From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from smtp.kernel.org (aws-us-west-2-korg-mail-1.web.codeaurora.org [10.30.226.201]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.subspace.kernel.org (Postfix) with ESMTPS id 97FD618858A; Tue, 4 Feb 2025 04:27:51 +0000 (UTC) Authentication-Results: smtp.subspace.kernel.org; arc=none smtp.client-ip=10.30.226.201 ARC-Seal:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1738643271; cv=none; b=K8Dd0HVy7ZmQ8MduJV0RSSKe4+SKz/SgI033ZN7SdVZIRPo6C0t1NurebfbhYhTo056FKWcrmkjWqFn4Ql1FLkYU09DekLplMpri8ZzrX2C6nJ+Dewr9OBq/zOE2vUzy5iypx49Anv8o3Tc6JJizGICxPi/58+UGD6iPO3EQppE= ARC-Message-Signature:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1738643271; c=relaxed/simple; bh=U/EzLb92gORbHR+90FsDZyMbFuAxD8VZ4ZDPbBb34VA=; h=Date:From:To:Cc:Subject:Message-ID:References:MIME-Version: Content-Type:Content-Disposition:In-Reply-To; b=YU8+KhYeVwpLzvpnR+MI7+2h4kLCBljNYHfv/osynW6prGHqsPcvGxRUq55JX1ocNdfq4RWV0Zy1N0UT46fKt2alXtw5eISWush4BQhyn56BNea1bNQiaMA82FdHCAiK04a+uYCR8iscFfbu495K86JK+9vo1O2NgmOIn2jxh1U= ARC-Authentication-Results:i=1; smtp.subspace.kernel.org; dkim=pass (2048-bit key) header.d=kernel.org header.i=@kernel.org header.b=NXLSx0ba; arc=none smtp.client-ip=10.30.226.201 Authentication-Results: smtp.subspace.kernel.org; dkim=pass (2048-bit key) header.d=kernel.org header.i=@kernel.org header.b="NXLSx0ba" Received: by smtp.kernel.org (Postfix) with ESMTPSA id BDB92C4CEDF; Tue, 4 Feb 2025 04:27:50 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=kernel.org; s=k20201202; t=1738643271; bh=U/EzLb92gORbHR+90FsDZyMbFuAxD8VZ4ZDPbBb34VA=; h=Date:From:To:Cc:Subject:References:In-Reply-To:From; b=NXLSx0baNRTpaFMrnJvAqnbs8yGNJ7HNpYb1lwD6TYhY6dXVPQYhGzio+xv7r80yj B5z8TGpCQvAWegcU1J8qML1ZDg8rpQTvrsjJfX4JfUh7eg8sUgQwY7TyCUTylT8N2f lZ/xKi2wEzMLtQu7vIU9D562asUlZIln/0bY4fHObk8haj0rmEnUgWYhIYEdLxxVH8 pX96UXU6ftwAtg8owKHBTEv+WPqZedCgyELGV1m53ghnl36A87Yr/Ffm3w7ezuTFr2 TsaC0Zb+/a66R2GFEnnokHRRvkyM0//o8/gPnmehcqYOBKRG+4XxZEHOiIKhuSdUCk /4xadKAFRqmjA== Date: Mon, 3 Feb 2025 20:27:49 -0800 From: Namhyung Kim To: Dmitry Vyukov Cc: irogers@google.com, acme@kernel.org, linux-perf-users@vger.kernel.org, linux-kernel@vger.kernel.org Subject: Re: [PATCH v3 7/8] perf report: Add latency and parallelism profiling documentation Message-ID: References: <4b49d5dd316f5d7119d8b16d5d70f13574158dc6.1738592865.git.dvyukov@google.com> Precedence: bulk X-Mailing-List: linux-kernel@vger.kernel.org List-Id: List-Subscribe: List-Unsubscribe: MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Disposition: inline In-Reply-To: <4b49d5dd316f5d7119d8b16d5d70f13574158dc6.1738592865.git.dvyukov@google.com> On Mon, Feb 03, 2025 at 03:30:42PM +0100, Dmitry Vyukov wrote: > Describe latency and parallelism profiling, related flags, and differences > with the currently only supported CPU-consumption-centric profiling. It doesn't seem to have descriptions for the --latency option (for perf record and report). Probably better to put them in the previous patch. > > Signed-off-by: Dmitry Vyukov > Cc: Namhyung Kim > Cc: Arnaldo Carvalho de Melo > Cc: Ian Rogers > Cc: linux-perf-users@vger.kernel.org > Cc: linux-kernel@vger.kernel.org > --- > .../callchain-overhead-calculation.txt | 5 +- > .../cpu-and-latency-overheads.txt | 85 +++++++++++++++++++ > tools/perf/Documentation/perf-report.txt | 49 +++++++---- > tools/perf/Documentation/tips.txt | 3 + > 4 files changed, 123 insertions(+), 19 deletions(-) > > diff --git a/tools/perf/Documentation/callchain-overhead-calculation.txt b/tools/perf/Documentation/callchain-overhead-calculation.txt > index 1a757927195ed..e0202bf5bd1a0 100644 > --- a/tools/perf/Documentation/callchain-overhead-calculation.txt > +++ b/tools/perf/Documentation/callchain-overhead-calculation.txt > @@ -1,7 +1,8 @@ > Overhead calculation > -------------------- > -The overhead can be shown in two columns as 