linux-kernel - Re: [PATCH v2 06/15] perf session: load data directory into tool process memory

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20201027122154.GF2900849@krava>
Date:   Tue, 27 Oct 2020 13:21:54 +0100
From:   Jiri Olsa <jolsa@...hat.com>
To:     Alexey Budankov <alexey.budankov@...ux.intel.com>
Cc:     Arnaldo Carvalho de Melo <acme@...nel.org>,
        Namhyung Kim <namhyung@...nel.org>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        Adrian Hunter <adrian.hunter@...el.com>,
        Andi Kleen <ak@...ux.intel.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Ingo Molnar <mingo@...hat.com>,
        linux-kernel <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v2 06/15] perf session: load data directory into tool
 process memory

On Tue, Oct 27, 2020 at 10:37:58AM +0300, Alexey Budankov wrote:
> 
> On 24.10.2020 18:43, Jiri Olsa wrote:
> > On Wed, Oct 21, 2020 at 07:01:19PM +0300, Alexey Budankov wrote:
> >>
> >> Read trace files located in data directory into tool process memory.
> >> Basic analysis support of data directories is provided for report
> >> mode. Raw dump (-D) and aggregated reports are available for data
> >> directories, still with no memory consumption optimizations. However
> >> data directories collected with --compression-level option enabled
> >> can be analyzed with little less memory because trace files are
> >> unmaped from tool process memory after loading collected data.
> >> The implementation is based on the prototype [1], [2].
> >>
> >> [1] git clone https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git -b perf/record_threads
> >> [2] https://lore.kernel.org/lkml/20180913125450.21342-1-jolsa@kernel.org/
> >>
> >> Suggested-by: Jiri Olsa <jolsa@...nel.org>
> > 
> > very loosely ;-) so there was a reason for all that reader refactoring,
> > so we could have __perf_session__process_dir_events function:
> > 
> >   https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git/commit/?h=perf/record_threads&id=308aa7cff1fed335401cfc02c7bac1a4644af68e
> 
> Nonetheless. All that are necessary parts to make threaded data streaming
> and analysis eventually merged into the mainline as joint Perf developers
> community effort.
> 
> > 
> > when reporting the threaded record data on really big servers,
> > you will run out of memory, so you need to read and flush all
> > the files together by smaller pieces
> 
> Yes, handling all that _big_ data after collection to make it
> helpful for analysis of performance issues is the other part
> of this story so that possible OOM should be somehow avoided.
> 
> > 
> > IMO we need to have this change before we allow threaded record
> 
> There are use cases of perf tool as a data provider, btw VTune is not
> the only one of them, and for those use cases threaded trace streaming
> lets its users get to their data that the users just were loosing before.
> This is huge difference and whole new level of support for such users.
> Post-process scripting around perf (e.g. Python based) will benefit
> from threaded trace streaming. Pipe mode can be extended to stream into
> open and passed fds using threads (e.g. perf record -o -fd:13,14,15,16).
> VTune-like tools can get performance data, load it into a (relational)
> DB files and provide analysis. And all that uses perf tool at its core.
> 
> I agree perf report OOM issue can exist on really-big servers but data 
> directories support for report mode for not-so-big servers and desktops
> is already enabled with this smaller change. Also really-big-servers
> come with really-big amount of memory and collection could possibly be
> limited to only interesting phases of execution so the issue could likely
> be avoided. At the same time threaded trace streaming could clarify on
> real use cases that are blocked by perf report OOM issue and that would
> clarify on exact required solution. So perf report OOM issue shouldn't
> be the showstopper for upstream of threaded trace streaming.

so the short answer is no, right? ;-) 

I understand all the excuses, but from my point of view we are
adding another pain point (and there's already few ;-) ) that
will make perf (even more) not user friendly

if we allow really friendly way to create huge data, we should
do our best to be able to process it as best as we can

jirka