Add a new --quiet option to 'perf stat'. This is useful with 'perf stat
record' to write the data only to the perf.data file, which can lower
measurement overhead because the data doesn't need to be formatted.
On my 4C desktop:
% time ./perf stat record -e $(python -c 'print ",".join(["cycles"]*1000)') -a -I 1000 sleep 5
...
real 0m5.377s
user 0m0.238s
sys 0m0.452s
% time ./perf stat record --quiet -e $(python -c 'print ",".join(["cycles"]*1000)') -a -I 1000 sleep 5
real 0m5.452s
user 0m0.183s
sys 0m0.423s
In this example it cuts the user time by 20%. On systems with more cores
the savings are higher.
Signed-off-by: Andi Kleen <andi@firstfloor.org>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexey Budankov <alexey.budankov@linux.intel.com>
Link: http://lore.kernel.org/lkml/20201027002737.30942-1-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The --for-each-cgroup option is a syntax sugar to monitor large number
of cgroups easily. Current command line requires to list all the events
and cgroups even if users want to monitor same events for each cgroup.
This patch addresses that usage by copying given events for each cgroup
on user's behalf.
For instance, if they want to monitor 6 events for 200 cgroups each they
should write 1200 event names (with -e) AND 1200 cgroup names (with -G)
on the command line. But with this change, they can just specify 6
events and 200 cgroups with a new option.
A simpler example below: It wants to measure 3 events for 2 cgroups ('A'
and 'B'). The result is that total 6 events are counted like below.
$ perf stat -a -e cpu-clock,cycles,instructions --for-each-cgroup A,B sleep 1
Performance counter stats for 'system wide':
988.18 msec cpu-clock A # 0.987 CPUs utilized
3,153,761,702 cycles A # 3.200 GHz (100.00%)
8,067,769,847 instructions A # 2.57 insn per cycle (100.00%)
982.71 msec cpu-clock B # 0.982 CPUs utilized
3,136,093,298 cycles B # 3.182 GHz (99.99%)
8,109,619,327 instructions B # 2.58 insn per cycle (99.99%)
1.001228054 seconds time elapsed
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20200924124455.336326-3-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add support for new "stat.big-num" boolean option.
This allows a user to set a default for "--no-big-num" for "perf stat"
commands.
--
$ perf config stat.big-num
$ perf stat --event cycles /bin/true
Performance counter stats for '/bin/true':
778,849 cycles
[...]
$ perf config stat.big-num=false
$ perf config stat.big-num
stat.big-num=false
$ perf stat --event cycles /bin/true
Performance counter stats for '/bin/true':
769622 cycles
[...]
--
There is an interaction with "--field-separator" that must be
accommodated, such that specifying "--big-num --field-separator={x}"
still reports an invalid combination of options.
Documentation for perf-config and perf-stat updated.
Signed-off-by: Paul Clarke <pc@us.ibm.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Link: http://lore.kernel.org/lkml/1589991815-17951-1-git-send-email-pc@us.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
For interval mode, the metric is printed after the '#' character if it
exists. But it's not calculated by the counts generated in this
interval.
See the following examples:
root@kbl-ppc:~# perf stat -M CPI -I1000 --interval-count 2
# time counts unit events
1.000422803 764,809 inst_retired.any # 2.9 CPI
1.000422803 2,234,932 cycles
2.001464585 1,960,061 inst_retired.any # 1.6 CPI
2.001464585 4,022,591 cycles
The second CPI should not be 1.6 (4,022,591/1,960,061 is 2.1)
root@kbl-ppc:~# perf stat -e cycles,instructions -I1000 --interval-count 2
# time counts unit events
1.000429493 2,869,311 cycles
1.000429493 816,875 instructions # 0.28 insn per cycle
2.001516426 9,260,973 cycles
2.001516426 5,250,634 instructions # 0.87 insn per cycle
The second 'insn per cycle' should not be 0.87 (5,250,634/9,260,973 is
0.57).
The current code uses a global variable 'rt_stat' for tracking and
updating the std dev of runtime stat. Unlike the counts, 'rt_stat' is not
reset for interval. While the counts are reset for interval.
perf_stat_process_counter()
{
if (config->interval)
init_stats(ps->res_stats);
}
So for interval mode, the 'rt_stat' variable should be reset too.
This patch resets 'rt_stat' before read_counters(), so the runtime stat
is only calculated by the counts generated in this interval.
With this patch:
root@kbl-ppc:~# perf stat -M CPI -I1000 --interval-count 2
# time counts unit events
1.000420924 2,408,818 inst_retired.any # 2.1 CPI
1.000420924 5,010,111 cycles
2.001448579 2,798,407 inst_retired.any # 1.6 CPI
2.001448579 4,599,861 cycles
root@kbl-ppc:~# perf stat -e cycles,instructions -I1000 --interval-count 2
# time counts unit events
1.000428555 2,769,714 cycles
1.000428555 774,462 instructions # 0.28 insn per cycle
2.001471562 3,595,904 cycles
2.001471562 1,243,703 instructions # 0.35 insn per cycle
Now the second 'insn per cycle' and CPI are calculated by the counts
generated in this interval.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Tested-By: Kajol Jain <kjain@linux.ibm.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jin Yao <yao.jin@intel.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/20200420145417.6864-1-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We have supported the event modifier "percore" which sums up the event
counts for all hardware threads in a core and show the counts per core.