'Children' and 'Self' when > -perf collects callchains. The 'self' overhead is simply calculated by > +The CPU overhead can be shown in two columns as 'Children' and 'Self' > +when perf collects callchains (and corresponding 'Wall' columns for > +wall-clock overhead). The 'self' overhead is simply calculated by > adding all period values of the entry - usually a function (symbol). > This is the value that perf shows traditionally and sum of all the > 'self' overhead values should be 100%. > diff --git a/tools/perf/Documentation/cpu-and-latency-overheads.txt b/tools/perf/Documentation/cpu-and-latency-overheads.txt > new file mode 100644 > index 0000000000000..3b6d637054651 > --- /dev/null > +++ b/tools/perf/Documentation/cpu-and-latency-overheads.txt > @@ -0,0 +1,85 @@ > +CPU and latency overheads > +------------------------- > +There are two notions of time: wall-clock time and CPU time. > +For a single-threaded program, or a program running on a single-core machine, > +these notions are the same. However, for a multi-threaded/multi-process program > +running on a multi-core machine, these notions are significantly different. > +Each second of wall-clock time we have number-of-cores seconds of CPU time. > +Perf can measure overhead for both of these times (shown in 'overhead' and > +'latency' columns for CPU and wall-clock time correspondingly). > + > +Optimizing CPU overhead is useful to improve 'throughput', while optimizing > +latency overhead is useful to improve 'latency'. It's important to understand > +which one is useful in a concrete situation at hand. For example, the former > +may be useful to improve max throughput of a CI build server that runs on 100% > +CPU utilization, while the latter may be useful to improve user-perceived > +latency of a single interactive program build. > +These overheads may be significantly different in some cases. For example, > +consider a program that executes function 'foo' for 9 seconds with 1 thread, > +and then executes function 'bar' for 1 second with 128 threads (consumes > +128 seconds of CPU time). The CPU overhead is: 'foo' - 6.6%, 'bar' - 93.4%. > +While the latency overhead is: 'foo' - 90%, 'bar' - 10%. If we try to optimize > +running time of the program looking at the (wrong in this case) CPU overhead, > +we would concentrate on the function 'bar', but it can yield only 10% running > +time improvement at best. > + > +By default, perf shows only CPU overhead. To show latency overhead, use > +'perf record --latency' and 'perf report': > + > +----------------------------------- > +Overhead Latency Command > + 93.88% 25.79% cc1 > + 1.90% 39.87% gzip > + 0.99% 10.16% dpkg-deb > + 0.57% 1.00% as > + 0.40% 0.46% sh > +----------------------------------- > + > +To sort by latency overhead, use 'perf report --latency': > + > +----------------------------------- > +Latency Overhead Command > + 39.87% 1.90% gzip > + 25.79% 93.88% cc1 > + 10.16% 0.99% dpkg-deb > + 4.17% 0.29% git > + 2.81% 0.11% objtool > +----------------------------------- > + > +To get insight into the difference between the overheads, you may check > +parallelization histogram with '--sort=latency,parallelism,comm,symbol --hierarchy' I think you need to omit latency in the sort key and recommend users to use --latency option instead. 'perf report --hierarchy --latency --sort=parallelism,comm,symbol' > +flags. It shows fraction of (wall-clock) time the workload utilizes different > +numbers of cores ('Parallelism' column). For example, in the following case > +the workload utilizes only 1 core most of the time, but also has some > +highly-parallel phases, which explains significant difference between > +CPU and wall-clock overheads: > + > +----------------------------------- > + Latency Overhead Parallelism / Command / Symbol > ++ 56.98% 2.29% 1 > ++ 16.94% 1.36% 2 > ++ 4.00% 20.13% 125 > ++ 3.66% 18.25% 124 > ++ 3.48% 17.66% 126 > ++ 3.26% 0.39% 3 > ++ 2.61% 12.93% 123 > +----------------------------------- > + > +By expanding corresponding lines, you may see what commands/functions run > +at the given parallelism level: > + > +----------------------------------- > + Latency Overhead Parallelism / Command / Symbol > +- 56.98% 2.29% 1 > + 32.80% 1.32% gzip > + 4.46% 0.18% cc1 > + 2.81% 0.11% objtool > + 2.43% 0.10% dpkg-source > + 2.22% 0.09% ld > + 2.10% 0.08% dpkg-genchanges > +----------------------------------- > + > +To see the normal function-level profile for particular parallelism levels > +(number of threads actively running on CPUs), you may use '--parallelism' > +filter. For example, to see the profile only for low parallelism phases > +of a workload use '--latency --parallelism=1-2' flags. > diff --git a/tools/perf/Documentation/perf-report.txt b/tools/perf/Documentation/perf-report.txt > index 87f8645194062..7e0ba990d71e8 100644 > --- a/tools/perf/Documentation/perf-report.txt > +++ b/tools/perf/Documentation/perf-report.txt > @@ -44,7 +44,7 @@ OPTIONS > --comms=:: > Only consider symbols in these comms. CSV that understands > file://filename entries. This option will affect the percentage of > - the overhead column. See --percentage for more info. > + the overhead and latency columns. See --percentage for more info. > --pid=:: > Only show events for given process ID (comma separated list). > > @@ -54,12 +54,12 @@ OPTIONS > --dsos=:: > Only consider symbols in these dsos. CSV that understands > file://filename entries. This option will affect the percentage of > - the overhead column. See --percentage for more info. > + the overhead and latency columns. See --percentage for more info. > -S:: > --symbols=:: > Only consider these symbols. CSV that understands > file://filename entries. This option will affect the percentage of > - the overhead column. See --percentage for more info. > + the overhead and latency columns. See --percentage for more info. > > --symbol-filter=:: > Only show symbols that match (partially) with this filter. > @@ -68,6 +68,16 @@ OPTIONS > --hide-unresolved:: > Only display entries resolved to a symbol. > > +--parallelism:: > + Only consider these parallelism levels. Parallelism level is the number > + of threads that actively run on CPUs at the time of sample. The flag > + accepts single number, comma-separated list, and ranges (for example: > + "1", "7,8", "1,64-128"). This is useful in understanding what a program > + is doing during sequential/low-parallelism phases as compared to > + high-parallelism phases. This option will affect the percentage of > + the overhead and latency columns. See --percentage for more info. > + Also see the `CPU and latency overheads' section for more details. > + > -s:: > --sort=:: > Sort histogram entries by given key(s) - multiple keys can be specified > @@ -87,6 +97,7 @@ OPTIONS > entries are displayed as "[other]". > - cpu: cpu number the task ran at the time of sample > - socket: processor socket number the task ran at the time of sample > + - parallelism: number of running threads at the time of sample > - srcline: filename and line number executed at the time of sample. The > DWARF debugging info must be provided. > - srcfile: file name of the source file of the samples. Requires dwarf > @@ -97,12 +108,14 @@ OPTIONS > - cgroup_id: ID derived from cgroup namespace device and inode numbers. > - cgroup: cgroup pathname in the cgroupfs. > - transaction: Transaction abort flags. > - - overhead: Overhead percentage of sample > - - overhead_sys: Overhead percentage of sample running in system mode > - - overhead_us: Overhead percentage of sample running in user mode > - - overhead_guest_sys: Overhead percentage of sample running in system mode > + - overhead: CPU overhead percentage of sample. > + - latency: latency (wall-clock) overhead percentage of sample. > + See the `CPU and latency