From: Namhyung Kim <namhyung@kernel.org>
To: Joe Mario <jmario@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>,
Jiri Olsa <jolsa@kernel.org>,
Peter Zijlstra <peterz@infradead.org>,
Ian Rogers <irogers@google.com>,
Adrian Hunter <adrian.hunter@intel.com>,
Ingo Molnar <mingo@kernel.org>,
LKML <linux-kernel@vger.kernel.org>,
linux-perf-users@vger.kernel.org,
Linus Torvalds <torvalds@linux-foundation.org>,
Stephane Eranian <eranian@google.com>,
Masami Hiramatsu <mhiramat@kernel.org>,
linux-toolchains@vger.kernel.org,
linux-trace-devel@vger.kernel.org,
Ben Woodard <woodard@redhat.com>,
Kees Cook <keescook@chromium.org>,
David Blaikie <blaikie@google.com>, Xu Liu <xliuprof@google.com>,
Kan Liang <kan.liang@linux.intel.com>,
Ravi Bangoria <ravi.bangoria@amd.com>
Subject: Re: [RFC 00/48] perf tools: Introduce data type profiling (v1)
Date: Wed, 8 Nov 2023 20:48:50 -0800 [thread overview]
Message-ID: <CAM9d7ciWY6kvVB-JkrJS7BZs_WMy-OZRgwLvBG9WQ-Zi00czfA@mail.gmail.com> (raw)
In-Reply-To: <82cd8b7e-bd46-49ed-9160-eabcfd4c3c20@redhat.com>
Hello,
On Wed, Nov 8, 2023 at 9:12 AM Joe Mario <jmario@redhat.com> wrote:
>
> Hi Namhyung:
>
> I've been playing with your datatype profile patch and it looks really promising.
> I think it would be a big help if it could be integrated into perf c2c.
Great! Yeah, I think we can collaborate on it.
>
> Perf c2c gives a great insight into what's contributing to cpu cacheline contention, but it
> can be difficult to understand the output. Having visuals with your datatype profile output
> would be a big help.
Exactly.
>
> I have a simple test program with readers and writers tugging on the data below:
>
> uint64_t hotVar;
> typedef struct __foo {
> uint64_t m1;
> uint64_t m2;
> uint64_t m3;
> } FOO;
>
> The rest of this reply looks at both your datatype output and c2c to see where they
> might compliment each other.
>
>
> When I run perf with your patches on a simple program to cause contention on the above data, I get the following:
>
> # perf mem record --ldlat=1 --all-user -- ./tugtest -r3 -r5 -r7 -r9 -r11 -w10 -w12 -w14 -w16 -b5 -H2000000
> # perf report -s type,typeoff --hierarchy --stdio
>
> # Samples: 26K of event 'cpu/mem-loads,ldlat=1/P'
> # Event count (approx.): 2958226
> #
> # Overhead Data Type / Data Type Offset
> # ........... ............................
> #
> 54.50% int
> 54.50% int +0 (no field)
> 23.21% long int
> 23.21% long int +0 (no field)
> 18.30% struct __foo
> 9.57% struct __foo +8 (m2)
> 8.73% struct __foo +0 (m1)
> 3.86% long unsigned int
> 3.86% long unsigned int +0 (no field)
> <snip>
>
> # Samples: 30K of event 'cpu/mem-stores/P'
> # Event count (approx.): 33880197
> #
> # Overhead Data Type / Data Type Offset
> # ........... ............................
> #
> 99.85% struct __foo
> 70.48% struct __foo +0 (m1)
> 29.34% struct __foo +16 (m3)
> 0.03% struct __foo +8 (m2)
> 0.09% long unsigned int
> 0.09% long unsigned int +0 (no field)
> 0.06% (unknown)
> 0.06% (unknown) +0 (no field)
> <snip>
>
> Then I run perf annotate with your patches, and I get the following:
>
> # perf annotate --data-type
>
> Annotate type: 'long int' in /home/joe/tugtest/tugtest (2901 samples):
> ============================================================================
> samples offset size field
> 2901 0 8 long int ;
>
> Annotate type: 'struct __foo' in /home/joe/tugtest/tugtest (5593 samples):
> ============================================================================
> samples offset size field
> 5593 0 24 struct __foo {
> 2755 0 8 uint64_t m1;
> 2838 8 8 uint64_t m2;
> 0 16 8 uint64_t m3;
> };
>
> Now when I run that same simple test using perf c2c, and I focus on the cachline that the struct and hotVar reside in, I get:
>
> # perf c2c record --all-user -- ./tugtest -r3 -r5 -r7 -r9 -r11 -w10 -w12 -w14 -w16 -b5 -H2000000
> # perf c2c report -NNN --stdio
> # <snip>
> #
> # ----- HITM ----- ------- Store Refs ------ ------ Data address ------ ---------- cycles ---------- Total cpu Shared
> # Num RmtHitm LclHitm L1 Hit L1 Miss N/A Offset Node PA cnt Code address rmt hitm lcl hitm load records cnt Symbol Object Source:Line Node{cpu list}
> #.... ....... ....... ....... ....... ....... ............ .... ...... ............ ........ ........ ........ ....... ..... .......... ....... ............. ....
