Skip to content

p2p.md

Latest commit

 

History

History
22 lines (16 loc) · 1020 Bytes

p2p.md

File metadata and controls

22 lines (16 loc) · 1020 Bytes
Raw

Recipe: P2P rollback (fighting / GGPO)

Peer-to-peer with no authoritative server. Each peer advances every tick on its own input plus a prediction of the remote's, and rolls back + replays when the real remote input arrives and differs.

const P2p = magnet.replication.P2p(State, Input, step, cap);

_ = peer.advance(local_input);             // predicts the remote, advances now
_ = peer.confirmRemote(tick, remote_input); // rolls back if the prediction was wrong
  • Like lockstep, this needs hard cross-machine determinism - step in fixed-point.
  • Unlike lockstep, it hides latency behind prediction instead of waiting for the peer.
  • Small input delay; tick 60 Hz.

To establish the connection through NATs first, see proto.conn.ice + host.pumpIce (nat_punch example); P2p then runs over whatever transport that yields. P2p is the simulation; ICE is the connectivity - they compose.

Runnable: examples/p2p_rollback.zig.