Re: [pcp] Process analysis

[Nicolas Michel <be.nicolas.michel@xxxxxxxxx>]

Thank you all for your ideas. We'll try the native web app ;)

The majority of ideas you propose, if I understand well, makes possible to measure the stats of predefined processes we have to configure. It is great when you know in advance which process and what in the process you want to watch.

Before, in case of a performance problem (a real case I had here), you first have to identify which process is the cause of the performance problem.

So I would wanted something more like a top with history, and with I/O stats (it only keeps says the stats of the tenth hungrier processes). I think it would be a great plugin, disabled by default of course. And I would want to check process stats every hour or 2 hours (it is a matter of reading /proc/* I suppose?). In case of problem, we could change the precision or even seing it "live" like when connecting to pmcd directly with pmchart.

I need more a by-process statistics than intra-kernel stats. That "magical" plugin ;) would able me to tell: it is the process XYZ with PID XXX that from 01:00pm to 02:00pm used the more the I/O, at a first glance. Then I would be able to watch that particular process more closely/precisely, with custom probes if necessary, which systemtap-pcp is for if I understood well.

I know collected have a feature that can help: you can have a timeline of processes forking. But collected is not well packaged and difficult to install/configure (and often crash). And it doesn't give me I/O, memory and CPU stats by-process.

2014-11-14 16:27 GMT+01:00 Frank Ch. Eigler <fche@xxxxxxxxxx>:

Hi -

On Fri, Nov 14, 2014 at 12:00:24PM +0530, Shirshendu Chakrabarti wrote:

> This can be a probable solution. You can use systemtap to trace the
> desired process. [...]Â Please see, AFAIK, there is no known
> integration with PCP for above solutions.

systemtap et al. are certainly possible sources of additional data for
PCP. We're in the process of building out a richer systemtap-pcp
interface. Do you have any more specific ideas about what new
per-process or intra-kernel statistics you need collected?

- FChE

--

Nicolas MICHEL