NetworkPerformanceSystem

For groups playing together across long distances. If your friends are 200ms+ away and the game feels like mush no matter how much bandwidth everyone has, this is aimed at you.

Why this is different from the other networking mods

Every other Valheim networking mod treats multiplayer performance as a bandwidth problem — compression, bigger buffers, higher send rates, filtered broadcasts. That is the right diagnosis for ten people in one base saturating a shared cap.

It does not explain why two friends on opposite sides of the world feel awful at 5% link utilisation. Two things in vanilla punish distance specifically, and neither is about bandwidth:

1. The send window is a fixed 10 KB of in-flight data. Throughput through a fixed window is window ÷ round-trip-time, so vanilla's window is correctly sized up to about 67ms and starves everything past it — roughly 41 KB/s at 250ms, no matter how fast your connection is. And it fails hard rather than gracefully: over the limit, that player receives nothing that tick. That is the freeze half of freeze-then-teleport.

2. Simulation authority is handed out by arrival order. Valheim gives each object to one peer to simulate, and everyone else sees it relayed through the host. Vanilla picks whoever happens to be first in an internal list — uncorrelated with who is fighting the creature and uncorrelated with latency. For a spread-out group that is close to the worst available choice, and it is why a mob can feel laggy to everyone at once.

This mod sizes the window per peer from measured round-trip time, and places authority to minimise how stale things look to the people actually watching. Notably, sizing by RTT is a no-op for players who were already fine — simply raising the constant for everybody, which is what the other mods do, hands a 20ms player tens of kilobytes of standing queue, which is latency added to someone who did not have a problem.

What it does

Installing

Server-only works. Vanilla clients get the send window, ownership, scheduler and station-request fixes with nothing installed on their end.

Installing on clients too adds latency compensation, live position reporting and the clean exit from a dead session for those clients, and lets a ship they own pass to whoever takes its helm. Mixed groups are fine — benefits are per-player, and a client without the mod behaves exactly as vanilla. There is no version lock: nobody gets kicked for not having it.

Works on dedicated servers and on player-hosted games.

