31 Flavors Of Cloud
Amazon's Elastic Compute Cloud, or EC2, is the template for infrastructure-as a service. Customers create an Amazon Machine Image that contains the software and applications they want to run. They then use APIs to allocate resources, such as memory and storage. Price, based on aspects such as the OS used and RAM and disk space allocated to each AMI, starts at 10 cents per hour.
Amazon's adding more services, sometimes at extra cost, such as public-facing IP addresses that map back to a customer's account rather than to individual server instances. This setup lets customers start and stop virtual servers without disrupting access to a Web application. A load-balancing service shares network traffic loads among virtual servers. Just last month, Amazon started offering VPN connections between a company's data center and a bank of dedicated IP addresses inside Amazon's cloud. More on that later.
Some cloud infrastructure rivals are looking to compete with Amazon on price. Linux-based virtual servers from Rackspace, for example, start at 1.5 cents per hour, compared with Amazon's 10 cents.
Other rivals are specializing. Skytap focuses on software testing and development rather than vanilla Web hosting, and it charges customers per month rather than per hour.
IBM's Computing on Demand targets high-performance computing, such as simulations for the automotive and aerospace industries, or genomic modeling for life sciences. Customers get dedicated physical servers rather than virtualized instances, though companies can run virtual machines on top of those servers. Customers pay a $5,700 annual membership fee, then pay per CPU used.
Conventional hosting providers Savvis and Unisys also let customers mix and match physical servers with virtualized instances. With Savvis' Dedicated Cloud Compute, companies can buy dedicated physical servers run by Savvis, which creates a virtualization layer on top of these servers so customers can start and stop any number of VMs on demand. But the price is fixed based on the number of servers.
PropertyRoom uses Dedicated Cloud Compute to host its online auction application. CTO Dave Banks' main goals were to reduce costs, which he did by outsourcing hardware and OS maintenance to Savvis, and cut the number of physical servers hosting the application from 14 to six, which he did by using virtualization.
Savvis also offers Open Cloud Compute, a multitenant model like EC2, where multiple customers have their VMs running on shared hardware. Or companies can use some of both models--what CTO Doerr calls "fixed-plus-burstable capacity." Rackspace is trying the same strategy, adding a dedicated cloud service this summer alongside its public cloud, multitenant service.
Amazon's new Virtual Private Cloud stretches the model even further by letting customers create an isolated set of EC2 resources that connect via a VPN back to their corporate infrastructure. In a typical EC2 deployment, virtualized instances have internal and external IP addresses, with the external address letting the virtual instance communicate with the Internet. With VPC, which costs 5 cents per hour plus charges by data volume, virtual instances route all their traffic through the customer's own network infrastructure, including security systems such as firewalls and intrusion detection systems, before connecting to the Internet. Note that EC2 instances in a VPC subnet don't run on hardware dedicated to a single customer; they're isolated from other EC2 instances by private IP addresses, not by physical machines.
All these hybrid models are emerging to serve customers that aren't simply looking for the lowest-cost-per-hour provider. Many CIOs are uncomfortable with the multitenant cloud model because they don't want to run their workloads on hardware shared with other customers, fearing privacy or security mishaps. By melding dedicated hardware with on-demand virtual machines, or in the Amazon VPC concept by creating isolated subnets of computing resources routed through a company's own security infrastructure, they're seeking the cost savings and performance of clouds while trying to assuage customers' fears.
The dedicated hardware model also helps providers set up more predictable cash flow, since services with dedicated hardware typically require a monthly or annual contract and impose a base starting price, with consumption pricing added on. With four of our 12 providers, multitenant services require no minimum contract; customers pay only for what they use.
Only one of the providers on our list, Skytap, requires customers to run a minimum number of server instances--five per month. The others are trying to entice customers willing to take only baby steps into the cloud.
It will take a lot of those baby steps before most companies are ready for Zimory's cloud computing approach, a unique cloud computing model among our 12 providers. The startup provides software that lets groups of customer companies create "community clouds," in which they buy and sell excess computing capacity from the private clouds they've built inside their own data centers.
The Zimory application measures resource consumption based on a variety of factors such as memory, CPU, and storage, and the community itself defines availability of computing capacity, pricing, service-level agreements, and security agreements. Based in Berlin, it has a few European customers, including Deutsche Telekom Laboratories and Fraunhofer ITWN, an application development company. It plans to set up U.S. offices later this year.