The 20-Minute Dedicated Server Verification: Exact Commands and Order-Form Fields to Check Before You Sign

You get roughly one chance to test a dedicated server before a 12–36 month commitment, and it usually comes as a short trial, an SSH shell, or a provider’s demo box. This is a time-boxed verification pass: what to read on the order form, what to run on the machine, and the numeric thresholds that separate real hardware from a virtualised reseller. Everything below fits in about twenty minutes and needs no tooling beyond the standard Linux utilities plus ipmitool.

Step 1 — Order-form fields to capture in writing

Ask for these in the pre-sales ticket, not on a phone call. A provider that will not put the numbers in writing usually cannot deliver them.

Field	What to look for	Red flag
Exact CPU model	Xeon Gold 5412U / EPYC 9354P / Ryzen 9 7950X3D	“Intel Xeon” or “modern CPU”
Port speed	10 Gbps guaranteed on gaming SKUs	“up to 1 Gbps” shared
Bandwidth commitment	Unmetered 10 Gbps or 100+ TB	10 TB then charge per GB
DDoS mitigation	Always-on, Gbps figure quoted	On-demand filtering only
OOB access	IPMI / iDRAC / iLO included by default	Paid add-on or ticket-only rescue
Hardware replacement SLA	4-hour guaranteed swap	“best effort, 24–48 h”
Power feeds	Dual A+B, redundant PSU	Single PSU, single feed
IP allocation	/29 usable plus private VLAN	Single IP with per-IP fee

Step 2 — Confirm it is real bare metal

sudo systemd-detect-virt                 # expect: none
lscpu | egrep 'Model name|Socket|Core\(s\)|MHz|Hypervisor|Flags'
nproc --all
sudo dmidecode -t memory | egrep 'Size|Speed|Type|Locator|Rank'
sudo dmidecode -s system-product-name    # real chassis model, e.g. PowerEdge R7615
lspci | egrep -i 'nvme|ethernet|raid'

systemd-detect-virt returning anything other than none means you are buying a VM with a dedicated label. dmidecode telling you the memory is single-rank DDR4-2666 when the SKU promised DDR5-4800 is the other common substitution, and it directly caps your game server’s memory bandwidth.

Step 3 — Storage tier and real throughput

lsblk -d -o NAME,ROTA,SIZE,MODEL            # ROTA=0 and NVMe in MODEL
sudo nvme list                              # expect /dev/nvme0n1, Gen4 or better
sudo hdparm -tT /dev/nvme0n1                # sequential read
sudo dd if=/dev/zero of=/root/verify.bin bs=1M count=4096 oflag=direct conv=fdatasync
sudo smartctl -a /dev/nvme0n1 | egrep 'Percentage Used|Power_On_Hours|Media_Wearout'
cat /sys/block/nvme0n1/queue/scheduler

Check	Acceptable 2026 value	Why it matters
Device type	NVMe SSD, ROTA=0	SATA SSD halves ARK and Palworld world-save speed
Sequential read	>3,000 MB/s (Gen4 >6,000)	Directly sets map-load time
Sequential write	>1,000 MB/s	Determines autosave stall length
Power_On_Hours	<20,000 h	A 40,000 h drive is near end of life
Percentage Used	<20%	Worn drives cause write latency spikes
I/O scheduler	none (NVMe)	mq-deadline adds overhead on NVMe

Step 4 — Network quality, not just speed

ethtool eth0 | grep -E 'Speed|Duplex'      # expect 10000Mb/s, Full
mtr -rwzbc500 1.1.1.1                      # loss and jitter over 500 packets
ping -A -c 2000 -i 0.2 <peer-ip> | tail -n 3
iperf3 -c <provider-speedtest> -P 4 -t 20
vmstat 1 5 | awk '{print $16}'             # steal time column
cat /proc/net/softnet_stat | awk '{print $2}'

Loss above 0.2% over 500 packets or jitter above 3 ms will be visible in-game on a 64-tick server. Steal time above 1% means another tenant is consuming your CPU, which is the single clearest sign of an oversubscribed host. ethtool reporting 1000Mb/s when the order form said 10 Gbps is a billing-grade discrepancy worth escalating immediately.

Step 5 — Out-of-band and power control

ipmitool -I lanplus -H <bmc-ip> -U admin -P '<pass>' chassis power status
ipmitool -I lanplus -H <bmc-ip> -U admin -P '<pass>' mc info
ipmitool -I lanplus -H <bmc-ip> -U admin -P '<pass>' sel list | tail -n 20
ipmitool -I lanplus -H <bmc-ip> -U admin -P '<pass>' sol activate

Test the reboot path deliberately: run chassis power cycle and time how long the box takes to come back to a responsive SSH prompt. Anything past 5 minutes on a dedicated machine points to a slow POST or a shared storage mount. If the provider has no OOB at all, ask for their rescue-mode response time in writing — when a kernel panics during a 200-player Rust wipe, a 30-minute support ticket is a lost population.

Step 6 — The five-minute load reality check

• Run your real workload, not a synthetic benchmark: start the actual game server, fill it to 70% of the intended slot count with bots, and watch tick and CPU.
• Add a 100 MB write every 60 seconds to the data directory to emulate autosaves and confirm the game process never stalls past 300 ms.
• Generate 500 Mbps of outbound UDP and verify throughput and a stable RTT with mtr running alongside.
• Check thermals via ipmitool sdr type temperature after 20 minutes of load; throttling above 85°C on a leased box usually means failed fans the provider has not noticed.

Red flags that justify walking away

• The trial box and the production box have different CPUs — always re-run Step 2 after provisioning.
• Bare-metal claim proven false by systemd-detect-virt.
• No written SLA on hardware replacement or a “best effort” clause.
• DDoS mitigation measured in Mbps rather than Gbps for a public game server.
• Autosave stalls visible in the trial, with no NVMe upgrade path offered.

Use this pass on any candidate, then cross-check the numbers against our dedicated server comparison before you commit. If you want the shortlist-first approach instead, start by reading how to compare dedicated server providers on price-per-spec, then come back and verify the winner with the commands above. Fifteen minutes of dmidecode and mtr output is worth more than any spec sheet.