Why twenty minutes is enough
A new dedicated server arrives with a clean operating system and a confident email. Neither tells you whether the RAID array is fully initialised, whether one DIMM is running at half speed, or whether the second power supply is actually plugged into a different feed. A short, ordered acceptance check catches almost every delivery defect before you put customer traffic on the box, and it gives you a timestamped baseline to argue from if something degrades later.
The sequence below is deliberately ordered from lowest-risk to highest. Read-only hardware inspection comes first, configuration review second, and any destructive or performance testing last so you never have to reinstall mid-check. Budget roughly twenty minutes of active work; if a step fails, stop and raise a ticket with the command output attached.
Minutes 0-5: platform identity and firmware
Start by confirming you received the machine you ordered. The vendor and model in the baseboard record should match the order form, and the BIOS or UEFI version should be recent enough to include the security patches and microcode revisions for your CPU generation.
dmidecode -t system— manufacturer, product name, serial number matched to your invoice.dmidecode -t bios— BIOS version and release date.lscpu— socket count, cores per socket, threads per core, max MHz.uname -randcat /etc/os-release— kernel and distribution against the order form.
If the BIOS date is more than a year old on a current-generation platform, ask the provider to update it before you proceed. Microcode updates matter for stability and for speculative-execution mitigations, and doing it now is far cheaper than doing it during a maintenance window later. Note the firmware versions of the RAID controller and NIC as well — storcli /c0 show and ethtool -i eth0 reveal both.
Minutes 5-9: memory integrity and ECC
Memory faults are the most common silent cause of mysterious crashes. Confirm the capacity, the module count, the speed, and crucially the ECC mode. A server running non-ECC memory, or ECC that is disabled in the BIOS, is not what you paid for.
dmidecode -t memory— total, per-slot size, configured speed, and rank count.lshw -short -C memory— a compact capacity and form-factor summary.journalctl -k | grep -i -E "edac|ecc|correctable"— kernel ECC messages since boot.grep -H . /sys/devices/system/edac/mc/mc*/ce_count— correctable error counter per controller.
Zero correctable errors on a fresh box is normal. A non-zero count on day one suggests a marginal DIMM and is worth escalating immediately. Also confirm the DIMMs are populating all memory channels: six modules across a four-channel platform is fine, but two modules on a four-channel platform halves your usable bandwidth.
Minutes 9-13: disk layout, SMART, and RAID state
This is where most delivery surprises hide. A RAID array can report “optimal” while still initialising in the background, and a drive can be brand new with a firmware bug. Check both layers.
| Check | Command | Pass condition |
|---|---|---|
| Block layout | lsblk -o NAME,SIZE,TYPE,MODEL |
Drives match order form |
| Drive health | smartctl -a /dev/sdX |
No reallocated or pending sectors |
| NVMe wear | nvme smart-log /dev/nvme0 |
Percentage used near zero |
| Software RAID | cat /proc/mdstat |
All arrays [UU], no resync |
| Hardware RAID | storcli /c0/vall show |
State Optimal, no init in progress |
Confirm the RAID level too, not just the health. A provider that shipped RAID 0 when you ordered RAID 1 has given you no fault tolerance at all, and the array will still report optimal. Finally, check whether a hot spare exists if your design called for one.
Minutes 13-17: network, offloads, and IP allocation
Network verification has three parts: the physical link, the software offloads, and the address plan. All three appear on the order form and all three are frequently wrong.
ethtool eth0— negotiated speed and duplex; confirm 10000Mb/s, not 1000Mb/s.ethtool -k eth0 | grep -E "tso|gso|gro|rx-checksum"— expected offloads enabled.ethtool -S eth0 | grep -i error— link and receive errors should be zero.ip -br addrandip route— count usable addresses versus the sheet.mtr -rwc 50 1.1.1.1— packet loss on the first hops toward your users.
Watch for a /29 that is presented as five usable addresses but turns out to be bridged rather than routed, which breaks any virtualisation or failover plan you had. If you ordered additional IPs, ping each one from outside before signing off.
Minutes 17-20: IPMI, power, and order-form field pitfalls
Remote management is the difference between a ten-minute fix and a four-hour outage. Confirm the BMC answers on its management address, that your account can power-cycle the box, and that sensor readings look sane.
ipmitool -I lanplus -H <bmc-ip> -U <user> sdr— PSU, fan, and temperature sensors.ipmitool ... chassis power status— confirm power state reporting works.ipmitool ... sdr type "Power Supply"— both PSUs present and healthy.
Order-form fields deserve their own scrutiny. The most common pitfalls are: “bandwidth” meaning port speed rather than committed rate; “5 IPs” meaning a /29 including network and broadcast; “RAID” left unspecified, defaulting to RAID 0; and “management” meaning a web panel rather than a true out-of-band KVM. Re-read the confirmation email line by line against the original spec, and put any correction in writing before the first invoice.
Recording the baseline
Save the full output of every command above into a dated file on the server and a copy off-box. When something breaks in six months, that baseline is the fastest way to prove whether a drive, a DIMM, or a link has degraded. Re-run the same script quarterly on production nodes so drift is visible early.
If you are still selecting a provider, our overview of high-traffic workloads covers which of these commitments to insist on in the contract, and how they differ across the dedicated server hosting tiers you will be comparing.




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