Internet Computer

How does the Internet Computer work, and where does my cloud engine fit?

All help topics

The Internet Computer (ICP) is a network of independent machines, called nodes, grouped into subnets; each subnet runs apps (canisters) as one replicated, tamperproof computer, and your cloud engine is a subnet of its own.

The pieces
  • Nodes. Physical machines in data centers, owned and run by independent node providers.
  • Subnets and replicas. Nodes are grouped into subnets. Every node of a subnet runs the replica software, keeps a full copy of each canister's code and data, and executes the same messages in the same order.
  • Consensus. The replicas agree on the order of messages, block by block. A change takes effect only once the subnet agrees, so a minority of faulty or malicious nodes cannot alter your app or its data.
  • Canisters. Your apps: WebAssembly code together with its own persistent state.
  • The NNS. The Network Nervous System is the network's governance system and registry: it records the nodes, which subnet each belongs to, and creates subnets. A subnet's responses are signed with its key, which can be checked against the NNS root key, so anyone can verify an answer came from the right nodes.
A cloud engine versus an ordinary subnet

Ordinary application subnets are shared: many developers' canisters run side by side on nodes the network assigns (typically 13 per subnet). A cloud engine is a dedicated subnet of the CloudEngine type:

  • You choose its nodes: their class, who provides them and where they are. Only your apps run on it.
  • Its cost schedule is free: nothing a canister does is metered in cycles. You pay for the nodes, from the engine's balance.
  • A few call rules differ as a result (no cycles attached to calls, bounded-wait calls across subnets), and services that live on other subnets, such as chain-key signing, vetKeys and exchange rates, are reached through a proxy canister.
  • It is a configuration, not a machine: nodes can be added, replaced or removed while your apps keep running.
  • The platform installs one canister of its own on every engine, the engine operator. It keeps the engine's team and roles and forwards your canisters' inference requests.
What replication means for capacity

Every node holds the whole app, so an engine's memory and compute are what one node provides, not the sum of its nodes. More nodes buy fault tolerance and more capacity for reads; a larger node class buys more memory and compute.

Learn more
Still stuck? Contact support