On Tue, 01 Mar 2005 at 13:23 +0000, enio@... wrote:
> John Levon writes:
> >On Sat, Feb 19, 2005 at 03:30:06AM +0100, Philippe Elie wrote:
> >>I missed the call from pp tools to flush image file to disk, I though it
> >>was done by the daemon through an opcontrol --dump. John what about this ?
> >I'm unhappy with this and waiting for a response from Scott.
> Since I implemented this, let me take a crack at it ...
Sorry for the late reply.
> In the previous notes there have been basically 3 proposals/preferences as
> to when to generate the ELF files.
> 1) The current implementation modifies the PP tools to generate the ELF
> files at the time they're needed. Doing so ensures that the generated ELF
> files match all collected anonymous sample information.
> - PP tools are modified.
> - Postpone file generation overhead head until the time files are actually
> 2) A proposed implementation is to have the daemon generate the ELF files
> - Must determine a "good" time interval to write files.
> - Added profiling overhead. File(s) must be written several times,
> depending on the interval chosen, during the profiling run.
> - PP tools are not modified.
> 3) Another proposed implementation is to generate the ELF files on demand:
> a) By the daemon at --dump time
> b) By a stand-alone command
> Both a and b would have the same end result, so I'll go under the
> assumption that --dump is generating the files.
> - What if users don't --dump before running PP tools?
use of pp tools w/o a --dump after the task being profiled finished is an
undefined behavior. We don't document, and can't, how many time sample
remains in kernel buffer
> - Daemon is still responsible for generating ELF files.
> - Fits in well with the --dump semantics.
> I chose 1) because I was trying to postpone the overhead of generating the
> files until the time they were actually needed.
> I personally don't believe the daemon should have to deal with generating
> the ELF files. I think that belongs in the PP tools, but we'll put that
> functionality wherever you think it belongs.
I see your point now, boring to choose the best way. John I'm about to
accept the patch w/o taking care about this point. The change needed to
switch from 1) to 3) are small and at the moment I can easily change my
mind from 1) to 3) then 3) to 1) ...