Re: [RFC] scripts: Include postprocessing script for memory allocation tracing

From: Michal Hocko
Date: Thu Oct 06 2016 - 08:01:09 EST


On Fri 23-09-16 10:07:09, Michal Hocko wrote:
> On Thu 22-09-16 11:30:36, Janani Ravichandran wrote:
> >
> > > On Sep 19, 2016, at 5:42 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > >
> > > On Tue 13-09-16 14:04:49, Janani Ravichandran wrote:
> > >>
> > >>> On Sep 12, 2016, at 8:16 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > >>
> > >> Iâm using the function graph tracer to see how long __alloc_pages_nodemask()
> > >> took.
> > >
> > > How can you map the function graph tracer to a specif context? Let's say
> > > I would like to know why a particular allocation took so long. Would
> > > that be possible?
> >
> > Maybe not. If the latencies are due to direct reclaim or memory compaction, you
> > get some information from the tracepoints (like mm_vmscan_direct_reclaim_begin,
> > mm_compaction_begin, etc). But otherwise, you donât get any context information.
> > Function graph only gives the time spent in alloc_pages_nodemask() in that case.
>
> Then I really think that we need a starting trace point. I think that
> having the full context information is really helpful in order to
> understand latencies induced by allocations.

Are you planning to pursue this path?
--
Michal Hocko
SUSE Labs