A launch day is not a scaled-up test night. It is a fault-tolerant burst event: thousands of clients install a 60–120 GB build, decrypt it, stream a first-time asset set, and connect within the same few hours, while your matchmaker, login service, and telemetry API absorb a request rate an order of magnitude above steady state. Bandwidth planning for game launches therefore has to be done in three separate currencies — egress throughput (Gbps), packet rate (Mpps), and cold-cache peak multiplies — and then mapped onto hardware and contract terms that do not fall over when the first two minutes are ugly.
This guide covers the arithmetic for a 10,000-concurrent-player launch, the failure modes that show up in the first ten minutes, and the commitments that keep a dedicated server platform inside budget.
Step 1: Separate the Three Traffic Streams
Launch traffic is never one number. It is three streams, each stressing a different part of the stack:
- Gameplay (UDP). Continuous, latency-critical, small packets. At 20–40 Hz snapshots, a player consumes roughly 15–40 Kbps downstream and 3–8 Kbps upstream; the server fans every player’s state out to every visible player, so server-side egress grows faster than linearly with density.
- Content delivery (TCP/HTTPS). Bursty and enormous. A 90 GB build delivered to 10,000 players in the first day is 900 TB of egress; if 60% of it lands in a three-hour window, that window alone carries about 540 TB, or roughly 50 GB/s if spread evenly — which is why it must not come from your game nodes.
- Control plane (HTTPS/JSON). Login, matchmaking, inventory, and telemetry. Low bytes, high packet count and high connection rate. This is the stream that eats CPU on the API tier, not bandwidth.
Step 2: The Per-Player Bandwidth Table
Use measured per-slot figures from your own load tests, but here is a defensible planning baseline for a session-based multiplayer title at 20 Hz:
| Stream | Per player | Server egress | Server ingress |
|---|---|---|---|
| Gameplay, 20 Hz, moderate density | 25 Kbps up / 5 Kbps down | 25 Kbps | 5 Kbps |
| Gameplay, 30 Hz, high density (24+ visible) | 60 Kbps up / 12 Kbps down | 60 Kbps | 12 Kbps |
| Voice, Opus 24 kbps | 24 Kbps up/down | 24 Kbps each direction | 24 Kbps |
| Login + telemetry burst at join | 400 KB one-off | 400 KB | 2 KB |
| Patch / asset download | 90 GB one-off | 90 GB (CDN) | — |
At 10,000 concurrent players on 30 Hz servers with voice off, peak gameplay egress is 10,000 × 60 Kbps ≈ 600 Mbps. Voice at 100% adoption doubles that. The naive conclusion — “1 Gbps is plenty” — is wrong, because it measures average bytes and ignores packet rate and the join stampede.
Step 3: Packet Rate Is the Real Ceiling
A 300-byte snapshot at 30 Hz means 30 packets per second per player — 300,000 pps at 10,000 players, well within a 10GbE port but not within every virtualised host:
| Metric | Single 128-slot instance | 10,000 CCU fleet |
|---|---|---|
| Outbound packets/s | ~4,000 pps | ~300,000 pps |
| Outbound bits/s | ~8 Mbps | ~600 Mbps |
| Conntrack entries | 128 UDP | 10,000–30,000 NAT flows |
| Peak join rate target | 8 joins/s | 600 joins/s |
Where the fleet meets the internet, a single IRQ-bound queue saturates around 1–1.5 Mpps on a modern core. Enable RSS multiqueue, XPS, and RPS on the aggregation tier, and verify with cat /proc/net/softnet_stat that no CPU’s drop column climbs. The kernel-side sequence is in game server network optimisation for UDP and NIC tuning.
Step 4: Model the Join Stampede, Not the Average
Capacity plans die at T+90 seconds, when a streamer’s audience and the press embargo lift at once:
- Login burst: 10,000 players hitting auth in 10 minutes = 17 auth/s, each with 2–4 dependent calls. Keep auth and matchmaking on separate cores from gameplay and give them connection pooling to a warm database.
- Instance spin-up: pre-warm 120% of expected slots. Cold-starting a survival-world instance can take 45–180 seconds of disk reads; a player who gives up at 30 seconds is a lost player, not a queued one.
- Content burst: push the day-one patch to the CDN 48 hours ahead and pre-seed regional PoPs. Origin egress should spike to single-digit Gbps, not hundreds.
- Retry amplification: every failed join retries 3–5 times. A 5% failure rate becomes a 25% request-rate increase against the same tier.
Step 5: Hardware and Port Commitments That Match the Model
For 10,000 CCU, the typical production shape is not one enormous machine — it is a small fleet behind a matchmaker:
| Tier | Count | Spec | Peak load |
|---|---|---|---|
| Game nodes | 8 | 16c/32t, 3.5 GHz+, 64–128 GB, 10GbE | ~6,000 pps each |
| Auth / matchmaking | 2 + 1 standby | 8c/16t, 32 GB, NVMe | 200 req/s each |
| Aggregation / LB | 2 (HA) | 8c, 16 GB, 2×10GbE bonded | <0.6 Gbps, <350k pps |
| Origin for CDN | 1 | 8c, 32 GB, NVMe RAID1 | 2–6 Gbps burst |
On the contract side, pick a provider that quotes unmetered or 95th-percentile burst on a 10GbE port and can document scrubbing capacity above your expected game traffic. Fixed 1 Gbps commitments look cheap until the CDN cache-miss day triples egress and you are billed per TB.
Step 6: The Pre-Launch Rehearsal Checklist
Bandwidth plans are only as good as the last test. Before launch, prove these six things on the real hardware, not a cloud approximation:
- Load-test at 140% of expected peak CCU with synthetic players and verify tick time stays under 45 ms p95.
- Run a simultaneous 10,000-connection auth storm and confirm your error budget stays under 1%.
- Burn 30 minutes of line-rate traffic on the aggregation port and watch for softirq drops.
- Simulate a game-node loss mid-match and confirm the matchmaker reschedules inside 60 seconds.
- Verify carrier BGP sessions and test failover to your secondary transit path.
- Re-check that monitoring at fleet scale does not itself consume a node’s headroom — sidecar agents cost measurable CPU under burst.
Common Mistakes That Sink Launch Days
Three patterns account for most launch incidents: serving patches from game servers, sizing for average CCU without a ramp plan, and running CDN origin and gameplay over one unprotected uplink.
Summary Budget for 10,000 Concurrent Players
Realistic launch envelope: 0.6–1.5 Gbps sustained gameplay egress, 300–500k pps, 2–6 Gbps CDN origin bursts, and roughly 2 TB/day of log and telemetry ingest. Provision 2× the expected peak on the spine, 120% of game slots pre-warmed, and a bandwidth contract with burst headroom rather than a fixed cap.


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