Reversing on the host turned the whole set around on the client, and the
locomotive ended up looking like something derailed making its way across the
scenery. Both come from the same place: a routine correction was calling
place_on_track(), which sets the axles from scratch and works from the
vehicle's own idea of which way it faces. That is right for a peer joining with
nothing to go on, and wrong for a train that is already running - the placement
reorients it.
A correction now slides the vehicle along the track it is already on, by the
difference between the two bogie offsets, which is the same operation the
physics performs on every step. When the two sides disagree about which piece
of track the vehicle is on, it is left alone until they agree; guessing there
is what put a locomotive beside the rails. place_on_track() is kept for the one
case it suits, a peer joining the session.
Also:
- taking a seat works with the game paused. The whole thing sat inside the
"has any time passed" branch, so with the world stopped the lobby button
changed the roster and nothing else happened
- the debug panel gets a Network section, shown only in a session: role, peer,
tick, bytes and messages each way with a running rate, two charts of kB/s so
the cost of a session can be read off, how far the worst vehicle is out of
place, and whether the state digest agrees. The separate "network" overlay
window is gone - everything it said lives there now, together with the lobby
and refusal messages that had nowhere to be seen
Until now the server ran the physics of everything and a client re-ran it from
the server's frame times, corrected a few times a second. That is the worst of
both: the client's own train never moved smoothly, its controls answered a
round trip late, and the two copies still sat metres apart between corrections.
Simulation authority now follows the train:
- the first person aboard a train runs its physics on their own machine, and
what they produce is what the rest of the session is told about it. A train
with nobody in it belongs to the host. Every peer works the owner out the
same way, from the crew of its lowest numbered car, so nobody has to be told
- a peer never applies corrections to a train it runs itself. Its own physics
is the authority on it, and overwriting that is precisely what made the
owner's ride stutter
- a client publishes the state of its own trains upstream, ten times a second.
The server takes it, filtered so a peer can only move the trains it is
entitled to move, and passes it on to everybody else in its own stream
- so the picture is: your train is smooth for you and may stutter for the
others; their trains may stutter for you and are smooth for them
Two things follow from it:
- commands for the train you run take effect the moment you press the key
instead of after a trip to the server and back. They are still sent, so
everybody else sees them, and the echo is dropped when it returns
- the client no longer replays the server's frame times out of a jitter buffer.
It runs on its own clock, like the host does, and simply carries out the
commands as they arrive. Playing back somebody else's frame times meant the
world ran at the wrong speed whenever the two machines drew at different
rates, and every buffered frame sat between the player and their controls
Also: the server only sends a frame when it has commands in it, or every tenth
one to carry the tick, instead of one per rendered frame.
Second live test. Four things, three of them the same mistake seen from
different angles: the session was reasoning about single vehicles where it
should have been reasoning about trains.
- a peer may now operate anything coupled into the set it is sitting in. This
is what "cabchangebackward rejected: not a member of the crew of that
vehicle" was: the crew was recorded against one car of a four-car unit, and
the cab switch addressed another
- every vehicle gets a network id, not only the ones worth driving. Commands
are addressed by that id, and a player has to be able to walk into any car
of their train. "peer 3 cannot take en71-017rb_4: no such vehicle in this
session" was the driving trailer of an EN71 - a full desk, no engine, and
nobody assigned to it, so the old test missed it. Having a driving position
is what counts now, engine or not
- the roster groups vehicles into the trains they are coupled into and the
lobby lists one row per train. Coupling is worked out afresh every time, so
a train that comes apart becomes two, and because a crew entry sits on the
car its player occupies, each half keeps the people who were riding in it
- walking to another car of your own train is not a new claim and asks nothing
of the crew capacity, which counts people per train
Positional correction never actually did anything: it identified a track by
name, and most track in a scenery has none, so the lookup returned nothing and
place_on_track() was never reached. The client sat 178 m out of place for
minutes, reporting four vehicles "put back on the rails" every time while
moving none of them. Tracks now travel as their index in the path table, which
comes from the scenery file and is therefore the same on every peer.
