Re: [PATCH 12/12] perf kvm: add live mode

[Arnaldo Carvalho de Melo <acme@ghostprotocols.net>]

Em Mon, Oct 08, 2012 at 02:54:10PM -0600, David Ahern escreveu:
> On 10/8/12 2:35 PM, Arnaldo Carvalho de Melo wrote:
> >Em Mon, Oct 08, 2012 at 11:17:36AM -0600, David Ahern escreveu:
> >Overall it looks good, I like the fact you're not doing it the way perf
> >script does, i.e. that "pipe mode" thing, more or less like I'm doing on
> >'perf trace', but I think this is the time to get rid of 'perf_session'
> >too for this specific "live mode" tools.
 
> pipe mode would never work for this case. I'm looking at nested virt
> case now and it shows 360,000 events per second.

Right, 'pipe mode' is an interesting construct, but its too inneficient
for same machine record/report, and for networking one we need to have
splice/sendfile, i.e. feed a socket directly in the kernel with what we
stash in mmap.
 
> >I'm almost getting there for 'trace' and 'top', just need to move the
> >ordered_samples code to evlist, then session, top, trace, kvm live and
> >script will not need to use 'perf_session' at all.
> 
> ok. you going to have that soon?

I hope to have it by the end of this week, perhaps earlier.
 
> >Also that 'perf_kvm__open_counters' really is just a third copy of code
> >in 'record' and 'top', right? You could move it to
> >'perf_ev{list,open}__open' and solve this dup :-)
> 
> there's a TO-DO at the top of that function for that reason ;-) I
> copied a fair bit from perf-stat. Now that it's working I can look
> at code consolidation.

Ok, just scratching again an old itch :-)
 
> >I applied the cleanups/simple stuff, will try to do some of the prep
> >work on ordered_samples to use on top and trace and then I think you
> >could look there so that we get to something that is useful for all
> >these !perf.data tools.
> 
> On my Westmere based server I am not seeing problems with timestamps
> and ordering (prior testing was on a Nehalem based server). It just
> weathered a 5 minute or so barrage of > 100,000 events per second
> (peaking at around 360k events/sec) with no problem.

Interesting, but does this involves consuming samples from multiple
CPUs?

- Arnaldo