For example,
# perf stat -e cpu/event=cpu-cycles,percore/ -a -A -- sleep 1
Performance counter stats for 'system wide':
S0-D0-C0 395,072 cpu/event=cpu-cycles,percore/
S0-D0-C1 851,248 cpu/event=cpu-cycles,percore/
S0-D0-C2 954,226 cpu/event=cpu-cycles,percore/
S0-D0-C3 1,233,659 cpu/event=cpu-cycles,percore/
This patch provides a new option "--percore-show-thread". It is used
with event modifier "percore" together to sum up the event counts for
all hardware threads in a core but show the counts per hardware thread.
This is essentially a replacement for the any bit (which is gone in
Icelake). Per core counts are useful for some formulas, e.g. CoreIPC.
The original percore version was inconvenient to post process. This
variant matches the output of the any bit.
With this patch, for example,
# perf stat -e cpu/event=cpu-cycles,percore/ -a -A --percore-show-thread -- sleep 1
Performance counter stats for 'system wide':
CPU0 2,453,061 cpu/event=cpu-cycles,percore/
CPU1 1,823,921 cpu/event=cpu-cycles,percore/
CPU2 1,383,166 cpu/event=cpu-cycles,percore/
CPU3 1,102,652 cpu/event=cpu-cycles,percore/
CPU4 2,453,061 cpu/event=cpu-cycles,percore/
CPU5 1,823,921 cpu/event=cpu-cycles,percore/
CPU6 1,383,166 cpu/event=cpu-cycles,percore/
CPU7 1,102,652 cpu/event=cpu-cycles,percore/
We can see counts are duplicated in CPU pairs (CPU0/CPU4, CPU1/CPU5,
CPU2/CPU6, CPU3/CPU7).
The interval mode also works. For example,
# perf stat -e cpu/event=cpu-cycles,percore/ -a -A --percore-show-thread -I 1000
# time CPU counts unit events
1.000425421 CPU0 925,032 cpu/event=cpu-cycles,percore/
1.000425421 CPU1 430,202 cpu/event=cpu-cycles,percore/
1.000425421 CPU2 436,843 cpu/event=cpu-cycles,percore/
1.000425421 CPU3 1,192,504 cpu/event=cpu-cycles,percore/
1.000425421 CPU4 925,032 cpu/event=cpu-cycles,percore/
1.000425421 CPU5 430,202 cpu/event=cpu-cycles,percore/
1.000425421 CPU6 436,843 cpu/event=cpu-cycles,percore/
1.000425421 CPU7 1,192,504 cpu/event=cpu-cycles,percore/
If we offline CPU5, the result is:
# perf stat -e cpu/event=cpu-cycles,percore/ -a -A --percore-show-thread -- sleep 1
Performance counter stats for 'system wide':
CPU0 2,752,148 cpu/event=cpu-cycles,percore/
CPU1 1,009,312 cpu/event=cpu-cycles,percore/
CPU2 2,784,072 cpu/event=cpu-cycles,percore/
CPU3 2,427,922 cpu/event=cpu-cycles,percore/
CPU4 2,752,148 cpu/event=cpu-cycles,percore/
CPU6 2,784,072 cpu/event=cpu-cycles,percore/
CPU7 2,427,922 cpu/event=cpu-cycles,percore/
1.001416041 seconds time elapsed
v4:
---
Ravi Bangoria reports an issue in v3. Once we offline a CPU,
the output is not correct. The issue is we should use the cpu
idx in print_percore_thread rather than using the cpu value.
v3:
---
1. Fix the interval mode output error
2. Use cpu value (not cpu index) in config->aggr_get_id().
3. Refine the code according to Jiri's comments.
v2:
---
Add the explanation in change log. This is essentially a replacement
for the any bit. No code change.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Tested-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/20200214080452.26402-1-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
It is useful to aggregate counts per die. E.g. Uncore becomes die-scope
on Xeon Cascade Lake-AP.
Introduce a new option "--per-die" to support per-die aggregation.
The global id for each core has been changed to socket + die id + core
id. The global id for each die is socket + die id.
Add die information for per-core aggregation. The output of per-core
aggregation will be changed from "S0-C0" to "S0-D0-C0". Any scripts
which rely on the output format of per-core aggregation probably be
broken.
For 'perf stat record/report', there is no die information when
processing the old perf.data. The per-die result will be the same as
per-socket.
