CPU usage spike with gosigar + metricbeat (process) on FreeBSD

[michbsd]

Hi,

I running the latest build on metricbeat on FreeBSD 10.1-REL.
As soon as activating it, specifically the process metricset - the CPU usage spikes to 40-50%.

The metricbeat devs suggest that this might be an issue with gosigar.

Please have a look at the thread here:
https://discuss.elastic.co/t/metricbeat-on-freebsd-using-a-lot-of-resources/55412/20

There is also a link to a pprof trace.

This is my current gosigar revision:
61dfed4

Let me know if you need anything else.

Thanks,

[michbsd]

Any news on this?

[ruflin]

@michbsd No updates so far from our side. Pinging @knz and @gregoryo2014 as they could also be active users of metricbeat on freebsd and it would be interesting if they have a similar problem?

[knz]

I've not seen this happening earlier. To me it looks as if the client code was polling the metrics too often. Since this code is going via /proc it will involve more machinery in the kernel than a simple sysctl/ioctl call.

[knz]

If you want to help you could first instrument / profile the code to see how often the API in gosigar is being invoked. If yous ee these usage spikes with a poll rate lower than one time per second, then there would be cause for concern.

[michbsd]

@knz sure I can help with that. Can you point me to a guide / doc / tutorial on doing this?

thanks

[knz]

print time.Time() at the beginning of every relevant function. Then look in the output of your program which times are being printed and how much time elapses between each print.

[michbsd]

@knz right - to I add this to all "process" related functions in metricbeats - print e.g function-name + time ?

[knz]

for example yes