0

What are recommended threshold for Linux performance metrics?

I'm collecting statistics from remote servers using rstatd.

On which values should I issue warnings? errors?

2 Answers 2

2

establish baseline under normal load and start from there. that's it.

and for stats gathering - you can also take a look at munin, cacti; for alerting - nagios, zabbix.

2
  • I am looking for more concrete, industry standard thresholds, like: CPU should be less than 80%, Average Load should be less than the overall number of CPUs.
    – yshalbar
    Oct 2, 2010 at 15:39
  • 1
    but there is no industry standard. load average of 16 can be quite ok for some workloads and totally unacceptable for another. same with swap usage - you might permit it for some servers and want to avoid it by all means for another. same for saturated one or few cores of cpu or network interfaces running at full speed. no standards - it all depends what do you expect from machine - if it's real time or batch processing and so on...
    – pQd
    Oct 2, 2010 at 16:46
1

As pQd already mentioned, it's best to know what's the baseline, intervene when the state deviates from there.

As for specific thresholds that shouldn't be exceeded if you want the server to be snappy:

  • load average shouldn't be more than the number of cores (physical, not HT created)
  • no swap usage
  • disk duty cycle shouldn't exceed about 80%

Though it all depends on the kind of workloads you're running. For example, you can have higher load avg if you have many low priority (with high niceness) jobs in the background.

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .