Re: pmlogger -u questions

[Nathan Scott <nathans@xxxxxxxxxx>]

----- Original Message -----
> ----- Original Message -----
> [...]
> Yep, agreed.  Actually I would say the log write I/O patterns appear to
> be noticeably improved for the more realistic sizes, and well worth the
> slight CPU overhead.
> 
> However, given this:
> > Test cases 10,000 samples at 10msec intervals.
> 
> Could we re-create that table at a 10sec logged interval?  Perhaps
> just for 100 samples - that'd certainly satisfy my curiosity anyway.

Pondering further overnight, and after some code digging - there's no point
in doing this, I think - please ignore.  The answers will be (effectively)
the same AIUI, as there is nothing in the __pmOptFetch APIs that'll change
behaviour based on a different delta - same number of writes should result.

There was an earlier comment about difficulty in discerning physical I/O
resulting from pmlogger writes from those issued by the rest of the system.
One way to isolate that (somewhat) is to write to a separate unused device.
This device can then be targetted by blktrace(1).  Other factors still play
an overwhelming role though - filesystem characteristics, like delalloc vs.
allocating space when pmlogger issues the write(2), memory utilisation, and
so on.

cheers.

--
Nathan