Committer notes:
Renamed 'die' variable to 'die_id' to fix the build in some systems:
CC /tmp/build/perf/builtin-script.o
cc1: warnings being treated as errors
builtin-stat.c: In function 'perf_env__get_die':
builtin-stat.c:963: error: declaration of 'die' shadows a global declaration
util/util.h:19: error: shadowed declaration is here
mv: cannot stat `/tmp/build/perf/.builtin-stat.o.tmp': No such file or directory
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lkml.kernel.org/n/tip-bsnhx7vgsuu6ei307mw60mbj@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
With this patch, we can use the 'percore' event qualifier in perf-stat.
root@skl:/tmp# perf stat -e cpu/event=0,umask=0x3,percore=1/,cpu/event=0,umask=0x3/ -a -A -I1000
1.000773050 S0-C0 98,352,832 cpu/event=0,umask=0x3,percore=1/ (50.01%)
1.000773050 S0-C1 103,763,057 cpu/event=0,umask=0x3,percore=1/ (50.02%)
1.000773050 S0-C2 196,776,995 cpu/event=0,umask=0x3,percore=1/ (50.02%)
1.000773050 S0-C3 176,493,779 cpu/event=0,umask=0x3,percore=1/ (50.02%)
1.000773050 CPU0 47,699,641 cpu/event=0,umask=0x3/ (50.02%)
1.000773050 CPU1 49,052,451 cpu/event=0,umask=0x3/ (49.98%)
1.000773050 CPU2 102,771,422 cpu/event=0,umask=0x3/ (49.98%)
1.000773050 CPU3 100,784,662 cpu/event=0,umask=0x3/ (49.98%)
1.000773050 CPU4 43,171,342 cpu/event=0,umask=0x3/ (49.98%)
1.000773050 CPU5 54,152,158 cpu/event=0,umask=0x3/ (49.98%)
1.000773050 CPU6 93,618,410 cpu/event=0,umask=0x3/ (49.98%)
1.000773050 CPU7 74,477,589 cpu/event=0,umask=0x3/ (49.99%)
In this example, we count the event 'ref-cycles' per-core and per-CPU in
one perf stat command-line. From the output, we can see:
S0-C0 = CPU0 + CPU4
S0-C1 = CPU1 + CPU5
S0-C2 = CPU2 + CPU6
S0-C3 = CPU3 + CPU7
So the result is expected (tiny difference is ignored).
Note that, the 'percore' event qualifier needs to use with option '-A'.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Tested-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jin Yao <yao.jin@intel.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1555077590-27664-4-git-send-email-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The -c option to enable multiplex scaling has been useless for quite
some time because scaling is default.
It's only useful as --no-scale to disable scaling. But the non scaling
code path has bitrotted and doesn't print anything because perf output
code relies on value run/ena information.
Also even when we don't want to scale a value it's still useful to show
its multiplex percentage.
This patch:
- Fixes help and documentation to show --no-scale instead of -c
- Removes -c, only keeps the long option because -c doesn't support negatives.
- Enables running/enabled even with --no-scale
- And fixes some other problems in the no-scale output.
Before:
$ perf stat --no-scale -e cycles true
Performance counter stats for 'true':
<not counted> cycles
0.000984154 seconds time elapsed
After:
$ ./perf stat --no-scale -e cycles true
Performance counter stats for 'true':
706,070 cycles
0.001219821 seconds time elapsed
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
LPU-Reference: 20190314225002.30108-9-andi@firstfloor.org
Link: https://lkml.kernel.org/n/tip-xggjvwcdaj2aqy8ib3i4b1g6@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Go over the tools/ files that are maintained in Arnaldo's tree and
fix common typos: half of them were in comments, the other half
in JSON files.
No change in functionality intended.
Committer notes:
This was split from a larger patch as there are code that is,
additionally, maintained outside the kernel tree, so to ease cherry
picking and/or backporting, split this into multiple patches.
In this particular case, it affects documentation, so may be interesting
to cherry pick as it is information that is presented to the user.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20181203102200.GA104797@gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Adding the support to read rusage data once the workload is finished and
display the system/user time values:
$ perf stat --null perf bench sched pipe
...
Performance counter stats for 'perf bench sched pipe':
5.342599256 seconds time elapsed
2.544434000 seconds user
4.549691000 seconds sys
It works only in non -r mode and only for workload target.
So as of now, for workload targets, we display 3 types of timings. The
time we meassure in perf stat from enable to disable+period:
5.342599256 seconds time elapsed
The time spent in user and system lands, displayed only for workload
session/target:
2.544434000 seconds user
4.549691000 seconds sys
Those times are the very same displayed by 'time' tool. They are
returned by wait4 call via the getrusage struct interface.
