Skip to content

Server Slow

A server is reported slow with no other detail. The interviewer watches whether the candidate clarifies what slow means, then runs an ordered checklist that places the bottleneck on a resource, rather than guessing or restarting blindly.

Symptom

"The server feels slow. Everything is sluggish. Figure out why."

Clarifying Questions

  • Slow how, in numbers? Higher request latency, a stalled batch job, and slow logins are different problems.
  • When did it start, and what changed? A regression at a deploy time points at code, not capacity.
  • One host or many? A fleet-wide slowdown points at a shared dependency, not this box.
  • Is it getting worse? A rising load average means the cause is still active.

Diagnostic Path

Run the 60-second checklist in order on web (Rocky Linux 10.2), letting each command rule a resource in or out. Read dmesg first, because a hardware or memory event names the cause outright.

1. Load and Recent Events

uptime
dmesg -T | tail -3

Output:

 23:00:11 up 5 min,  0 user,  load average: 1.02, 0.24, 0.08

A load average rising across the three windows means the pressure is growing now. dmesg clean of OOM or I/O errors rules out the loudest causes.

2. Run Queue and CPU Split

vmstat 1 3

Output:

 r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy id wa st
 4  0      0 343284  10780 430336    0    0   681   710  127    0  2  1 97  0  0
 3  0      0 343284  10780 430336    0    0     0     0  511  168 100  0  0  0  0

The run queue r is 3 to 4 on two CPUs, and us is 100 with wa and st at 0. This is CPU saturation in user code, not I/O or swap, so the memory and disk branches can be skipped for now.

3. Which Process, and Balance

mpstat -P ALL 1 1 | tail -3
pidstat 1 1 | head -6

Output:

22:59:52       1  100.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00
23:00:11        0      1466   61.39    0.00    0.00   38.61   61.39     1  stress-ng-cpu
23:00:11        0      1465   77.23    0.00    0.00   22.77   77.23     1  stress-ng-cpu

Both CPUs are pinned, and pidstat names the consumers with high %wait, meaning more runnable threads than cores. If instead one core were hot and the rest idle, the work would be single-threaded and more cores would not help.

4. Rule Out Memory and Disk

free -h | head -2
iostat -xz 1 1 | tail -3

Output:

Mem:           975Mi       371Mi       344Mi       3.4Mi       400Mi       603Mi
vda             20.77    343.27      0.35   11.10    361.25      0.86     0.02   2.98

available memory is healthy and swap is not paging, so memory is not the limit. The disk shows %util 2.98 and aqu-sz 0.02, so I/O is idle. Both branches are clear, confirming the CPU finding.

Root Causes

Branch Evidence Fix
CPU saturated vmstat r > cores, us high, pidstat names it Scale out, optimise, or throttle the process
Single-threaded One mpstat core at 100%, rest idle Parallelise; more cores do not help
Memory Low available, vmstat si/so active Fix the leak or add RAM
Disk I/O iostat %util near 100, high aqu-sz/await Find the writer, reduce I/O
Steal st non-zero and persistent Move to a larger or dedicated instance
Application Host resources clear, latency still high Look at the app, locks, dependencies

Fix

For the captured CPU-bound case, identify and act on the process rather than rebooting. A CPU-heavy batch job can be reniced or moved; a runaway loop needs a code fix.

renice +10 -p $(pgrep -n stress-ng)     # lower its priority
# or cap it in a slice: systemd-run --scope -p CPUQuota=50% ...

If the checklist is clean but users still see slowness, the bottleneck is above the host: the application, a lock, or a downstream dependency.

Prevention

  • Keep sar history running, so a slowdown can be compared against a baseline.
  • Alert on saturation (run queue, disk queue) and on user-facing latency, not on raw utilization.
  • Set CPU and memory limits on batch and background work, so it cannot starve the main service.

Captured on Rocky Linux 10.2 on an iximiuz Labs FlexBox microVM, kernel 6.1.167, 2026-09. Load was generated with stress-ng for the capture.