overheads' section for more details. > + - overhead_sys: CPU overhead percentage of sample running in system mode > + - overhead_us: CPU overhead percentage of sample running in user mode > + - overhead_guest_sys: CPU overhead percentage of sample running in system mode > on guest machine > - - overhead_guest_us: Overhead percentage of sample running in user mode on > + - overhead_guest_us: CPU overhead percentage of sample running in user mode on > guest machine > - sample: Number of sample > - period: Raw number of event count of sample > @@ -125,8 +138,8 @@ OPTIONS > - weight2: Average value of event specific weight (2nd field of weight_struct). > - weight3: Average value of event specific weight (3rd field of weight_struct). > > - By default, comm, dso and symbol keys are used. > - (i.e. --sort comm,dso,symbol) > + By default, overhead, comm, dso and symbol keys are used. > + (i.e. --sort overhead,comm,dso,symbol). > > If --branch-stack option is used, following sort keys are also > available: > @@ -201,9 +214,9 @@ OPTIONS > --fields=:: > Specify output field - multiple keys can be specified in CSV format. > Following fields are available: > - overhead, overhead_sys, overhead_us, overhead_children, sample, period, > - weight1, weight2, weight3, ins_lat, p_stage_cyc and retire_lat. The > - last 3 names are alias for the corresponding weights. When the weight > + overhead, latency, overhead_sys, overhead_us, overhead_children, sample, > + period, weight1, weight2, weight3, ins_lat, p_stage_cyc and retire_lat. > + The last 3 names are alias for the corresponding weights. When the weight > fields are used, they will show the average value of the weight. > > Also it can contain any sort key(s). > @@ -289,7 +302,7 @@ OPTIONS > Accumulate callchain of children to parent entry so that then can > show up in the output. The output will have a new "Children" column > and will be sorted on the data. It requires callchains are recorded. > - See the `overhead calculation' section for more details. Enabled by > + See the `Overhead calculation' section for more details. Enabled by > default, disable with --no-children. > > --max-stack:: > @@ -442,9 +455,9 @@ OPTIONS > --call-graph option for details. > > --percentage:: > - Determine how to display the overhead percentage of filtered entries. > - Filters can be applied by --comms, --dsos and/or --symbols options and > - Zoom operations on the TUI (thread, dso, etc). > + Determine how to display the CPU and latency overhead percentage > + of filtered entries. Filters can be applied by --comms, --dsos, --symbols > + and/or --parallelism options and Zoom operations on the TUI (thread, dso, etc). > > "relative" means it's relative to filtered entries only so that the > sum of shown entries will be always 100%. "absolute" means it retains > @@ -627,6 +640,8 @@ include::itrace.txt[] > --skip-empty:: > Do not print 0 results in the --stat output. > > +include::cpu-and-latency-overheads.txt[] > + > include::callchain-overhead-calculation.txt[] > > SEE ALSO > diff --git a/tools/perf/Documentation/tips.txt b/tools/perf/Documentation/tips.txt > index 67b326ba00407..f6f71e70ff2cb 100644 > --- a/tools/perf/Documentation/tips.txt > +++ b/tools/perf/Documentation/tips.txt > @@ -62,3 +62,6 @@ To show context switches in perf report sample context add --switch-events to pe > To show time in nanoseconds in record/report add --ns > To compare hot regions in two workloads use perf record -b -o file ... ; perf diff --stream file1 file2 > To compare scalability of two workload samples use perf diff -c ratio file1 file2 > +For latency profiling, try: perf record/report --latency > +For parallelism histogram, try: perf report --hierarchy --sort latency,parallelism,comm,symbol Ditto. Thanks, Namhyung > +To analyze particular parallelism levels, try: perf report --latency --parallelism=32-64 > -- > 2.48.1.362.g079036d154-goog >