Committer notes:
Had to rename some variables to avoid this on older systems such as
centos:6:
builtin-stat.c: In function 'print_footer':
builtin-stat.c:1831: warning: declaration of 'stime' shadows a global declaration
/usr/include/time.h:297: warning: shadowed declaration is here
Committer testing:
# perf stat --null time perf bench sched pipe
# Running 'sched/pipe' benchmark:
# Executed 1000000 pipe operations between two processes
Total time: 5.526 [sec]
5.526534 usecs/op
180945 ops/sec
1.00user 6.25system 0:05.52elapsed 131%CPU (0avgtext+0avgdata 8056maxresident)k
0inputs+0outputs (0major+606minor)pagefaults 0swaps
Performance counter stats for 'time perf bench sched pipe':
5.530978744 seconds time elapsed
1.004037000 seconds user
6.259937000 seconds sys
#
Suggested-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20180605121313.31337-1-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently print count interval for performance counters values is
limited by 10ms so reading the values at frequencies higher than 100Hz
is restricted by the tool.
This change makes perf stat -I possible on frequencies up to 1KHz and,
to some extent, makes perf stat -I to be on-par with perf record
sampling profiling.
When running perf stat -I for monitoring e.g. PCIe uncore counters and
at the same time profiling some I/O workload by perf record e.g. for
cpu-cycles and context switches, it is then possible to observe
consolidated CPU/OS/IO(Uncore) performance picture for that workload.
Tool overhead warning printed when specifying -v option can be missed
due to screen scrolling in case you have output to the console
so message is moved into help available by running perf stat -h.
Signed-off-by: Alexey Budankov <alexey.budankov@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/b842ad6a-d606-32e4-afe5-974071b5198e@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Auto-merge for these events was disabled when auto-merging of non-alias
events was disabled in commit 63ce844 (perf stat: Only auto-merge events
that are PMU aliases).
Non-merging of legacy events is preserved:
$ perf stat -ag -e cache-misses,cache-misses sleep 1
Performance counter stats for 'system wide':
86,323 cache-misses
86,323 cache-misses
1.002623307 seconds time elapsed
But prefix or glob matching auto-merges the events created:
$ perf stat -a -e l3cache/read-miss/ sleep 1
Performance counter stats for 'system wide':
328 l3cache/read-miss/
1.002627008 seconds time elapsed
$ perf stat -a -e l3cache_0_[01]/read-miss/ sleep 1
Performance counter stats for 'system wide':
172 l3cache/read-miss/
1.002627008 seconds time elapsed
As with events created with aliases, auto-merging can be suppressed with
the --no-merge option:
$ perf stat -a -e l3cache/read-miss/ --no-merge sleep 1
Performance counter stats for 'system wide':
67 l3cache/read-miss/
67 l3cache/read-miss/
63 l3cache/read-miss/
60 l3cache/read-miss/
1.002622192 seconds time elapsed
Signed-off-by: Agustin Vega-Frias <agustinv@codeaurora.org>
Acked-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Timur Tabi <timur@codeaurora.org>
Cc: linux-arm-kernel@lists.infradead.org
Change-Id: I0a47eed54c05e1982ca964d743b37f50f60c508c
Link: http://lkml.kernel.org/r/1520345084-42646-4-git-send-email-agustinv@codeaurora.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Starting on v4.12 event parsing code for dynamic pmu events already
supports prefix-based matching of multiple pmus when creating dynamic
events. E.g., in a system with the following dynamic pmus:
mypmu_0
mypmu_1
mypmu_2
mypmu_4
passing mypmu/<config>/ as an event spec will result in the creation of
the event in all of the pmus. This change expands this matching through
the use of fnmatch so glob-like expressions can be used to create events
in multiple pmus. E.g., in the system described above if a user only
wants to create the event in mypmu_0 and mypmu_1, mypmu_[01]/<config>/
can be passed.
Signed-off-by: Agustin Vega-Frias <agustinv@codeaurora.org>
Acked-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: linux-arm-kernel@lists.infradead.org
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Timur Tabi <timur@codeaurora.org>
Change-Id: Icb25653fc5d5239c20f3bffdfdf4ab4c9c9bb20b
Link: http://lkml.kernel.org/r/1520454947-16977-1-git-send-email-agustinv@codeaurora.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When using -G with one cgroup and -e with multiple events, only the
first event gets the correct cgroup setting, all events from the second
onwards will track system-wide events.
If the user wants to track multiple events for a specific cgroup, the
user must give parameters like the following:
$ perf stat -e e1 -e e2 -e e3 -G test,test,test
This patch simplify this case, just type one cgroup:
$ perf stat -e e1 -e e2 -e e3 -G test
$ mkdir -p /sys/fs/cgroup/perf_event/empty_cgroup
$ perf stat -e cycles -e cache-misses -a -I 1000 -G empty_cgroup
Before:
1.001007226 <not counted> cycles empty_cgroup
1.001007226 7,506 cache-misses
After:
1.000834097 <not counted> cycles empty_cgroup
1.000834097 <not counted> cache-misses empty_cgroup
Signed-off-by: weiping zhang <zhangweiping@didichuxing.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20180129154805.GA6284@localhost.didichuxing.com
[ Improved the doc text a bit, providing an example for cgroup + system wide counting ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Introduce a new option to print counts after N milliseconds and update
'perf stat' documentation accordingly.