Known limitations

• On a dedicated server, contested objects go to the lowest-latency player present rather than to the server itself - except in the zones the server has actually loaded, which are the ones around the world origin. There, whenever two or more players are present, the server wins contested objects (it is zero hops from everyone, so that is the lowest possible staleness) and simulates them. That is intended and is a CPU cost to plan for on a busy spawn hub; set Allow Host As Owner to false to place purely on players. The server never takes ownership of zones it has not loaded, because owning something it is not simulating would freeze it.
• Ownership arbitration re-places two kinds of object, by two different rules: creatures and other simulated, moving things go to whoever minimises staleness for everyone watching, and interactables that do not move — pickables, ore deposits, rocks, trees, logs, destructibles — go to whoever is nearest. Everything else — building pieces, containers, fermenters, smelters, cooking stations, portals — follows vanilla exactly: owned by whoever arrived first, re-owned only when that player leaves. That exclusion is deliberate. A station's owner is where the game sends item-insert RPCs, and the game removes the item from your inventory before sending, so moving the owner while a player is mid-insert loses the item outright. An interactable spends nothing before its RPC and keeps all its state in the object, so the worst a mistimed move can cost is one keypress.
• Moving an interactable opens a brief window — one send tick — in which two players' copies of the world disagree about who owns it. If both happen to hit the same bush or rock inside that window, it can be picked or damaged twice. Two things make that very unlikely: an object is never taken from an owner standing within reach of it, so for anything picked or hit in melee only one of the two can be interacting at all; and every move is pushed to the players nearby at the front of the next send. Ranged damage into a destructible from well outside that reach is the case not covered. Set Interactive Object Ownership to false to keep vanilla's rule for these objects.
• Requests for these objects are not re-addressed the way station requests are (see above), so a pick or a swing sent to an owner that has just changed still goes nowhere and needs a second press. That matches vanilla's behaviour whenever an owner walks away; extending the re-addressing to cover them is a later change.
• A ship is handed to its helmsman by the machine that currently owns it, so that machine needs the mod: a listen host, a dedicated server, or a client with it installed. A ship owned by a player without the mod stays with them, as in vanilla. The host never takes a moving ship from its owner itself, because the owner's updates already on their way would drag ownership back. It does give an abandoned ship to the player at its helm. The throttle press that takes the helm can be lost if the ship's former owner receives it just after the handoff; press again.
• Latency compensation is dead reckoning: an entity that stops abruptly will overshoot slightly and settle back. Lower LatencyCompensationStrength if you find it distracting; 0 disables it.
• Requires the Steam backend. Crossplay/PlayFab connections do not report round-trip time, and without a measurement every mechanism falls back to vanilla behaviour rather than guessing.
• If the log shows RttSampling disabled, the server's Steam interface could not be queried and everything runs as vanilla; nps_stats shows the reason.
• Honest send windows mean a full server can actually use its bandwidth: 50 players at the default 150 KB/s target is ~60 Mbit/s of upload worst case. Backpressure degrades gracefully if the link is smaller, but provision the server's uplink for the player count rather than assuming vanilla's artificially starved usage.
• On large servers the two knobs that matter are Frame Budget Ms (how much of each server frame the send path may use - under load the per-peer send rate degrades gracefully instead of the frame time growing without bound; nps_stats shows the effective rate) and Max Reassigns Per Pass, which auto-scales with player count so ownership converges at the same per-player rate on a full server as in a small group. The Load Penalty Ms setting spreads contested objects across low-latency peers instead of piling everything on the single lowest-ping player - leave it on unless you specifically want pure staleness placement.
• Max Players raises vanilla's cap of 10, but a crossplay server cannot go past 128 whatever it is set to - PlayFab's lobbies do not hold more, and crossplay clients join that lobby before the server ever sees them. Steam-only servers have no such ceiling. Crossplay (PlayFab) peers also never report a round-trip time, so they are priced at Unmeasured Peer RTT Ms for ownership and rendered at vanilla by other clients.
• Connection Timeout Seconds is for players who get dropped mid-join or during a hitch on a weak link. It does not make a slow connection faster — it stops both ends declaring it dead while it is still working. Its old cost, that a player who is genuinely gone kept ownership of everything they were simulating for the whole window, is what Evict Ghost Owners now removes: the slot is still held, the objects are not. What remains is the slot itself, so on a full server a long timeout still means a departed player's place takes that long to free up. The one timeout a server cannot set for a client is Connect Timeout Seconds, which covers the handshake before anything has been synced — whoever cannot get connected has to raise that one themselves.
• Ghost detection is at its sharpest on Steam sockets, where the transport is asked directly and answers immediately. Crossplay (PlayFab) peers report nothing, so they are judged on silence alone — correct, just up to Ghost Owner Evict Seconds slower. If every peer goes quiet at once, that is read as a fault at the host's end rather than as everybody leaving, and nobody is evicted until one of them answers.
• Raising Max Players is not free and nothing here makes it free: each added player costs the host upload and CPU against every other player. Frame Budget Ms and the Steam transport ceiling below are what decide whether a bigger number is actually playable - read the two bullets above before setting one.
• Some mods wait for a peer's socket send queue to fall under a fixed byte count before they send (Jotunn's CustomRPC, ServerSync and every mod bundling it, ConditionalConfigSync) and disconnect the peer after 30 seconds if it never does. That number was sized against vanilla; a latency-sized window legitimately keeps more in flight past ~100 ms RTT. Jotunn's limit is a field, and it is raised automatically to Max Window Bytes plus 20000 on both ends. The others have theirs compiled in, so the Compatibility settings briefly bring a backlogged peer's queue under it every Queue Drain Interval Seconds - a few percent of throughput, only while that peer is backlogged. If a mod with its own threshold still times out, lower Queue Drain Floor Bytes or the interval; nps_stats shows both limits and how often the drain ran.
• The Allocation settings are churn reductions, not a fix for running out of memory, and it is worth being precise about the difference. Mono's collector runs more often the faster objects are created, so removing allocations from the ZDO path means fewer collections and less CPU spent on each one. It does not reduce how much is live at any moment - on a very large world that is millions of field tables, one or two per ZDO, and it is a function of how much world has been generated. No mod changes that, and it is what the ceiling is actually set by. If your server is hitting a memory wall, these lengthen the interval between incidents and make the server cheaper to run in between; they do not remove the wall. Reducing generated world area is what does that.
• All five Allocation settings ship off. They are byte-identical to vanilla by construction - same values, same wire format, same order - but they are IL-level changes to the hottest paths in the game, so they are opt-in for their first release: enable them one at a time and soak each. Three of them need a restart to turn on, because the mod refuses to install a hook on a method called for every replicated object in every packet unless someone has asked for it; turning any of them off applies immediately. nps_stats says which are running, and tells you plainly when one is switched on in the config but was not installed this session.
• Limit Relay By Distance ships off. It judges "too far to have the object loaded" from each player's negotiated simulation distance plus one zone of slack, so it is lossless for players whose game loads what the server agreed to. A client-side mod that loads more of the world than that could miss building damage or effects at the very edge of its view. Leave it off and read the "out of range" line in nps_stats first: if that number is small, it is not worth turning on.

Incompatible with

Other networking mods that rewrite the same code: FiresGhettoNetworking, VBNetTweaks, SkadiNet, BetterNetworking, Smoothbrain's Network, NetworkTweaks, WarheimNetwork, TimeoutLimit. BepInEx will refuse to load this mod alongside them rather than leave you half-patched. TimeoutLimit in particular is no longer something you give up: the Connection Timeout settings cover the same ground, and unlike a timeout raised at one layer only, they move both of the places vanilla times out at.

