Chapter 12. Building Your Own Cloud with Eucalyptus
administration (e.g., credential distribution, revocation, role-definition policies),
or it can be integrated with the data center’s existing Active Directory or LDAP
installation.
12.3 Single-cluster Eucalyptus Cloud
We illustrate the process of deploying a Eucalyptus private cloud in a single
computational cluster. One node (machine) in the cluster acts as the
head node
that hosts all of the web services that compose the Eucalyptus control plane.
In this configuration, all nodes except the head node host VMs. We call the
nodes that host VMs
worker nodes
. Cloud requests (made via HTTPS or the
Management Console) are fielded by the various services on the head node and,
once authenticated and determined to be feasible, are forwarded to one or more NCs
running on worker nodes for actuation. Similarly, when a request is terminated,
the head node s ends notice of the termination to all NCs that must deallocate
resources associated with the request. The request is fully terminated when all
NCs report successful deallocation.
This configuration is useful for a supported production deployment in many
academic or research settings where a moderate-sized user po pul ation (e.g., an
instructional class, research group, or development team) shares a cluster, also
of moderate size (tens to hundreds of nodes). Note that the scalability of this
configuration is typically determined by the number of nodes and not the total
number of cores (separate CPUs) that each node comprises. Also, from a reliability
perspective, all VMs remain active and network reachable in the event the head
node fails or goes off line. No new cloud requests can be serviced while the head
node is down and some storage abstractions cease to function; but VM activity,
network connectivity, and access to ephemeral storage (which is local to each VM)
are not interrupted with a head node failure. Further, functionality is completely
restored when the head node is restored to functionality. Thus, this configuration,
which is rela tively simple to deploy and is portable to a wide variety of hardware
configurations, is capable of long-duration VM hosting.
A single-cluster configuration typically requires little data-center support:
commodity servers connected to a publicly routable subnet are sufficient to support
a cloud. The cloud administration effort required for such an installation is also low:
once the cloud is deployed, the cloud administrator is responsible for issuing us er
credenti als, man agin g resource quotas, and setting instance type configurations.
In an academic setting, this b urden is usually budgeted as a small fraction of a
local system administrator’s available time.
267