Show below is the output of the new option for perf stat.
$ perf stat --time 2000 -e cycles -a
Performance counter stats for 'system wide':
157,260,423 cycles
2.003060766 seconds time elapsed
We can print the count deltas after N milliseconds with this new
introduced option. This option is not supported with "-I" option.
In addition, according to Kangliang's patch(19afd10410), the
monitoring overhead for system-wide core event could be very high if the
interval-print parameter was below 100ms, and the limitation value is
10ms.
So the same warning will be displayed when the time is set between 10ms
to 100ms, and the minimal time is limited to 10ms. Users can make a
decision according to their spcific cases.
Committer notes:
This actually stops the workload after the specified time, then prints
the counts.
So I renamed the option to --timeout and updated the documentation to
state that it will not just print the counts after the specified time,
but will really stop the 'perf stat' session and print the counts.
The rename from 'time' to 'timeout' also fixes the build in systems
where 'time' is used by glibc and can't be used as a name of a variable,
such as centos:5 and centos:6.
Changes since v3:
- none.
Changes since v2:
- modify the time check in __run_perf_stat func to keep some consistency
with the workload case.
- add the warning when the time is set between 10ms to 100ms.
- add the pr_err when the time is set below 10ms.
Changes since v1:
- none.
Signed-off-by: yuzhoujian <yuzhoujian@didichuxing.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Milian Wolff <milian.wolff@kdab.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/1517217923-8302-3-git-send-email-ufo19890607@gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Introduce a new option to print counts for fixed number of times and
update 'perf stat' documentation accordingly.
Show below is the output of the new option for perf stat.
$ perf stat -I 1000 --interval-count 2 -e cycles -a
# time counts unit events
1.002827089 93,884,870 cycles
2.004231506 56,573,446 cycles
We can just print the counts for several times with this newly
introduced option. The usage of it is a little like 'vmstat', and it
should be used together with "-I" option.
$ vmstat -n 1 2
procs ---------memory-------------- --swap- ----io-- -system-- ------cpu---
r b swpd free buff cache si so bi bo in cs us sy id wa st
0 0 0 78270544 547484 51732076 0 0 0 20 1 1 1 0 99 0 0
0 0 0 78270512 547484 51732080 0 0 0 16 477 1555 0 0 100 0 0
Changes since v3:
- merge interval_count check and times check to one line.
- fix the wrong indent in stat.h
- use stat_config.times instead of 'times' in cmd_stat function.
Changes since v2:
- none.
Changes since v1:
- change the name of the new option "times-print" to "interval-count".
- keep the new option interval specifically.
Signed-off-by: yuzhoujian <yuzhoujian@didichuxing.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Milian Wolff <milian.wolff@kdab.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/1517217923-8302-2-git-send-email-ufo19890607@gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add generic support for standalone metrics specified in JSON files to
perf stat. A metric is a formula that uses multiple events to compute a
higher level result (e.g. IPC).
Previously metrics were always tied to an event and automatically
enabled with that event. But now change it that we can have standalone
metrics. They are in the same JSON data structure as events, but don't
have an event name.
We also allow to organize the metrics in metric groups, which allows a
short cut to select several related metrics at once.
Add a new -M / --metrics option to perf stat that adds the metrics or
metric groups specified.
Add the core code to manage and parse the metric groups. They are
collected from the JSON data structures into a separate rblist. When
computing shadow values look for metrics in that list. Then they are
computed using the existing saved values infrastructure in stat-shadow.c
The actual JSON metrics are in a separate pull request.
% perf stat -M Summary --metric-only -a sleep 1
Performance counter stats for 'system wide':
Instructions CLKS CPU_Utilization GFLOPs SMT_2T_Utilization Kernel_Utilization
317614222.0 1392930775.0 0.0 0.0 0.2 0.1
1.001497549 seconds time elapsed
% perf stat -M GFLOPs flops
Performance counter stats for 'flops':
3,999,541,471 fp_comp_ops_exe.sse_scalar_single # 1.2 GFLOPs (66.65%)
14 fp_comp_ops_exe.sse_scalar_double (66.65%)
0 fp_comp_ops_exe.sse_packed_double (66.67%)
0 fp_comp_ops_exe.sse_packed_single (66.70%)
0 simd_fp_256.packed_double (66.70%)
0 simd_fp_256.packed_single (66.67%)
0 duration_time
3.238372845 seconds time elapsed
v2: Add missing header file
v3: Move find_map to pmu.c
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170831194036.30146-7-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Implementing a new --smi-cost mode in perf stat to measure SMI cost.
During the measurement, the /sys/device/cpu/freeze_on_smi will be set.
