Dedicated Gaming Server Benchmark Thresholds: CPU, NVMe and Network

A benchmark is only useful if it has a pass threshold and a command that produces the number. “Feels fast” and “runs well” are not acceptance criteria. This page lists the exact thresholds a gaming node must clear, the command to measure each one, and what a failure predicts for tick rate at 50 and 100 players.

The Four Thresholds That Predict Tick Rate

Measure four things before players arrive: CPU single-thread performance, NVMe random-write IOPS and p99 latency, network round-trip and jitter, and sustained disk throughput during a simulated save. Each has a hard pass line. Any node that passes all four will hold its target tick rate unless mods or entity counts are pathological.

MetricPASSFAILCommand
Single-thread scoregeomean 1,800 (R23 equiv.)<1,000sysbench cpu –threads=1 run
1T event latencyp99 < 12 ms>25 mssysbench cpu –threads=1 –time=60 run
NVMe 4K randwriteIOPS >150k, p99 <500 us<40k or p99 >2 msfio randwrite bs=4k
Network RTT / jitter<30 ms / <3 ms>60 ms / >10 msmtr –report -c 100
Save-window stallp99 <30 ms>120 msfio mixed + save loop

CPU Single-Thread: Commands and Pass Line

Run sysbench cpu --threads=1 --time=60 run three times and take the median events per second. On current Genoa or Sapphire Rapids cores expect 3,800-4,500 events/s; a v3 Xeon lands near 1,900-2,300. The important number is the p99 event latency from the same run: under 12 ms is PASS, over 25 ms means the hypervisor is stealing cycles or the core is throttling.

Check steal time during the run with vmstat 1 10 and cat /proc/stat | grep cpu. More than 2% steal on a “dedicated” node is a fail even if the throughput number looks acceptable, because steal arrives in bursts that coincide exactly with save flushes. Cross-check with a fixed-workload single-thread compile and compare node to node.

For Minecraft specifically, the number that matters is milliseconds per tick with a reference chunk load. Time /forge tps or spark against a 5,000-chunk pre-generated area and require mean tick under 22 ms at 40 players. If the host cannot supply a pre-generated world, time world generation separately — it is disk and CPU bound and hides true tick cost.

NVMe: IOPS, Queue Depth and the p99 That Matters

The acceptance command: fio --name=rw --filename=/data/test --size=4G --rw=randwrite --bs=4k --iodepth=32 --numjobs=4 --runtime=120 --time_based --percentile_list=50:99:99.9. Read the iodepth-1 result separately, because game saves are largely synchronous: --iodepth=1 --numjobs=8. PASS is over 150,000 IOPS at depth 32 with p99 under 500 µs.

An SATA SSD in the same test lands near 60,000 IOPS and 250 µs at depth 32 — which looks survivable until you add a second instance on the same drive. Shared image files on a NAS backend will show p99 in the 3-10 ms range and fail outright. Verify device type first with lsblk -d -o NAME,ROTA,MODEL; ROTA=1 means spinning rust regardless of what the invoice says.

Network: RTT, Jitter and Loss Under Load

Run mtr --report --report-cycles=200 -rw 1.1.1.1 from the node and from three player locations. PASS is under 30 ms median to the player population, under 3 ms jitter, and zero loss on the first hop. A first hop over 2 ms usually means the node sits behind an oversubscribed ToR or a hypervisor vswitch that batches packets.

Ping is not enough: run a 60-second flood on UDP with hping3 -2 -i u2000 --fast target against a second node and watch for reordering. Better still, install a tick-timing mod and log round-trip plus jitter per player. Player-side stutter with a clean 20 ms ping is jitter, not latency, and no amount of extra bandwidth fixes it. The network path is usually the difference between a 8 Gbps and a 1 Gbps quote in practice.

Baseline Reference Numbers for Comparison

Keep a reference node so results mean something. On a 2024 EPYC 9354 bare-metal node with enterprise NVMe we measured 4,180 events/s single-thread, NVMe 210,000 IOPS at depth 32 with p99 310 µs, and 18 ms RTT with 1.9 ms jitter. A consumer-grade VPS on shared storage measured 2,050 events/s, 38,000 IOPS with p99 4.6 ms, and 31 ms with 9 ms jitter.

The gap between those two nodes is the entire meaning of “dedicated”. If your candidate node scores closer to the second column, no amount of RAM or player-count tuning will hold 20 TPS at 80 players. Record your reference numbers once and compare every quote against them rather than against an abstract threshold.

Putting Results Into an Acceptance Gate

Run all four tests on a fresh node within the first hour of provisioning, before importing a world. Record results in a per-node sheet with date and firmware version. If any test fails, file it before you spend six hours migrating; most providers will move you to another chassis rather than argue, and a dated failure log is the argument.

Re-test after kernel updates, because a NIC driver or firmware change can shift p99 latency by an order of magnitude. Keep a reference run from month one so you can tell whether month nine degradation is the disk, the neighbours or your own world growth. The full hardware selection rationale sits in the dedicated server hardware pages.

What Failing Each Threshold Costs You

The thresholds are not arbitrary. A single-thread event p99 over 25 ms means roughly one in a hundred ticks blows its 50 ms budget at 20 Hz, which players see as periodic rubber-banding. NVMe p99 over 2 ms is longer than a full tick, so a save flush becomes a visible freeze rather than a slip. Both are measurable long before players are online.

A CPU fail shows as tick spikes every 5 minutes matching your save interval; a disk fail shows the same but worse at 50+ players. A network fail shows up as rubber-banding on players with good pings. Correlating the three timestamps — save window, tick spike, player complaint — takes a failing node from “feels bad” to a specific fixable component.

  • CPU p99 fail: move the instance or fix host CPU pinning
  • NVMe p99 fail: block passthrough missing or shared controller
  • RTT/jitter fail: rerouting needed, not more RAM
  • Save stall fail: split save to a separate NVMe volume

Bookmark the measured baseline next to your hosting evaluation notes and compare every new node against it. Two nodes that look identical on paper rarely measure identical, and the sheet is how you find out in hour one instead of week six.

Leave a Reply