The use of clouds, both provided by commercial and R&E providers, are attractive to research groups that need limited time on processing resources. Integrating a cloud solution into your DMZ relies on knowing the network path into the cloud, and the performance expectations for data mobility.
Types of Cloud
Clouds come in many varieties, from commercial offerings at Amazon and Microsoft to R&E funded efforts such as Jetstream, Clouldlab, and Chameleon Cloud. Choosing the cloud provider that best fits your scientific workload is often the first step. From a networking perspective, the next choice comes down to how your users will interface with the cloud: figuring out the data mobility profile. The common case is to treat a cloud like any other computing resource; data will be sent in and results will be achieved. Depending on the relationship you have with the cloud provider, there may be charges for ingress or egress of data.
Establishing a peering with a cloud provider either directly, or via your regional or backbone network, is an efficient way to manage traffic flow. In doing so, you will avoid traversing potentially congested transit networks and thus increase the probability of a clean data transfer.
Tools such as Globus backend the data transfer engines on many cloud portals. It is recommended that the programatic API be used to facilitate automated data movement from local to remote sources when possible.
Validating performance to a cloud resource can be difficult due to the nature of network and host virtualization. In the general case, it is possible to see high levels of performance when a peering relationship is established with a cloud provider to eliminate sources of network friction, and the machines that are purchased within the cloud are given more available system resources.