The measurement can be done with one counter (unhalted core cycles), and
two free running MSR counters (IA32_APERF and SMI_COUNT).
In practice, the percentages of SMI core cycles should be more useful
than absolute value. So the output will be the percentage of SMI core
cycles and SMI#. metric_only will be set by default.
SMI cycles% = (aperf - unhalted core cycles) / aperf
Here is an example output.
Performance counter stats for 'sudo echo ':
SMI cycles% SMI#
0.1% 1
0.010858678 seconds time elapsed
Users who wants to get the actual value can apply additional
--no-metric-only.
Signed-off-by: Kan Liang <Kan.liang@intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Robert Elliott <elliott@hpe.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1495825538-5230-3-git-send-email-kan.liang@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The uncore PMU has a lot of duplicated PMUs for different subsystems.
When expanding an uncore alias we usually end up with a large
number of identically named aliases, which makes perf stat
output difficult to read.
Automatically sum them up in perf stat, unless --no-merge is specified.
This can be default because only the uncores generally have duplicated
aliases. Other PMUs have unique names.
Before:
% perf stat --no-merge -a -e unc_c_llc_lookup.any sleep 1
Performance counter stats for 'system wide':
694,976 Bytes unc_c_llc_lookup.any
706,304 Bytes unc_c_llc_lookup.any
956,608 Bytes unc_c_llc_lookup.any
782,720 Bytes unc_c_llc_lookup.any
605,696 Bytes unc_c_llc_lookup.any
442,816 Bytes unc_c_llc_lookup.any
659,328 Bytes unc_c_llc_lookup.any
509,312 Bytes unc_c_llc_lookup.any
263,936 Bytes unc_c_llc_lookup.any
592,448 Bytes unc_c_llc_lookup.any
672,448 Bytes unc_c_llc_lookup.any
608,640 Bytes unc_c_llc_lookup.any
641,024 Bytes unc_c_llc_lookup.any
856,896 Bytes unc_c_llc_lookup.any
808,832 Bytes unc_c_llc_lookup.any
684,864 Bytes unc_c_llc_lookup.any
710,464 Bytes unc_c_llc_lookup.any
538,304 Bytes unc_c_llc_lookup.any
1.002577660 seconds time elapsed
After:
% perf stat -a -e unc_c_llc_lookup.any sleep 1
Performance counter stats for 'system wide':
2,685,120 Bytes unc_c_llc_lookup.any
1.002648032 seconds time elapsed
v2: Split collect_aliases. Rename alias flag.
v3: Make sure unsupported/not counted is always printed.
v4: Factor out callback change into separate patch.
v5: Move check for bad results here
Move merged check into collect_data
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170320201711.14142-3-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Boris asked for default -a option in case we monitor only uncore events.
While implementing that I thought it might be actually useful to make it
overall default.
Running 'perf stat' will now collect system wide data.
Committer note:
Testing it:
# perf stat
^C
Performance counter stats for 'system wide':
3571.559178 cpu-clock (msec) # 4.000 CPUs utilized
3,346 context-switches # 0.937 K/sec
277 cpu-migrations # 0.078 K/sec
57,271 page-faults # 0.016 M/sec
4,535,633,835 cycles # 1.270 GHz
6,389,736,516 instructions # 1.41 insn per cycle
1,541,293,875 branches # 431.547 M/sec
14,526,396 branch-misses # 0.94% of all branches
0.892950118 seconds time elapsed
#
Requested-and-Acked-by: Borislav Petkov <bp@alien8.de>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20170217170034.GB15389@krava
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add basic plumbing for TopDown in perf stat
TopDown is intended to replace the frontend cycles idle/ backend cycles
idle metrics in standard perf stat output. These metrics are not
reliable in many workloads, due to out of order effects.
This implements a new --topdown mode in perf stat (similar to
--transaction) that measures the pipe line bottlenecks using
standardized formulas. The measurement can be all done with 5 counters
(one fixed counter)
The result are four metrics:
FrontendBound, BackendBound, BadSpeculation, Retiring
that describe the CPU pipeline behavior on a high level.
The full top down methology has many hierarchical metrics. This
implementation only supports level 1 which can be collected without
multiplexing. A full implementation of top down on top of perf is
available in pmu-tools toplev. (http://github.com/andikleen/pmu-tools)
The current version works on Intel Core CPUs starting with Sandy Bridge,
and Atom CPUs starting with Silvermont. In principle the generic
metrics should be also implementable on other out of order CPUs.
TopDown level 1 uses a set of abstracted metrics which are generic to
out of order CPU cores (although some CPUs may not implement all of
them):
topdown-total-slots Available slots in the pipeline
topdown-slots-issued Slots issued into the pipeline
topdown-slots-retired Slots successfully retired
topdown-fetch-bubbles Pipeline gaps in the frontend
topdown-recovery-bubbles Pipeline gaps during recovery
from misspeculation
These metrics then allow to compute four useful metrics:
FrontendBound, BackendBound, Retiring, BadSpeculation.
