Locations, node groups and capacity

Group nodes into locations, choose how new servers are placed, build ordered node groups for billing products, and set how much of each node may be sold.

Every node belongs to a location, usually one datacenter. Customers and billing products order servers in a location, and the panel picks the node. A node group is an ordered list of nodes, possibly across locations, that a billing product can point at instead. Both stay within what each node may sell, which you set per node.

Locations#

  1. Go to Infrastructure › Locations › Add location.
  2. Enter a Name (such as Bucharest) and choose the Country (its flag is shown beside the node).
  3. Choose the Placement rule (below).
  4. Optionally turn on High availability and set Dead after (minutes). See High availability.
  5. Save.

Delete removes the location from the panel; it is refused while the location has nodes.

Placement rules#

The placement rule decides which node of the location a new server goes on, among those that can take it. A plan may set its own rule (the plan form's Placement rule), which then wins over the location's.

Rule Picks
Most memory left (default) the node with the most memory left to sell; spreads servers by memory
Most storage left the node with the most room on the storage the server's disk would go on
Fewest servers the node with the fewest servers, whatever their size
Round-robin each node in turn, by name, skipping those that cannot take the server
Fill first the node with the least memory left that still fits: fills one node before the next

A node can take a server only when it is online and answering, takes new servers (see Maintenance), has free addresses of the kind the plan gives (IPv4, IPv6, NAT), has a storage that fits the disks (with the plan's storage tag, if it has one), and has room in memory, vCPUs and disk. Every attempt, with the nodes looked at and why each was passed over, is in System › Placement log.

When no node of the location has room, resellers and customers read "This location has no room for a server of this size right now."

Node groups#

A node group is a name, a description and an ordered list of nodes. A server created in a group goes on the first node of the list that is online, answering, has the addresses the plan gives and room for the server; when that one is full, the next. It is what a billing product points at when you want to fill nodes in a fixed order.

To make one:

  1. Go to Infrastructure › Node groups › Add node group.
  2. Enter a Name (such as Europe) and an optional Description.
  3. Under Nodes, pick a node in Node to add and click Add; repeat. Use Up, Down and Remove to set the order. Placement follows it.
  4. Save.

Rules:

  • A node may be in several groups, and nodes of different locations may share a group. A server takes the location of the node it lands on.
  • A server is created in a group from a plan. A plan limited to some locations skips the group's nodes elsewhere.
  • The server keeps no reference to the group, so a group can always be deleted. A billing product that names a deleted group can no longer create servers until it names another.
  • With no node left with room: "Group Europe has no node with room for a server of this size right now."

Administrators create a server in a group from Create server with Place on: A node group. Billing modules list groups through the API (GET /node-groups) and send node_group_id when creating a server. See Billing modules. A node's page names its groups in its header.

Capacity and overselling#

A KVM server takes host memory as it uses it, and its vCPUs share the host's threads, so a node can be promised more than it has. Each node says how much may be sold, on its Settings tab under Capacity (see Node settings):

Resource What may be sold Default
Memory (memory − memory kept for the node) × memory sell % 4 GB kept, 100%
vCPUs CPU threads × vCPUs per thread 4 per thread
Storage each storage's size × its sell % 100%

What counts as sold is the size of every server on the node, in any state. The node's Overview shows, for memory, vCPUs and server storage, the node's figure, what is for servers, what is sold and what is in use now. A node sold beyond what it may is marked Oversold on its page, in the nodes list and on the dashboard. Nothing is stopped or moved.

Caps (most servers, memory, vCPUs, disk) can limit a node further; the smaller of the ratio and the cap applies. A node over a cap shows Over its cap.

How the limits are applied:

  • Resellers and customers only get a node where the server fits all of them.
  • Administrators choosing a node by hand are told what a server would oversell, with the figures, and can create it with Create anyway after ticking I know this oversells the node. The audit entry records by how much.
  • Everyone: a server never gets more vCPUs than its node has CPU threads.
  • Upgrades and disk resizes need room on the disk's own storage.

Storage is counted per storage: a plan's disks go on one storage that fits them together, and each storage has its own sell percent. A thin pool or ZFS storage sold above 100% is over-provisioned; see LVM and LVM-thin and ZFS. A shared storage (NFS or Ceph) is counted once across all the nodes attached to it.

Choosing the ratios#

  • Memory is the resource that hurts when oversold: when servers really use it the node swaps, and the kernel kills a server to free memory. Keep memory at 100% unless you know your servers' use, and keep enough for the node itself.
  • vCPUs are shared in time; 4 per thread is a common starting point for general-purpose servers. Lower it for CPU-heavy plans.
  • Storage fills as customers write. With qcow2 or thin disks an empty disk takes little space, so selling above 100% works until servers fill their disks. Watch the storage alerts.