> #
> ---------------------------------------------------------------
> 0 1094 2008 17071 13762 0 0x406100
> ---------------------------------------------------------------
> 0.00% 0.20% 0.00% 0.00% 0.00% 0x8 1 1 0x401355 0 978 1020 2962 4 [.] writer tugtest tugtest.c:129 0{10,12,14,16}
> 0.00% 0.00% 0.12% 0.02% 0.00% 0x8 1 1 0x401360 0 0 0 23 4 [.] writer tugtest tugtest.c:129 0{10,12,14,16}
> 68.10% 60.26% 0.00% 0.00% 0.00% 0x10 1 1 0x401505 2181 1541 1393 5813 5 [.] reader tugtest tugtest.c:163 1{3,5,7,9,11}
> 31.63% 39.34% 0.00% 0.00% 0.00% 0x10 1 1 0x401331 1242 1095 936 3393 4 [.] writer tugtest tugtest.c:127 0{10,12,14,16}
> 0.00% 0.00% 40.03% 40.25% 0.00% 0x10 1 1 0x40133c 0 0 0 12372 4 [.] writer tugtest tugtest.c:127 0{10,12,14,16}
> 0.27% 0.15% 0.00% 0.00% 0.00% 0x18 1 1 0x401343 834 1136 1032 2930 4 [.] writer tugtest tugtest.c:128 0{10,12,14,16}
> 0.00% 0.05% 0.00% 0.00% 0.00% 0x18 1 1 0x40150c 0 933 1567 5050 5 [.] reader tugtest tugtest.c:163 1{3,5,7,9,11}
> 0.00% 0.00% 0.06% 0.00% 0.00% 0x18 1 1 0x40134e 0 0 0 10 4 [.] writer tugtest tugtest.c:128 0{10,12,14,16}
> 0.00% 0.00% 59.80% 59.73% 0.00% 0x20 1 1 0x401516 0 0 0 18428 5 [.] reader tugtest tugtest.c:163 1{3,5,7,9,11}
>
> With the above c2c output, we can see:
> - the hottest contended addresses, and the load latencies they caused.
> - the cacheline offset for the contended addresses.
> - the cpus and numa nodes where the accesses came from.
> - the cacheline alignment for the data of interest.
> - the number of cpus and threads concurrently accessing each address.
> - the breakdown of reads causing HITM (contention) and writes hitting or missing the cacheline.
> - the object name, source line and line number for where the accesses occured.
> - the numa node where the data is allocated.
> - the number of physical pages the virtual addresses were mapped to (e.g. numa_balancing).
>
> What would really help the c2c output be more usable is if it had a better visual to it.
> It's likely the current c2c output can be trimmed a bit.
>
> Here's one idea that incorporates your datatype info, though I'm sure there are better ways, as this may get unwieldy.:
>
> # ----- HITM ----- ------- Store Refs ------ ------ Data address ------ ---------- cycles ---------- Total cpu Shared
> # Num RmtHitm LclHitm L1 Hit L1 Miss N/A Offset Node PA cnt Code address rmt hitm lcl hitm load records cnt Symbol Object Source:Line Node{cpu list}
> #.... ....... ....... ....... ....... ....... ............ .... ...... ............ ........ ........ ........ ....... ..... .......... ....... ............. ....
