Received-SPF: pass (google.com: domain of linux-kernel+bounces-90943-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
Precedence: bulk
MIME-Version: 1.0
References: <20240228005230.287113-1-namhyung@kernel.org> <ZeXTmdhu3Y_gC9ma@x1>
In-Reply-To: <ZeXTmdhu3Y_gC9ma@x1>
From: Namhyung Kim <namhyung@kernel.org>
Date: Mon, 4 Mar 2024 08:36:30 -0800
Message-ID: <CAM9d7chEoab71v8=th7NAWcsPPZV+oGiJPGfiCcjdS3L+u13dA@mail.gmail.com>
Subject: Re: [PATCH 0/4] perf annotate: Improve memory usage for symbol histogram
To: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Ian Rogers <irogers@google.com>, Jiri Olsa <jolsa@kernel.org>, 
	Adrian Hunter <adrian.hunter@intel.com>, Peter Zijlstra <peterz@infradead.org>, 
	Ingo Molnar <mingo@kernel.org>, LKML <linux-kernel@vger.kernel.org>, 
	linux-perf-users@vger.kernel.org, Andi Kleen <ak@linux.intel.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

Hi Arnaldo,

On Mon, Mar 4, 2024 at 5:58=E2=80=AFAM Arnaldo Carvalho de Melo <acme@kerne=
l.org> wrote:
>
> On Tue, Feb 27, 2024 at 04:52:26PM -0800, Namhyung Kim wrote:
> > This is another series of memory optimization in perf annotate.
>
> > When perf annotate (or perf report/top with TUI) processes samples, it
> > needs to save the sample period (overhead) at instruction level.  For
> > now, it allocates an array to do that for the whole symbol when it
> > hits any new symbol.  This comes with a lot of waste since samples can
> > be very few and instructions span to multiple bytes.
>
> > For example, when a sample hits symbol 'foo' that has size of 100 and
> > that's the only sample falls into the symbol.  Then it needs to
> > allocate a symbol histogram (sym_hist) and the its size would be
>
> >   16 (header) + 16 (sym_hist_entry) * 100 (symbol_size) =3D 1616
>
> > But actually it just needs 32 (header + sym_hist_entry) bytes.  Things
> > get worse if the symbol size is bigger (and it doesn't have many
> > samples in different places).  Also note that it needs separate
> > histogram for each event.
>
> > Let's split the sym_hist_entry and have it in a hash table so that it
> > can allocate only necessary entries.
>
> > No functional change intended.
>
> I tried this before/after this series:
>
>   $ time perf annotate --stdio2 -i perf.data.annotate
>
> For:
>
>   perf record -e '{cycles,instructions,cache-misses}' make -k CORESIGHT=
=3D1 O=3D/tmp/build/$(basename $PWD)/ -C tools/perf install-bin
>
> And found these odd cases:
>
>   $ diff -u before after
>   --- before    2024-02-28 15:38:25.086062812 -0300
>   +++ after     2024-02-29 14:12:05.606652725 -0300
>   @@ -2450826,7 +2450826,7 @@
>                                 =E2=86=93 je       1c62
>                                 =E2=86=92 call     operator delete(void*)=
@plt
>                                 { return _M_dataplus._M_p; }
>   -                       1c62:   mov      0x13c0(%rsp),%rdi
>   +  0.00   0.00 100.00   1c62:   mov      0x13c0(%rsp),%rdi
>                                 if (_M_data() =3D=3D _M_local_data())
>                                   lea      0x13d0(%rsp),%rax
>                                   cmp      %rax,%rdi
>   @@ -2470648,7 +2470648,7 @@
>                                   mov      %rbx,%rdi
>                                 =E2=86=92 call     operator delete(void*)=
@plt
>                                 using reference =3D T &;
>   -  0.00   0.00 100.00  11c65:   mov      0x8(%r12),%rax
>   +                      11c65:   mov      0x8(%r12),%rax
>                                 size_t size() const { return Size; }
>                                   mov      0x10(%r12),%ecx
>                                   mov      %rax,%rbp
>   $
>
>
> This is a large function:

Thanks for the test!  I think it missed the cast to 64-bit somewhere.
I'll check and send v2 soon.

Thanks,
Namhyung


>
> Samples: 574K of events 'anon group { cpu_core/cycles/u, cpu_core/instruc=
tions/u, cpu_core/cache-misses/u }', 4000 Hz, Event count (approx.): 614695=
751751, [percent: local period]$
> clang::CompilerInvocation::ParseCodeGenArgs(clang::CodeGenOptions&, llvm:=
:opt::ArgList&, clang::InputKind, clang::DiagnosticsEngine&, llvm::Triple c=
onst&, std::__cxx11::basic_string<char, std
> Percent
>
> Probably when building the BPF skels in tools/perf/
>
> - Arnaldo