UNIX for Dummies Questions & Answers

Hello

Is there a way to get CPU usage on UNIX environments (SunOS, Linux, HP-UX, AIX) without executing any commands?

The reason I ask is because I am asked to research any potential methods to monitor performance on UNIX environments but not use any commands because they want to 'poll' this every second and if the command itself drains some usage, then the monitoring tool would do more harm than good.

Any help is appreciated. Thanks.

You have to read data in kernel counters get what you want. That means you expend cpu cycles getting the information. There is no way around that.

You can minimize the issue by having a process running continually, waking up periodically, then sampling and going back to sleep. This uses a lot less resource than creating a new process (i.e., running a command, which may be what the 'others' are referring to) every time you issue a command.

However, this seems really odd to me. The impact of this kind of thing is minimal unless you get really heavy-handed with monitoring. Those tools are designed with the idea they need to have minimal system impact. Linux and Solaris have had kernel changes in the past just to accomodate the idea of minimizing resource use when monitoring.

Solaris has a tool, dtrace, which was also designed specifically with monitoring and NOT impacting the system it is monitoring.

On the other hand, some debuggers and profilers can really impact the system. So some performance monioring tools do degrade performance.

Overall, I think you are barking up the wrong tree - your worry is overblown somewhat.

How do you do something without doing anything? That's a very Zen question.

I don't think there's any commandless command capable of doing anything like a command.

Thanks for the replies.

I took look at gangalia and it seems that they access kernel stats exposed via /proc or kstat. i guess this is as 'light' as possible without degrading performance.

Depends how they wrote it, it does. Doing it without CPU is still kind of a fallacy.

Hi.
So from my point of view, if you are concerned with observing the work, then you must also be concerned about the generation of the data that you wish to analyze.

This suggests that the only measurement allowable would be completely external to the system in question. For example, using a stop watch, one could time a computational load, but could not depend on any other data from the system, because that would have an impact. Farther, one should disable the generation of data that records any data except that of computational answers -- no capture of CPU time, IO time, nothing except the results of computation.

So the answer seems to be easy: use any of the available commands that are not part of the system under consideration, the measurement needs to be wholly external. The measurement of a known computation can be recorded, then a multiple of that load, and the calculation of system characteristics can be done externally.

Good luck ... cheers, drl