Add a new --topdown options to enable events. When --topdown is
specified set up events for all topdown events supported by the kernel.
Add topdown-* as a special case to the event parser, as is needed for
all events containing -.
The actual code to compute the metrics is in follow-on patches.
v2: Use standard sysctl read function.
v3: Move x86 specific code to arch/
v4: Enable --metric-only implicitly for topdown.
v5: Add --single-thread option to not force per core mode
v6: Fix output order of topdown metrics
v7: Allow combining with -d
v8: Remove --single-thread again
v9: Rename functions, adding arch_ and topdown_.
v10: Expand man page and describe TopDown better
Paste intro into commit description.
Print error when malloc fails.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/1464119559-17203-1-git-send-email-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add a new mode to only print metrics. Sometimes we don't care about the
raw values, just want the computed metrics. This allows more compact
printing, so with -I each sample is only a single line. This also
allows easier plotting and processing with other tools.
The main target is with using --topdown, but it also works with -T and
standard perf stat. A few metrics are not supported.
To avoiding having to hardcode all the metrics in the code it uses a two
pass approach: first compute dummy metrics and only print the headers in
the print_metric callback. Then use the callback to print the actual
values.
There are some additional changes in the stat printout code to handle
all metrics being on a single line.
One issue is that the column code doesn't know in advance what events
are not supported by the CPU, and it would be hard to find out as this
could change based on dynamic conditions. That causes empty columns in
some cases.
The output can be fairly wide, often you may need more than 80 columns.
Example:
% perf stat -a -I 1000 --metric-only
1.001452803 frontend cycles idle insn per cycle stalled cycles per insn branch-misses of all branches
1.001452803 158.91% 0.66 2.39 2.92%
2.002192321 180.63% 0.76 2.08 2.96%
3.003088282 150.59% 0.62 2.57 2.84%
4.004369835 196.20% 0.98 1.62 3.79%
5.005227314 231.98% 0.84 1.90 4.71%
v2: Lots of updates.
v3: Use slightly narrower columns
v4: Add comment
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/1457049458-28956-6-git-send-email-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
With all the recently added fields in the perf stat CSV output we should
finally document them in the man page. Do this here.
v2: Fix fields in documentation (Jiri)
v3: fix order of fields again (Jiri)
v4: Change order again.
v5: Document more fields (Jiri)
v6: Move time stamp first
v7: More fixes (Jiri)
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/1457049458-28956-5-git-send-email-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add 'perf stat record' command support. It creates simple (header only)
perf.data file ATM.
The record command could be specified anywhere among stat options. All
stat command options are valid for stat record command with '-o' option
exception. If specified for record command it denotes the perf data file
name.
Committer note:
Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
while avoiding that older tools show confusing messages, for instance,
with sample_type = 0, we get:
$ perf stat record usleep 1
Performance counter stats for 'usleep 1':
0.630237 task-clock (msec) # 0.528 CPUs utilized
1 context-switches # 0.002 M/sec
0 cpu-migrations # 0.000 K/sec
52 page-faults # 0.083 M/sec
978,312 cycles # 1.552 GHz
671,931 stalled-cycles-frontend # 68.68% frontend cycles idle
<not supported> stalled-cycles-backend
646,379 instructions # 0.66 insns per cycle
# 1.04 stalled cycles per insn
131,046 branches # 207.931 M/sec
7,073 branch-misses # 5.40% of all branches
0.001193240 seconds time elapsed
$ oldperf evlist
WARNING: The perf.data file's data size field is 0 which is unexpected.
Was the 'perf record' command properly terminated?
non matching sample_type
$
While with sample_type set to PERF_SAMPLE_IDENTIFIER, after we re-run 'perf
stat record usleep' we get:
$ oldperf evlist
WARNING: The perf.data file's data size field is 0 which is unexpected.
Was the 'perf record' command properly terminated?
task-clock
context-switches
cpu-migrations
page-faults
cycles
stalled-cycles-frontend
stalled-cycles-backend
instructions
branches
branch-misses
$
Which at least shows the names of the events in the perf.data file.
Additionally, such files, when passed to 'perf report' will produce:
$ oldperf report --stdio
WARNING: The perf.data file's data size field is 0 which is unexpected.
Was the 'perf record' command properly terminated?
Warning:
Kernel address maps (/proc/{kallsyms,modules}) were restricted.
Check /proc/sys/kernel/kptr_restrict before running 'perf record'.
As no suitable kallsyms nor vmlinux was found, kernel samples
can't be resolved.
Samples in kernel modules can't be resolved as well.
Error:
The perf.data file has no samples!
