Re: [RFCv2 00/48] perf tools: Add threads to record command

[Jiri Olsa]

On Fri, Sep 14, 2018 at 02:13:07PM +0200, Ingo Molnar wrote:
> 
> * Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> > On Fri, Sep 14, 2018 at 01:47:25PM +0200, Jiri Olsa wrote:
> > > On Fri, Sep 14, 2018 at 01:15:28PM +0200, Peter Zijlstra wrote:
> > > > On Fri, Sep 14, 2018 at 11:40:22AM +0200, Ingo Molnar wrote:
> > > > > In fact keeping the files separate has scalability advantages for 'perf report' and similar 
> > > > > parsing tools: they could read all the streams in a per-CPU fashion already, from the very 
> > > > > beginning.
> > > > 
> > > > Also writing to different files from different CPUs is good for record,
> > > > less contention on the inode state (which include pagecache).
> > > 
> > > maybe I should explain a little bit more on this
> > > 
> > > we write to different (per-cpu) files during the record,
> > > and at the end of the session, we take them and store
> > > them inside perf.data
> > 
> > How long does it take to combine that? If we generated a lot of data,
> > that could take a fair amount of time, no?

yep.. fair amount ;-) wasn't that bad in my tests,
but could be evil on some really big server

> > I feel that record should not mysteriously 'hang' when it is done. It
> > used to do that at some point because of that stupid .debug crap, but
> > acme fixed that I think.
> 
> Agreed - plus at the report stage it would be advantageous to be able to *read* per-cpu files 
> as well.
> 
> If we do things smartly them report will create similar NUMA affinity as the record session 
> used.

ok, separate files it is

jirka