> #
> ---------------------------------------------------------------
> 0 1094 2008 17071 13762 0 0x406100
> ---------------------------------------------------------------
> uint64_t hotVar: tugtest.c:38
> 0.00% 0.20% 0.00% 0.00% 0.00% 0x8 1 1 0x401355 0 978 1020 2962 4 [.] writer tugtest tugtest.c:129 0{10,12,14,16}
> 0.00% 0.00% 0.12% 0.02% 0.00% 0x8 1 1 0x401360 0 0 0 23 4 [.] writer tugtest tugtest.c:129 0{10,12,14,16}
> struct __foo uint64_t m1: tugtest.c:39
> 68.10% 60.26% 0.00% 0.00% 0.00% 0x10 1 1 0x401505 2181 1541 1393 5813 5 [.] reader tugtest tugtest.c:163 1{3,5,7,9,11}
> 31.63% 39.34% 0.00% 0.00% 0.00% 0x10 1 1 0x401331 1242 1095 936 3393 4 [.] writer tugtest tugtest.c:127 0{10,12,14,16}
> 0.00% 0.00% 40.03% 40.25% 0.00% 0x10 1 1 0x40133c 0 0 0 12372 4 [.] writer tugtest tugtest.c:127 0{10,12,14,16}
> struct __foo uint64_t m2: tugtest.c:40
> 0.27% 0.15% 0.00% 0.00% 0.00% 0x18 1 1 0x401343 834 1136 1032 2930 4 [.] writer tugtest tugtest.c:128 0{10,12,14,16}
> 0.00% 0.05% 0.00% 0.00% 0.00% 0x18 1 1 0x40150c 0 933 1567 5050 5 [.] reader tugtest tugtest.c:163 1{3,5,7,9,11}
> 0.00% 0.00% 0.06% 0.00% 0.00% 0x18 1 1 0x40134e 0 0 0 10 4 [.] writer tugtest tugtest.c:128 0{10,12,14,16}
> struct __foo uint64_t m3: tugtest.c:41
> 0.00% 0.00% 59.80% 59.73% 0.00% 0x20 1 1 0x401516 0 0 0 18428 5 [.] reader tugtest tugtest.c:163 1{3,5,7,9,11}
>
> And then it would be good to find a clean way to incorporate your sample counts.
I'm not sure we can get the exact source line for the data type/fields.
Of course, we can aggregate the results for each field. Actually you
can use `perf report -s type,typeoff,symoff --hierarchy` for something
similar. :)
>
> On a related note, is there a way the accesses could be broken down into read counts
> and write counts? That, with the above source line info for all the accesses,
> helps to convey a picture of "the affinity of the accesses".
Sure, perf report already supports showing events in a group
together. You can use --group option to force grouping
individual events. perf annotate with --data-type doesn't have
that yet. I'll update it in v2.
>
> For example, while it's normally good to separate read-mostly data from hot
> written data, if the reads and writes are done together in the same block of
> code by the same thread, then keeping the two data symbols in the same cacheline
> could be a win. I've seen this often. Your datatype info might be able to
> make these affinities more visible to the user.
>
> Thanks for doing this. This is great.
> Joe
Thanks for your feedback!
Namhyung
prev parent reply other threads:[~2023-11-09 4:49 UTC|newest]
Thread overview: 96+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-10-12 3:50 [RFC 00/48] perf tools: Introduce data type profiling (v1) Namhyung Kim
2023-10-12 3:50 ` [PATCH 01/48] perf annotate: Move raw_comment and raw_func_start Namhyung Kim
2023-10-12 3:50 ` [PATCH 02/48] perf annotate: Check if operand has multiple regs Namhyung Kim
2023-11-27 19:05 ` Arnaldo Carvalho de Melo
2023-10-12 3:50 ` [PATCH 03/48] perf tools: Add util/debuginfo.[ch] files Namhyung Kim
2023-10-12 3:50 ` [PATCH 04/48] perf dwarf-aux: Fix die_get_typename() for void * Namhyung Kim
2023-11-04 10:52 ` Masami Hiramatsu
