“Rubber-banding” is the most over-diagnosed fault in game server operations. Players report being pulled back across the map, and the reflex is to add bandwidth, move the data centre, or blame the ISP. Three very different subsystems produce identical-sounding complaints: client prediction correcting from late snapshots, the server skipping ticks under CPU or disk pressure, and path jitter no amount of hardware will fix.
A server fault trace tells you where a packet was lost, not which component caused the loss. The sequence below attaches each symptom to a measurement you can run in minutes on a live node whose specs you already trust. If the node itself turns out to be undersized rather than faulty, start from the dedicated server hosting options before spending money on network fixes.
Step 1: Classify the Symptom
Three distinct phenomena get called rubber-banding. They have different root causes and different owners — network, server CPU, or disk — and the first job is to decide which one your players are describing:
| Symptom | What players see | Most likely origin |
|---|---|---|
| Positional snap-back | Player moves, then is teleported back 1–3 m | Client prediction correcting from a delayed snapshot |
| Uniform delay | Everything happens ~200 ms late, consistently | Path latency or a low tick rate; not a fault |
| Stall then catch-up | Freeze for 0.5–3 s, then all movement resolves at once | Server tick hitches — CPU stall or blocking disk I/O |
| Spike-only | Fine until a fight or a build completes | Burst cost: entity spawn, save, or garbage collection |
If players report snap-back after every movement, suspect client prediction lag rather than the server. If they report freezes that resolve at the same moment for everyone on the map, the server missed ticks — skip to Step 3 and ignore the network checks.
Step 2: Separate Netcode From Network
Client prediction hides up to (RTT / 2 + interpolation delay) of latency. At 60 ms RTT and 100 ms interpolation, corrections of up to about 2 m while sprinting are normal and invisible. Persistent snap-back means one of three things: snapshots arrive out of order, the server sends state at a lower rate than the client predicts at, or packets are dropped on the path.
- Verify the tick rate actually being served. Many titles reduce tick adaptively under load, and a drop from 30 to 20 Hz shows up as repeated small corrections rather than one big jump.
- Measure server-side send rate per client with
ss -u -i: a client receiving fewer snapshot bytes than its session average is being throttled by socket buffer pressure. - Compare server send timestamp to client arrival timestamp on a sample of 200 packets. One-way jitter above 15 ms with zero loss still produces corrections.
For the kernel-side values that remove the throttling component — socket buffer sizes, multiqueue, and interrupt affinity — see the guide to UDP, kernel, and NIC tuning for lower game latency.
Step 3: Catch the Tick Hitch
If the freeze resolves simultaneously for every player, the server stopped simulating. Capture the tick-time distribution and read it against this table:
| Tick p99 | Diagnosis | Next check |
|---|---|---|
| < 25 ms | Server fine; look at path or client | Jitter and loss per hop |
| 25–40 ms | Headroom shrinking; population-driven | Per-core CPU utilisation, steal time |
| 40–80 ms | Real hitching (0.5–3 s freezes) | Disk iowait, GC pauses, save writes |
| > 80 ms or exactly periodic | Blocking operation on the game thread | Autosave, world save, log flush, backup |
Line up the spike timestamps against your save schedule. A 400 ms p99 tick every 15 minutes that aligns exactly with autosave is disk I/O on the game thread — the single most common cause of “random” rubber-banding on hardware that is otherwise perfectly healthy.
Step 4: Rule Out Disk I/O
Run iostat -x 5 during a hitch. The signature of a disk-induced hitch is %iowait above 10% while the game process sits in uninterruptible sleep (D state). A SATA SSD save write with 200–500 µs p99 latency becomes a 20–40 ms tick stall once a 200 MB world save queues behind it; NVMe p99 under 200 µs keeps the stall invisible.
- Put world saves on a separate NVMe device from the game binary and logs.
- Reduce autosave frequency and write via atomic rename so the game thread never waits on fsync.
- Check for an active RAID rebuild or a failed controller cache battery — both cause periodic multi-second stalls.
- Confirm the filesystem is below 80% full; fragmented allocation raises write latency sharply at high occupancy.
Step 5: Check Steal Time and CPU Placement
On a virtualised node, sustained steal time above 2% produces exactly the stall-then-catch-up pattern, because the hypervisor deschedules the game thread while clients keep predicting. Sample it per vCPU under load. A dedicated machine removes this class of fault completely, which is why tick-sensitive titles usually end up on bare metal rather than shared compute. Also confirm that the game thread and its network thread sit on the same CCD; a cross-socket hop adds latency to every snapshot fan-out on the server side.
Step 6: The Client Side of the Equation
Not every rubber-band is yours. Client frame time, packet loss on the player’s Wi-Fi, and anti-cheat agents that hook network calls all produce corrections. Ask for an MTR from the player to the server and, if the game exposes it, their jitter stat. If 30% of a player base reports rubber-banding and they share an ISP or city, the problem is a transit path, not your node — the fix is a second peering or a different regional PoP.
A 10-Minute Diagnostic Runbook
- T+0: record tick p50/p95/p99 alongside live player count.
- T+1:
mpstat -P ALL 1 60— look for one core above 90% (single-thread ceiling) or elevated %steal. - T+3:
iostat -x 5— check %iowait and await on the save device. - T+5:
ss -u -i— send/recv queues and drops per client. - T+7: correlate spikes with the save schedule, backup window, and log rotation.
- T+10: rerun the same window with autosave disabled to confirm or rule out disk I/O conclusively.
Most rubber-banding on a well-provisioned server comes down to one of three things: an autosave colliding with a tick, a single-thread CPU ceiling reached before the player-count limit, or steal time on a shared host. All three are visible within ten minutes using the commands above, and all three are fixed by configuration or hardware placement rather than by adding bandwidth.

Leave a Reply
You must be logged in to post a comment.