Also from the test:
- the camera jumped in and out of the cab on the client whenever the host
reversed, because the correction was forcing the server's CabActive and
CabOccupied onto it. Which cab somebody sits in is theirs alone - two people
sharing a vehicle each have their own - so it is no longer replicated
- the correction no longer touches the control positions of the train the local
player is working. Their crew's commands reach every peer on their own, and
snapping a lever back a fraction of a second after it was moved is what made
the reverser look dead
- the frame buffer was sized for the old lockstep, up to sixty frames deep.
Every frame in it is input lag, so it is down to twelve with a target of one
- asking the AI of a vehicle to step aside is given five tries and then said
once, instead of every thirty frames forever; and a resync request now has a
real cooldown
First live test showed three things wrong at once.
A client could not take a vehicle. Walking up to a locomotive and pressing the
enter-vehicle key posts entervehicle, which stage 4 classed as a world-wide
command and refused outright ("reserved for the session host"), so the client
never got a cab and every later control command died locally for want of one.
Walking into a cab is a request for a seat, not a world command: the authority
now turns it into a crew claim, for the host through the same path as for a
remote peer, and answers it the way the lobby button is answered.
Nothing but commands was ever replicated. The server sent dt and the commands
it ran, and the client simulated the rest for itself - so a pantograph the host
raised before the client joined, the contents of the memory cells the signalling
reads, brake pressures and switch positions simply never crossed the wire. The
server now streams the authoritative state a few times a second: whatever has
changed since the last one, and the whole of it every so often so a peer heals
from anything it missed. The snapshot carries the vehicle's inputs - line
breaker, battery, converter and compressor permits, pantograph valves, cab,
lights, doors - and lets the receiving peer derive the rest, plus the memory
cells and the switches.
Desync detection demanded that two state digests be equal, on a design where
the client deliberately runs its own physics. They never can be, so it reported
a desync permanently and asked for a resync every five seconds forever, which
fixed nothing because the next step disagreed again. The digest is a log line
now; what is measured instead is how far out of place the routine correction
found us, and a full resync is only asked for when even that is not catching up.
Also: a consist is put back on the rails as a whole or not at all. Correcting
one vehicle while its neighbour stayed put is what was stretching couplers
until they tore - the last test moved six vehicles out of eight, every time.
And the lobby window collapsed to nothing, because its size_max was left at the
-1 default while size_min was set, so imgui was handed a maximum smaller than
the minimum.
Joining meant replaying backbuffer.bin from the very first frame of the
session, so the longer a session ran, the longer it took to get in - and the
whole history had to be kept for it.
- network/snapshot: a chunked container with a version of its own. Every
section carries its length, so a peer built before a section existed reads
a newer snapshot and simply walks past what it does not know. Version 1
carries the session (tick, clock, weather, the state of the random engine),
the vehicles (position on the track, speed, direction, controls, brake
pressures, whether the AI is driving) and the crew roster. Switches,
memcells, signals and the event queue are the next sections
- joining is now: handshake -> load the scenario -> CLIENT_READY -> the server
takes a snapshot on the spot and sends it -> the peer goes live. Because the
snapshot is taken at the moment the peer becomes active there is no backlog
to replay at all; the frame by frame catch-up is gone
- the server refuses a peer that loaded a different scenario, with a readable
reason instead of a silent desync
- reconnecting into a session that is already loaded asks for a fresh snapshot
straight away rather than waiting for a loader that will not run again
- a vehicle is only put back on the rails when it really is somewhere else; the
arithmetic that turns the recorded bogie offset back into the nose distance
place_on_track() expects mirrors what that function itself does
- the session log is still written out - it is useful when picking a desync
apart - but nothing depends on replaying it any more
