SmartDataCenter (SDC) is designed with a very specific vision of how a datacenter should be organized. The overall architecture diagram is as follows:
The following sections explain the diagram in more detail.
SmartOS as hypervisor
All systems in SDC use SmartOS as their hypervisor. This
means that most of the Global Zone's operating
system filesystem is read-only. Notable exceptions to this are
VMs are stored, and
/opt, where agents are installed.
The Global Zone is only ever used to run operator infrastructure. All customer activity is run in non-global zones. Non-agent SDC services are run in non-global zones - see the services section below.
SDC is architected so that the initial system booted in a datacenter is the headnode, which hosts all of the service VMs. This system is not intended to host customer VMs - it is for the administration of the rest of the nodes.
All non-headnode machines in the datacenter are compute nodes, which are meant for provisioning customer VMs. These nodes are booted over the network from the headnode.
SDC communicates with the Compute Nodes through agents that run in the Global Zone on each compute node. These agents serve a variety of functions including control operations for customer VMs, monitoring and log collection. The ur agent is the initial agent used for bootstrapping all other agents.
See the reference for more information about agents.
SDC requires two separate layer 2 subnets for operation. The admin network is used for internal SDC communication and network booting compute nodes. Since this is the control plane for SDC, customers are not allowed to provision VMs on this network. For security reasons, this network should not be routable.
The other required network is the external network. This is used for SDC VMs that require access to the internet. For example, the Global Zone may require internet access for NTP, and IMGAPI can be configured to fetch images from Joyent image servers. This is also the network that administrative endpoints (Admin UI and cloudapi) can be configured to listen on to avoid giving their users access to the admin network.
Customer networks are separate from the two standard SDC networks -- the number of them is limited by the number of VLANs.
Compute node Global Zones are not given externally routable IP addresses -- only addresses on the admin network. This is meant to minimize potential intrusion points and avoid using public IPs that could be used by customers. VMs provisioned on those compute nodes can still have nics with externally routable addresses, just not the Global Zone.
SDC employs a service-oriented architecture. Operator functionality is split into a number of different services, which can be administered and updated independently. The services are distributed as virtual machine images. This allows the service and its local dependencies to be updated as one unit, and allows for simple upgrades and rollbacks.
See the reference for more information about services.
External facing services
The external facing services are the administrative endpoints for SDC. CloudAPI provides an API for customers to administer their VMs, and Admin UI provides a web API for operators to administer SDC. They both communicate with the internal services to perform administrative actions.
These services are stateless - they rely on the data tier services for persistence.
The internal services offer a variety of functionality related to the operation of SDC. The external facing services and APIs communicate with them to perform tasks, and they in turn communicate with the agents on the Compute Nodes. Most of the services are REST-based, though notable exceptions include:
- dhcpd: used to TFTP boot Compute Nodes and assign them IP addresses
- rabbitmq: provides an AMQP message queue for communicating with Compute Node agents
- sdc: provides CLI tooling and miscellaneous administrative services
These services are stateless -- they rely on the data tier services for persistence.
Data tier services
The data tier services offer storage for the internal and external facing services:
These services use manatee, a Highly Available PostgreSQL service, for their storage.