|To:||Stan Cox <scox@xxxxxxxxxx>|
|Subject:||Re: [pcp] python bindings|
|From:||Nathan Scott <nathans@xxxxxxxxxx>|
|Date:||Mon, 25 Jun 2012 10:25:03 +1000 (EST)|
|Cc:||pcp@xxxxxxxxxxx, Michael Werner <mtw@xxxxxxxxxxxxxx>|
Looked at this a little bit more on the weekend - there were a few places where
the default output didn't quite match up with collectl output for me, so pushed a
change to change that behaviour a little. There was a few fields that looked like
they needed to be converted to KB, and inspection of the collectl source shows
they're using 1024 bytes-per-KB - so I've changed over a few places.
I noticed a couple of things I didn't tackle that might be good to do too, but are
a bit beyond my dodgey python skills. There's a "metric-type" data structure,
with hard-coded metric-to-type mappings. Instead of hard-coding names of
metrics known to not need rate conversion, we could use the metric desc struct -
pmLookupDesc(3). This structure has a 'semantics' field (pmDesc.sem) and we
can test for PM_SEM_COUNTER. That would work for any metric (existing /
newly added) and the metric-type structure would be unnecessary.
We should be able to implement collectl playback mode fairly easily (using PCP
archives). My python-fu is not up to the task yet though - I see the pmContext()
call but not clear how to give it a filename to put it into archive mode. Any clues?
I'd like to add pmlogger and pmlogconf configuration files which will let someone
easily get recording of the necessary metrics setup.
|<Prev in Thread]||Current Thread||[Next in Thread>|