2023-10-12 3:50 ` [PATCH 05/48] perf dwarf-aux: Move #ifdef code to the header file Namhyung Kim
2023-11-04 10:59 ` Masami Hiramatsu
2023-10-12 3:50 ` [PATCH 06/48] perf dwarf-aux: Add die_get_scopes() helper Namhyung Kim
2023-11-05 9:50 ` Masami Hiramatsu
2023-10-12 3:50 ` [PATCH 07/48] perf dwarf-aux: Add die_find_variable_by_reg() helper Namhyung Kim
2023-11-05 9:48 ` Masami Hiramatsu
2023-10-12 3:50 ` [PATCH 08/48] perf dwarf-aux: Factor out __die_get_typename() Namhyung Kim
2023-11-05 9:07 ` Masami Hiramatsu
2023-11-06 4:01 ` Namhyung Kim
2023-10-12 3:50 ` [PATCH 09/48] perf dwarf-regs: Add get_dwarf_regnum() Namhyung Kim
2023-11-05 8:36 ` Masami Hiramatsu
2023-11-06 4:12 ` Namhyung Kim
2023-10-12 3:50 ` [PATCH 10/48] perf annotate-data: Add find_data_type() Namhyung Kim
2023-10-12 3:50 ` [PATCH 11/48] perf annotate-data: Add dso->data_types tree Namhyung Kim
2023-10-12 3:50 ` [PATCH 12/48] perf annotate: Factor out evsel__get_arch() Namhyung Kim
2023-10-12 3:50 ` [PATCH 13/48] perf annotate: Add annotate_get_insn_location() Namhyung Kim
2023-10-23 16:38 ` Arnaldo Carvalho de Melo
2023-10-24 19:10 ` Namhyung Kim
2023-10-26 5:26 ` Namhyung Kim
2023-10-26 19:37 ` Arnaldo Carvalho de Melo
2023-10-12 3:50 ` [PATCH 14/48] perf annotate: Implement hist_entry__get_data_type() Namhyung Kim
2023-10-12 3:50 ` [PATCH 15/48] perf report: Add 'type' sort key Namhyung Kim
2023-10-23 16:53 ` Arnaldo Carvalho de Melo
2023-10-24 19:11 ` Namhyung Kim
2023-10-12 3:50 ` [PATCH 16/48] perf report: Support data type profiling Namhyung Kim
2023-10-12 3:50 ` [PATCH 17/48] perf annotate-data: Add member field in the data type Namhyung Kim
2023-10-12 3:50 ` [PATCH 18/48] perf annotate-data: Update sample histogram for type Namhyung Kim
2023-10-12 3:50 ` [PATCH 19/48] perf report: Add 'typeoff' sort key Namhyung Kim
2023-10-12 3:50 ` [PATCH 20/48] perf report: Add 'symoff' " Namhyung Kim
2023-10-12 3:50 ` [PATCH 21/48] perf annotate: Add --data-type option Namhyung Kim
2023-10-12 3:50 ` [PATCH 22/48] perf annotate: Add --type-stat option for debugging Namhyung Kim
2023-10-23 17:28 ` Arnaldo Carvalho de Melo
2023-10-23 17:40 ` Arnaldo Carvalho de Melo
2023-10-24 19:12 ` Namhyung Kim
2023-10-12 3:50 ` [PATCH 23/48] perf annotate: Add --insn-stat " Namhyung Kim
2023-10-12 3:50 ` [PATCH 24/48] perf annotate-data: Parse 'lock' prefix from llvm-objdump Namhyung Kim
2023-10-12 3:50 ` [PATCH 25/48] perf annotate-data: Handle macro fusion on x86 Namhyung Kim
2023-10-12 3:50 ` [PATCH 26/48] perf annotate-data: Handle array style accesses Namhyung Kim
2023-10-12 3:50 ` [PATCH 27/48] perf annotate-data: Add stack operation pseudo type Namhyung Kim
2023-10-12 3:50 ` [PATCH 28/48] perf dwarf-aux: Add die_find_variable_by_addr() Namhyung Kim
2023-11-06 15:25 ` Masami Hiramatsu
2023-11-09 5:36 ` Namhyung Kim
2023-10-12 3:50 ` [PATCH 29/48] perf annotate-data: Handle PC-relative addressing Namhyung Kim
2023-10-12 3:50 ` [PATCH 30/48] perf annotate-data: Support global variables Namhyung Kim
2023-10-12 3:50 ` [PATCH 31/48] perf dwarf-aux: Add die_get_cfa() Namhyung Kim
2023-11-07 0:50 ` Masami Hiramatsu
2023-11-08 5:28 ` Namhyung Kim
2023-10-12 3:50 ` [PATCH 32/48] perf annotate-data: Support stack variables Namhyung Kim
