the end of the year. Startups developing new applications should pick the best solutions provider based on the strength of their stack offering and usage pricing. Some cloud layers can easily be abstracted to best-of-breed solutions.

I hope you enjoyed this summary of the world of cloud hosting! There is a lot going on and this post just scratches the surface of how our computing world is changing.

6 replies on “The anatomy of cloud computing”

  1. Given the BOSS, SearchMonkey, Y!Pipes, Y!OS, and YAP services, i think it is fair to think also of Yahoo as a company offering cloud computing services.
    They just not opened them up for general use outside of Yahoo’s Web properties.
    And with Rasmus Lerdorf they are employing THE PHP guy, similar to Google employing Guido van Rossum.

  2. Finally someone who really understands what cloud computing is. We see too many article that can’t make the difference between cloud computing and software as a service. Thanks.
    Would be nice to also address new scalable database technologies such as Hadoop in the cloud – e.g. Cloudera -.

    1. Jean,
      I see Cloudera’s distribution for Hadoop as a deployed software instance that could accelerate growth by packaging its distribution for popular generic cloud platforms.
      Cloudera’s first announcement integrated Hadoop software into popular package management systems RPM and yum. They could package their Hadoop software inside a full software stack bundled as an AMI for deployment on EC2 if they wish.
      I believe Alexa has already released their Hadoop AMI package used in their web crawls.

  3. So the question is would/is cloud computing cheaper than lets say GoDaddy?
    I can buy multiple servers from GoDaddy at a price per month, but is that cost more or less than Microsoft/Amazon/Google.
    That’s the question I want answered.

    1. Scott,
      If you run applications on Google App Engine under your daily allotted quota your monthly payment is zero. Monthly fees and total cost of ownership will vary by installation. Multi-tenant shared hosting accounts and cloud computing instances such as Windows Azure or Google App Engine are very different offerings from an infrastructure point of view.

Comments are closed.