Most dedicated hosting quotes are written to be hard to compare. “Dual Xeon, 64GB RAM, NVMe storage, 1Gbps port, 99.9% SLA” fits almost any configurable box on any provider’s price sheet, so the only visible difference is the number at the bottom. The vetting work is converting each of those phrases into a specific model, a specific measured number, or a specific contractual term, and refusing to sign until you have all three in writing.
The stakes are modest until they are not. A box that looks like a bargain at 40% less per month but ships SATA SSD instead of NVMe will choke a database at 8k random write IOPS, and you will discover it during a traffic spike rather than during the 24-hour benchmark window you did not run. Ask every question below before ordering, and expect the answers to appear in the quote email.
Start with the board and CPU, not the core count. “Xeon Silver” covers three generations that differ by 30% in single-thread performance, and single-thread is what most PHP and game-tick workloads actually consume. Ask for the exact processor SKU (e.g. Xeon Silver 4314, EPYC 7443P), the base and boost clocks with all cores loaded, and whether the board is a single-socket or dual-socket platform with slots free. Then ask for RAM speed and channel count: 8×16GB DDR4-3200 across eight channels is a far better platform for a database than 4×32GB DDR4-2666 in quad channel, at identical capacity. Providers that cannot answer this in one email are reselling someone else’s hardware. The dedicated server plans that publish these specs are the ones worth comparing line by line.
Make them prove the disk, not just name it
NVMe is a controller and a bus, not a performance figure. The same PCIe 4.0 drive can deliver 6,000 MB/s sequential and still fall to 90k random 4k write IOPS under a mixed queue, and a provider that oversubscribes a consumer drive across twelve VMs will deliver a fraction of that. Ask for a fresh fio run on the actual node: fio --rw=randwrite --bs=4k --iodepth=32 --numjobs=4 --direct=1. You want p99 latency under 1ms sustained for five minutes, and you want to see it before the disk has a week of neighbour noise on it.
- Storage model: enterprise NVMe (e.g. PM9A3, D7-P5520) or consumer drive?
- Controller RAID level, and is the write cache battery-backed?
- Measured 4k random read and write IOPS at QD32, plus p99 latency
- Drive endurance rating (DWPD) — order tables delete-and-insert heavily
Port speed is not bandwidth guarantee
A 10Gbps port means your NIC negotiates 10Gbps; it does not mean you can push 10Gbps. Ask for committed bandwidth in Mbps, whether it is metered or unmetered, what the 95th percentile billing rule is, and whether inbound is free. “Guaranteed 1Gbps unmetered” is a legitimate offer; “10Gbps port, 200Mbps committed” is also legitimate, and the two are priced differently for good reasons. For a game server pushing 0.4Mbps per player, 500 committed Mbps is 1,000 players; for a video backend it is nothing.
Also ask about transit mix and whether the path to your players is optimised or best-effort. A single-carrier blend adds 15–30ms to transatlantic traffic compared with a tier-1 mix, and for a competitive shooter that is the difference between playable and refunded.
Do the SLA credit math yourself
99.9% monthly uptime permits 43 minutes and 12 seconds of downtime. 99.95% permits 21m 36s. Read what counts as downtime: many SLAs exclude scheduled maintenance, upstream carrier failures, and anything caused by the customer’s own load. Then read the credit: a 10% one-month credit on a $200 server is $20, and if the box is down for six hours you have lost far more than $20 in players, ad spend, or cart abandonment.
| Monthly uptime | Allowed downtime | Typical credit |
|---|---|---|
| 99.9% | 43m 12s | 5–10% |
| 99.95% | 21m 36s | 10–25% |
| 99.99% | 4m 19s | 25–100% |
| Add-on: hardware replace <4h | — | negotiated |
IPMI, rescue mode, and who actually has the console
Ask whether the quote includes full IPMI or a restricted KVM-over-IP session, how you get in when the OS will not boot (rescue image? ISO mount?), and how long those take to enable after a ticket. A provider that requires a support ticket and a two-hour wait for a rescue boot is a provider you cannot operate a production box on at 2am. Confirm the console is available from day one, not “on request”.
Test it before you trust it: open IPMI, mount an ISO, boot it, and time the whole loop while the machine is still empty of customer data.
Replacement hardware terms
The question is not whether hardware fails but what happens in the four hours after it does. Ask for the stated replacement SLA in hours, whether spare drives and PSUs are stocked in the same facility, whether CPU or RAM failure means a board swap (full node migration) or a parts swap, and whether you get a replacement of equal or better spec. Put a number on it: four hours is the usual floor for a competent provider, 24-hour is not acceptable for a production game server.
Also confirm data migration: if the node is replaced, does the provider move your disks, or do you restore from backup? Get that in the quote too, because it determines whether you need a 10TB off-site backup budget.
The five-line summary email
Send one final message asking the provider to confirm, in plain text: CPU SKU and all-core clock, RAM config and channel count, drive model with measured 4k IOPS at QD32, port speed with committed Mbps, IPMI and rescue access terms, replacement SLA hours, and SLA credit percentages. If all seven come back clean, you have a comparable quote. If three come back as marketing language, you have your answer.
Vet this way once and the comparison stops being about price sheets. When the hardware spec and SLA terms are explicit, the remaining decision is purely whether the measured numbers fit your workload.
Leave a Reply
You must be logged in to post a comment.