2023-10-12 3:50 ` [PATCH 33/48] perf dwarf-aux: Check allowed DWARF Ops Namhyung Kim
2023-11-07 9:32 ` Masami Hiramatsu
2023-11-08 5:34 ` Namhyung Kim
2023-10-12 3:50 ` [PATCH 34/48] perf dwarf-aux: Add die_collect_vars() Namhyung Kim
2023-11-08 10:52 ` Masami Hiramatsu
2023-11-09 5:05 ` Namhyung Kim
2023-10-12 3:50 ` [PATCH 35/48] perf dwarf-aux: Handle type transfer for memory access Namhyung Kim
2023-11-08 10:57 ` Masami Hiramatsu
2023-10-12 3:50 ` [PATCH 36/48] perf annotate-data: Introduce struct data_loc_info Namhyung Kim
2023-12-03 16:22 ` Athira Rajeev
2023-12-05 0:10 ` Namhyung Kim
2023-12-05 7:17 ` Athira Rajeev
2023-10-12 3:51 ` [PATCH 37/48] perf map: Add map__objdump_2rip() Namhyung Kim
2023-10-12 3:51 ` [PATCH 38/48] perf annotate: Add annotate_get_basic_blocks() Namhyung Kim
2023-10-12 3:51 ` [PATCH 39/48] perf annotate-data: Maintain variable type info Namhyung Kim
2023-10-12 3:51 ` [PATCH 40/48] perf annotate-data: Add update_insn_state() Namhyung Kim
2023-10-12 3:51 ` [PATCH 41/48] perf annotate-data: Handle global variable access Namhyung Kim
2023-10-12 3:51 ` [PATCH 42/48] perf annotate-data: Handle call instructions Namhyung Kim
2023-10-12 3:51 ` [PATCH 43/48] perf annotate-data: Implement instruction tracking Namhyung Kim
2023-10-12 3:51 ` [PATCH 44/48] perf annotate: Parse x86 segment register location Namhyung Kim
2023-10-12 3:51 ` [PATCH 45/48] perf annotate-data: Handle this-cpu variables in kernel Namhyung Kim
2023-10-12 3:51 ` [PATCH 46/48] perf annotate-data: Track instructions with a this-cpu variable Namhyung Kim
2023-10-12 3:51 ` [PATCH 47/48] perf annotate-data: Add stack canary type Namhyung Kim
2023-10-12 3:51 ` [PATCH 48/48] perf annotate-data: Add debug message Namhyung Kim
2023-10-12 6:03 ` [RFC 00/48] perf tools: Introduce data type profiling (v1) Ingo Molnar
2023-10-12 16:19 ` Namhyung Kim
2023-10-12 18:33 ` Ingo Molnar
2023-10-12 20:45 ` Namhyung Kim
2023-10-12 9:11 ` Peter Zijlstra
2023-10-12 16:41 ` Namhyung Kim
[not found] ` <CADzB+2mu98v9EUsA1Y-wVDSrXT2kznKi87Tb6QdN5y4mMFNsyg@mail.gmail.com>
2023-10-25 5:58 ` Namhyung Kim
2023-10-12 9:15 ` Peter Zijlstra
2023-10-12 16:52 ` Namhyung Kim
2023-10-13 14:15 ` Arnaldo Carvalho de Melo
2023-10-23 21:58 ` Andi Kleen
2023-10-24 19:16 ` Namhyung Kim
2023-10-25 2:09 ` Andi Kleen
2023-10-25 5:51 ` Namhyung Kim
2023-10-25 20:01 ` Andi Kleen
2023-11-08 17:12 ` Joe Mario
2023-11-09 4:48 ` Namhyung Kim [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CAM9d7ciWY6kvVB-JkrJS7BZs_WMy-OZRgwLvBG9WQ-Zi00czfA@mail.gmail.com \
--to=namhyung@kernel.org \
--cc=acme@kernel.org \
--cc=adrian.hunter@intel.com \
--cc=blaikie@google.com \
--cc=eranian@google.com \
--cc=irogers@google.com \
--cc=jmario@redhat.com \
--cc=jolsa@kernel.org \
--cc=kan.liang@linux.intel.com \
--cc=keescook@chromium.org \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-perf-users@vger.kernel.org \
--cc=linux-toolchains@vger.kernel.org \
--cc=linux-trace-devel@vger.kernel.org \
--cc=mhiramat@kernel.org \
--cc=mingo@kernel.org \
--cc=peterz@infradead.org \
--cc=ravi.bangoria@amd.com \
--cc=torvalds@linux-foundation.org \
--cc=woodard@redhat.com \
--cc=xliuprof@google.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).