# To display the perf.data header info, please use --header/--header-only options.
#
$
Which is confusing and can be solved by just adding the kernel mmap record,
which will also remove that warning about the data size field being equal to
zero, after generating the mmap record:
$ perf stat record usleep 1
Performance counter stats for 'usleep 1':
0.600796 task-clock (msec) # 0.478 CPUs utilized
1 context-switches # 0.002 M/sec
0 cpu-migrations # 0.000 K/sec
54 page-faults # 0.090 M/sec
886,844 cycles # 1.476 GHz
582,169 stalled-cycles-frontend # 65.65% frontend cycles idle
<not supported> stalled-cycles-backend
638,344 instructions # 0.72 insns per cycle
# 0.91 stalled cycles per insn
130,204 branches # 216.719 M/sec
7,500 branch-misses # 5.76% of all branches
0.001255897 seconds time elapsed
$ oldperf evlist
task-clock
context-switches
cpu-migrations
page-faults
cycles
stalled-cycles-frontend
stalled-cycles-backend
instructions
branches
branch-misses
$ oldperf report --stdio
Error:
The perf.data file has no samples!
# To display the perf.data header info, please use --header/--header-only options.
#
[acme@zoo linux]$
No warnings, sensible output about what are the events in the perf.data file and also
a "file has no samples" message, which indeed it doesn't.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Tested-by: Kan Liang <kan.liang@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: htp://lkml.kernel.org/r/1446734469-11352-3-git-send-email-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The --interval-print parameter was limited to 100ms. However, for
example, 10ms is required to do sophisticated bandwidth analysis using
uncore events.
The test shows that the overhead of the system-wide uncore monitoring
with 10ms interval is only ~2%. So this patch reduces the minimal
interval-print allowd to 10ms.
But 10ms may not work well for all cases. For example, when the
cpus/threads number is very large, for system-wide core event monitoring
the overhead could be high.
To handle this issue, a warning will be displayed when the
interval-print is set between 10ms to 100ms. So users can make a
decision according to their specific cases.
# perf stat -e uncore_imc_1/cas_count_read/ -a --interval-print 10 -- sleep 1
print interval < 100ms. The overhead percentage could be high in some
cases. Please proceed with caution.
# time counts unit events
0.010200451 0.10 MiB uncore_imc_1/cas_count_read/
0.020475117 0.02 MiB uncore_imc_1/cas_count_read/
0.030692800 0.01 MiB uncore_imc_1/cas_count_read/
0.040948161 0.02 MiB uncore_imc_1/cas_count_read/
0.051159564 0.00 MiB uncore_imc_1/cas_count_read/
Signed-off-by: Kan Liang <kan.liang@intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/1443776674-42511-1-git-send-email-kan.liang@intel.com
[ Added warning about overhead when using sub 100ms intervals to the man page ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently all the -p option PID arguments tasks values get aggregated
and printed as single values.
Adding --per-tasks option to print values per task.
$ perf stat -e cycles,instructions --per-thread -p 30190,30242
^C
Performance counter stats for process id '30190,30242':
cat-30190 0 cycles
yes-30242 3,842,525,421 cycles
cat-30190 0 instructions
yes-30242 10,370,817,010 instructions
1.143155657 seconds time elapsed
Also works under interval mode:
$ perf stat -e cycles,instructions --per-thread -p 30190,30242 -I 1000
# time comm-pid counts unit events
1.000073435 cat-30190 89,058 cycles
1.000073435 yes-30242 3,360,786,902 cycles (100.00%)
1.000073435 cat-30190 14,066 instructions
1.000073435 yes-30242 9,069,937,462 instructions
2.000204830 cat-30190 0 cycles
2.000204830 yes-30242 3,351,667,626 cycles
2.000204830 cat-30190 0 instructions
2.000204830 yes-30242 9,045,796,885 instructions
^C 2.771286639 cat-30190 0 cycles
2.771286639 yes-30242 2,593,884,166 cycles
2.771286639 cat-30190 0 instructions
2.771286639 yes-30242 7,001,171,191 instructions
It works only with -t and -p options, otherwise following error is
printed:
$ perf stat -e cycles --per-thread -I 1000 ls
The --per-thread option is only available when monitoring via -p -t options.
-p, --pid <pid> stat events on existing process id
-t, --tid <tid> stat events on existing thread id
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/1435310967-14570-23-git-send-email-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add support to perf stat to print the basic transactional execution statistics:
Total cycles, Cycles in Transaction, Cycles in aborted transsactions
using the in_tx and in_tx_checkpoint qualifiers.
Transaction Starts and Elision Starts, to compute the average transaction
length.
This is a reasonable overview over the success of the transactions.
Also support architectures that have a transaction aborted cycles
counter like POWER8. Since that is awkward to handle in the kernel
abstract handle both cases here.
Enable with a new --transaction / -T option.
This requires measuring these events in a group, since they depend on each
other.
This is implemented by using TM sysfs events exported by the kernel
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Arnaldo Carvalho de Melo <acme@infradead.org>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1377128846-977-5-git-send-email-andi@firstfloor.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
