Burn-In and Acceptance Testing a Server

Components rarely fail politely on schedule; they fail young or old. Burn-in exists to compress the young failures into the one week when they cost nothing — before data, workloads and expectations arrive. For refurbished machines the logic doubles: you are validating both the hardware and the refurbisher's story about it.

Day one: memory, overnight

RAM errors are the failures most worth finding early, because their symptoms downstream — corrupt files, crashed VMs, impossible bugs — never point back at the DIMM. Run the platform's built-in diagnostics for a first pass, then MemTest86 (or memtester under Linux) for a full overnight: every module, multiple passes. On ECC platforms also watch the BMC's correctable-error counters — a slot logging corrections during test is a slot to reseat, cross-swap and, if it follows the DIMM, return. Zero tolerance is the correct policy on day one.

Day two: storage, surface and truth

Record SMART baselines for every drive (hours, reallocated/pending sectors, SSD wear), then exercise: long self-tests on each device, followed by a real write/read pass — badblocks on drives that hold nothing yet, or a filesystem-level fill-and-verify. What you are hunting: pending sectors that appear under load, drives that reset mid-transfer (cable, backplane or controller conversations), and rebuild behavior — pull a member from a test array and watch the spare take over, because discovering a broken rebuild path in production is the expensive version of this test.

Day three: heat soak and power honesty

Load CPUs (stress-ng, Prime95) alongside disk activity for several hours while logging BMC temperatures and fan response. Healthy: CPUs plateauing in vendor range, fans stepping up then holding, no thermal-event log entries. Suspicious: one core running hot (paste/heatsink seat), fans oscillating (sensor or shroud), inlet creeping (room, not server). Then the tests everyone skips: pull one power cord under load — the machine should not blink and the BMC should log the loss; swap and repeat for the other PSU. Finish with a cold-boot cycle series — some marginal components pass every warm test and fail the third power cycle.

Sign-off: the record that pays later

Close by clearing test artifacts, updating firmware if not already current, and saving three things: the SMART baselines, a BMC log export, and a one-line config inventory. That file turns future anomalies into diffs (“reallocated went 0→12 since March”) and warranty conversations into evidence. Total cost: three quiet days and some electricity. Total value: every failure that now happens on a workbench instead of a Tuesday night.