Telegraf config to collect per-CPU cycles and instructions

cpu_cycles_and_instructions_count.conf.toml

# Telegraf configuration to collect per-CPU cycles and instructions
# using perf(1).  You can then calculate CPI / IPC from the metrics.

# If Telegraf is not running as UID 0 (aka "root") you can either use
# sudo(1) with the perf(1) command, raise Telegraf capability, or set
# kernel.perf_event_paranoid = 0.  But be warned: "Hic sunt dracones".

[[inputs.exec]]

  name_prefix = "cpu_"
  name_suffix = "_sample_1s"

  interval = "10s"

  commands = [
    "perf stat -x, -A -e cycles,instructions --timeout 1000 --log-fd 1"
  ]

  data_format = "csv"

  csv_column_names = ["cpu", "count", "dummy", "probe"] # the rest is ignored
  csv_tag_columns = ["cpu"]
  csv_measurement_column = "probe"

  fielddrop = ["dummy"]

Sample:

cpu_cycles_sample_1s,cpu=CPU0,host=host.example.com count=1296679258i 1656065203000000000
cpu_instructions_sample_1s,cpu=CPU0,host=host.example.com count=999403573i 1656065203000000000

Enjoy!