system.cpu.load and Windows Agent

[jpomorski]

Hi,

Could somebody explain me how Zabbix Agent in Windows get system.cpu.load value? I'm asking about versions 3.0 and higher.
Previously I used version 2.2.7 and values were more or less comparable to values from "System\Processor Queue Length" counter. After upgrade to 3.0.4, I receive much higher values.

[allexpetrov]

Hi,

Could somebody explain me how Zabbix Agent in Windows get system.cpu.load value? I'm asking about versions 3.0 and higher.
Previously I used version 2.2.7 and values were more or less comparable to values from "System\Processor Queue Length" counter. After upgrade to 3.0.4, I receive much higher values.

Hi, could you please show the values (from Windows as from Zabbix) ? As well as paste the your key.

Thank you,
Alexey!

[jpomorski]

Hi Alexey

Values from Zabbix



Values from Performance Monitor on monitored machine



Item configuration



Jarek

[allexpetrov]

Hi Jarek,
please expand the item.

Thanks,
Alex!

[jpomorski]

[jan.garaj]

system.cpu.load - windows implementation: https://www.zabbix.org/websvn/wsvn/z...fo/win32/cpu.c:

 if (NULL == (tmp = get_rparam(request, 1)) || '\0' == *tmp || 0 == strcmp(tmp, "avg1"))
        {
                ret = get_perf_counter_value(collector->cpus.queue_counter, 1 * SEC_PER_MIN, &value, &error);
        }

So cpus.queue_counter is used. But that's 1 minute average in your case != 1 second average in Performance Monitor case.

[jpomorski]

Jan, thank you for above link.

I've done some investigation, and IMO the problem is related to power management settings in Windows. If power plan is set as Balanced or Power saver, then on less loaded server,s some cores are parked (https://support.microsoft.com/en-us/...havior-changes). After change power plan on High performance, core parking is disabled, and values from Zabbix Agent are comparable to values from Performance Monitor.

I'm doing code review of Zabbix Agent, but I'm not dev, so it takes me a lot of time... Maybe someone who knows better Zabbix code, can do it faster and better, and can confirm (or not) my guesswork...

Jarek

[jan.garaj]

Please try to use item:

system.cpu.load[percpu,avg1]

percpu = total load divided by online CPU count = normalized CPU load

https://www.zabbix.com/documentation/3.0/manual/config/items/itemtypes/zabbix_agent

[jpomorski]

After change key to system.cpu.load[percpu,avg1], returned values are divided by 2 (server has two 1 core vCPUs) but still too high beside to values from Performance Monitor.