Verified compatible (different layers, no overlap): LeanNet, Compress, Scenic. ReturnToSender is fine too — it already fixes the scheduler, so that one mechanism stands down and the rest keeps working. EnRoute and BetterZeeRouter are fine in the same way: both rework the routed-RPC relay, so relay filtering and relay send reuse stand down when either is present and everything else keeps working. Valheim Plus is fine too: it has its own player limit, so this mod's Player Limit settings are ignored when it is installed and V+'s maxPlayers is the one in force. ClientGhostWatchdog is fine as well: it does the client-side half of the ghost handling itself and against its own timeout, so this mod's client watchdog stands down and leaves that mod's behaviour in charge. The host-side half — evicting a ghost from ownership, which that mod does not do — keeps running either way.

Credits

The ghost handling started from ClientGhostWatchdog by DreamWraith, which spotted that a disconnected client keeps playing into a world that is no longer there, and settled on warning at the halfway mark and leaving cleanly at the deadline. No code is shared, and the two halves here differ in what they ask — the transport's own verdict rather than the game's ping timer, and a host-side question that a client-only mod could not reach — but the observation that the game needs a second, independent opinion on whether a peer is still there is theirs.

Changelog

See CHANGELOG.md.

Changelog

1.7.0

• Players on a high-ping connection no longer see the world pause every 8 seconds while there is a lot going on.
  • Queue Drain Interval Seconds and Queue Drain Floor Bytes are no longer read and can be deleted from the config.
  • nps_stats "Third-party send queue thresholds" shows how many outside reads were adjusted and how much is kept out of them for each player.
• A creature with exactly one player near it now belongs to that player, whatever anyone else's latency is. New Ownership setting Creature Proximity Ownership (default on), with Creature Proximity Radius (default 48 m).
  • nps_stats gains a "proximity" line: kept, pulled, and rescues it redirected.
• Network monitoring: a recorder a server owner can switch on when something is wrong, to send with a bug report. New Monitoring section, Enable Network Monitoring (default off). See "Reporting a network problem" in the README.
  • Records each change of a creature's owner and whether it held, messages delivered to a player who no longer owned the target, how regularly each player's creatures report in, and each player's ping and connection quality. Clients with the mod add what only they can see, at up to 2 KB per second in small batches that always wait for the game's own traffic, and a player can refuse with AllowMonitoringUpload.
  • No player names, platform ids, addresses or chat. Players appear only as the session number the game gives them.
  • Costs nothing while off: its hooks are installed when it is switched on and removed when it is switched off, with no restart either way.
  • Written to BepInEx/NpsMonitoring, compressed, and capped by Max Disk MB (default 4096). Recording stops at the cap; nothing already recorded is deleted.
  • nps_stats gains a "Network monitoring" block.

1.6.0

• A player who has stopped answering no longer freezes everything they were simulating. New Connection Timeout setting Evict Ghost Owners (default on), with Ghost Owner Evict Seconds (default 10).
  • "Stop trusting this peer to simulate" is a different question and is now asked separately. A peer that goes quiet loses its objects to the players who are actually there, while keeping its slot for the full timeout so it can come back. Nothing is disconnected any sooner, and if it returns it competes for ownership again on the next pass.
  • Detection asks Steam directly - SteamNetConnectionRealTimeStatus_t.m_eState, which reports a dead link before anything closes the socket - and falls back to a silence timer for crossplay peers, which report no state. The timer discounts main-thread stalls rather than counting a frozen frame as the peers going quiet.
  • If every peer goes quiet at once, that is treated as a fault at the host's end and nobody is evicted until one answers.
  • This is what makes raising Connection Timeout Seconds safe: the wait now costs only the slot.
• Clients are no longer left playing a world the server has dropped them from. New Client config setting EnableGhostWatchdog (default on).
  • A warning with a countdown appears halfway to the timeout actually in force and clears itself if the connection comes back; at the deadline, or as soon as the transport calls the link dead, the character is saved and the game returns to the menu with an explanation rather than the generic "disconnected".
  • No timeout of its own - it follows the same deadline the game is using, so it cannot disagree with the server's setting.
  • Stands down automatically when ClientGhostWatchdog is installed.
• nps_stats gains a "Peer liveness" block: who is quiet, for how long, what the transport says, and the eviction and watchdog counters.
• Credit: the ghost handling started from ClientGhostWatchdog by DreamWraith. No code is shared; the observation that the game needs a second opinion on whether a peer is still there is theirs.
• Picking berries, mining ore and chopping trees no longer lag when another player owns the object. Ownership arbitration is now tiered: moving objects are placed by latency as before, and interactable-but-stationary ones by distance. New Ownership setting Interactive Object Ownership (default on).
  • Pickables, ore deposits, rocks, trees, logs and destructibles go to whoever is standing nearest them. These ask their owner to do the work and every handler checks it is the owner, so an object owned by another player costs four network legs per keypress - and placing by ping rather than by distance would not have helped, because nothing about a berry bush changes between interactions.
  • Building pieces, containers, crafting stations and portals are still never moved off a player who is present, which is what keeps the fermenter and smelter fix above intact.
  • An object is never taken from an owner still within reach of it, never taken for a player further away than Interactive Claim Radius, and each move is pushed to nearby players immediately.
  • An unowned patch - one nobody has visited - now goes to the nearest player rather than the lowest-ping one, on the first pass and with no delay.
  • Tuning: Interactive Claim Radius, Interactive Challenge Margin, Interactive Min Hold Seconds, Interactive Max Reassigns Per Pass. The tier has its own budget so a zone full of creatures cannot starve it, or be starved by it.
  • nps_stats gains per-tier ownership counters.
• Ships are now simulated by the player at the helm, so steering is no longer choppy when someone else on board owned the ship. New Ownership setting Ship Ownership Follows Helmsman (default on).
  • The ship's current owner hands it over when the helm is taken, the same way the game already hands over saddles and carts. This needs the mod on whichever machine owns the ship; a ship owned by a player without it behaves as in vanilla.
  • The host gives an abandoned ship (its owner left or disconnected) to the player at its helm first.
  • nps_stats gains a "Ship helm" block with the handoff and helm rescue counts.

1.5.0

• Updates default send timing to prevent alternating delays from the dedicated server
• New Allocation section: four opt-in changes that remove short-lived objects from the ZDO network path, reducing how often the garbage collector runs. None of them alters a byte on the wire or a value in a ZDO.
  • ZDO Deserialize Fast Path (default on) - reads a received ZDO's fields directly instead of through the fourteen delegates the game allocates for every one.
  • Packet Read Fast Path (default on) - reads each incoming ZDO payload straight into the buffer that is about to hold it.
  • Send Package Reuse (default on) - reuses the two packet buffers the send path builds, instead of constructing and discarding both on every send to every peer.
  • RPC Invoke Fast Path (default on) - calls an incoming RPC's handler directly when its signature is the common one, instead of going through reflection for every message.
  • Relay Send Reuse (default on): the host writes each relayed RPC once and hands the same bytes to every player it goes to, instead of rebuilding and re-copying the whole message per player.
  • nps_stats gains an "Allocation relief" block with a counter per mechanism and the game's own per-second ZDO send/receive rates to read them against.
• New Routed RPC setting Limit Relay By Distance (default on): stops relaying object RPCs such as building damage, fragments and ward flashes to players who visited that object earlier but are now too far away to have it loaded.

1.4.2

• Fixes Jotunn-based mods (Jotunn's own config sync included) disconnecting a peer with a 30 second "sending timeout" on higher-latency links (#3)
• New Compatibility section for mods with that threshold compiled in (ServerSync and mods bundling it, ConditionalConfigSync): a peer in sustained backlog has its queue briefly brought under the threshold every Queue Drain Interval Seconds so they can get their packet out.
  • nps_stats gains a "Third-party send queue thresholds" block showing Jotunn's limit as read back from Jotunn and the drain counters per peer.

1.4.1

• Fixes the player limit not applying to crossplay on dedicated servers.
• The Steam server browser now shows the configured limit for dedicated servers.

1.4.0

• Fixes items vanishing when inserted into a fermenter, smelter, cooking station, fireplace or similar in multiplayer. This is a vanilla bug made worse with more players.
  • Station item requests are now delivered to the object's current owner. This is best paired with the Valheim Community Patch, which addresses the other half of this.
• nps_stats ownership block gains a static held line: objects left with a present but not lowest-latency owner because they do not move.
• Valheim Plus compatibility: when V+ is installed its player limit is the one in force and this mod's Player Limit settings stand down instead of fighting over the same patch sites.

1.3.0

• Removes admin requirement from server commands
  • nps_stats_collect requires devcommands
• Adds support for Playfab servers to increase connection limits beyond vanilla 10

1.2.1

• Bumps the maximum connection timeout limit to 600s

1.2.0

Configurable connection timeouts. Vanilla drops a quiet connection after 30 seconds, which is not enough for a slow or distant link mid-join, and it enforces that at two layers independently - so raising only one of them changes nothing. The new Connection Timeout section writes both, plus the handshake timeout Steam applies before a connection exists and the longer allowance the game already gives itself during a crossplay world transfer. Everything ships at its vanilla value, so enabling the section on its own changes nothing; nps_stats shows what is actually in force.

1.1.0

Initial release!